AI Mindset · Model Cheatsheets

Claude

The Senior Colleague Who Tells You When It Is Not Sure

Claude has moved beyond careful chat. Anthropic released Fable 5.1 and Mythos 5.1 on September 1, 2026: 1M token context by default, 128k maximum output, always-on adaptive thinking, the same $10 / $50 headline price as Fable 5, and cache reads four times cheaper. Code, Cowork, Tag, Design, Slides, Docs, Claude in Chrome and the Microsoft 365 integrations put it inside the places work happens. Two things belong in the procurement file ahead of the feature list: the 30-day retention requirement on 5.1, which a per-workspace switch in Console can satisfy without moving a whole estate off zero data retention, and two breaking changes that fail a working integration rather than warn it.

Verified September 18, 2026Fable 5.1 and Mythos 5.1 launched Sept 1, 20265.1 cache hits at $0.25 per MTokSonnet 5 stays $2 / $10; the rise did not happen30-day retention on 5.1, with a per-workspace overrideEFS promises ZDR-equivalent privacy from this fallCode · Cowork · Chrome · Tag · Design

1 / Meet Claude

The colleague who will not pretend the work is done

Claude’s most useful personality trait is judgment. It is unusually willing to identify missing evidence, challenge the brief and keep uncertainty visible while still moving the task forward.

Personality

Careful without being passive

Claude reads deeply, preserves nuance and usually prefers an explicit caveat to a neat but unsupported conclusion.

Strong source fidelity
Structured, professional writing
Useful self-review and error detection

Deploy it for

Work where judgment matters

Long documents, complex coding, regulated analysis, technical reviews and deliverables that need a defensible reasoning trail.

Board and client materials
Repository-scale engineering
Financial, legal and scientific workflows

Choose Claude when

The brief may itself be wrong

Claude is especially valuable when the input is incomplete, internally inconsistent or likely to push a model toward a polished fiction.

Ask it to inspect the premise
Invite direct disagreement
Require uncertainty to remain visible

Brief, then challenge

Analyze this proposal for the executive team. Before recommending anything, identify what is missing, internally inconsistent or too weakly evidenced to support a decision.

Claude responds well when permission to challenge the brief is explicit.

Truth-preserving synthesis

Create the strongest defensible synthesis from these sources. Separate: verified facts, reasonable inferences, unresolved questions and claims the material cannot support.

This keeps uncertainty useful instead of burying it in a disclaimer.

2 / What’s Current

A new top tier, a retention switch most buyers miss, and a lineup that moved under both

Anthropic released Claude Fable 5.1 and Claude Mythos 5.1 on September 1, 2026: model ids claude-fable-5-1 and claude-mythos-5-1, 1M token context by default, 128k maximum output and always-on adaptive thinking. Anthropic’s own framing is blunt: “They’re the world’s most advanced models for coding and knowledge work.” Treat that as a vendor claim. The headline price is unchanged from Fable 5 at $10 / $50 per million tokens on the standard tier, so the interesting number is further along the row: a cache hit on 5.1 costs $0.25 per million tokens against $1 on Fable 5, which is 0.025x base input where every other Claude model charges 0.1x. Two breaking changes ride along with the release, and the retention requirement adds a third failure path that is a configuration gate rather than a code change. The condition underneath all of it is retention: the 30-day requirement on 5.1 is real, it is not an unfixable gate, and Anthropic published both an interim zero data retention offer and a successor privacy program on the same day as the launch. One September 1 non-event matters too: the scheduled Sonnet 5 price rise did not occur. And between September 3 and September 18, no Claude model retired.

The 30-day retention requirement, and the workspace switch that unlocks it

Give this its own line in the evaluation, and read the override before the requirement. On Anthropic’s direct terms, Claude Fable 5.1 and Claude Mythos 5.1 do require 30-day data retention, and Anthropic’s September 1 API release note still says both models require it and are not available under zero data retention unless expressly authorized by Anthropic. Everything in this card describes that route; the terms Microsoft publishes for the same model as a subprocessor are a separate question, set out in its own callout further down this section. Read to that point, the requirement looks like a hard gate on a zero data retention organization. It is not, and the fix is a configuration flag Anthropic documents itself. Its wording: “Organizations with a ZDR arrangement can make these models available in a specific workspace by enabling 30-day retention for that workspace only. Other workspaces in the organization keep zero data retention.” The path is self-serve. In Claude Console, open Settings and then Workspaces, select the workspace, open its Privacy controls tab, and turn on 30-day data retention. It also answers the mixed-estate question. An estate with some traffic on 5.1 under retention and some on Opus 5 without it needs a routing rule a compliance reviewer can read rather than one buried in application code. The workspace boundary is that rule, and it is Anthropic’s rather than something you have to invent. The failure mode is readable too: a request from an organization or workspace without retention enabled returns a 400 invalid_request_error with the message “In order to access this model, your organization or workspace must have data retention enabled.” The retention page is a living document with no publication date, so read it yourself before any of this goes into a control description.

What “deleted after 30 days” actually says, and what a Covered Model is

The shorthand “deleted after 30 days” overstates the guarantee. Anthropic’s wording is that prompts and model completions “are retained for at least 30 days and then automatically deleted, unless they are subject to a safety investigation or we are legally required to maintain them.” At least 30 days rather than exactly 30, with a safety-investigation carve-out and a legal-requirement carve-out. Neither carve-out carries a duration in that wording, and the durations a privacy review actually stops on were read somewhere else. Microsoft’s living page “Anthropic models in Microsoft Online Services,” stamped September 10, 2026, is describing Anthropic’s own terms when it says that where Anthropic’s trust and safety classifiers identify a potential violation of Anthropic’s Usage Policy, Anthropic may retain content, meaning inputs and outputs, for up to two years, and trust and safety classification scores for up to seven years. Read those two figures for what they are: Microsoft describing Anthropic, on a page that is rewritten in place, reached here through the sibling Microsoft 365 Copilot guide in this set. Anthropic does not restate either duration on its own retention pages, so confirm both there before a data protection impact assessment rests on them. Two more things belong in the same paragraph. The models that carry the requirement are what Anthropic calls Covered Models, which it defines as Claude models whose capabilities, for example in software engineering, agentic workflows, scientific reasoning or cybersecurity, represent a substantial step up from prior generations and create elevated risk if misused. Four are designated: Mythos 5.1 and Fable 5.1, both on August 31, 2026, and Mythos 5 and Fable 5, both on June 9, 2026. And retained data does not always sit in the same place. On the Claude API, including the Claude Platform on AWS, Anthropic handles retained data. On Amazon Bedrock and Google Cloud’s Agent Platform, retained data stays within your cloud provider’s environment. That is the data-residency answer, and it is platform-specific rather than uniform. The Covered Models article carries no publication date, only a relative “last updated” line, so it can anchor content but not a date; confirm it directly.

Enterprise Frontier Safeguards and the interim zero data retention offer

Anthropic published this on September 1, 2026, the same day as the 5.1 launch, which is the main reason it goes unread. Enterprise Frontier Safeguards, EFS, is described as giving customers complete privacy, which Anthropic explicitly equates with a zero data retention policy, by having customers store their data on their own cloud infrastructure rather than on Anthropic’s systems. Anthropic says it is rolling out in phases, starting this fall, and that it will be supported on Claude Code, Claude Enterprise, the Claude Platform, Amazon Bedrock, the Claude Platform on AWS, Google’s Agent Platform and Microsoft Foundry. No dated milestone is published for any phase, so “this fall” is the whole of the schedule you have. In the meantime, Anthropic says that until EFS is available, eligible customers will be able to use Fable 5.1 with zero data retention, and the dedicated September 1 article extends that to Fable 5 as well and describes the arrangement as available for a limited time. The same article confirms the origin of the policy: Anthropic introduced 30-day data retention starting with Fable 5. Treat the interim offer as something to ask your account team for by name, with a date attached, rather than as a benefit you already hold.

Anthropic’s own pages disagree about zero data retention on 5.1

This is a live first-party conflict rather than one stale page, and it deserves the same handling as the Mythos access question: read both sides, settle neither on your own. On one side, the two September 1 news pages say eligible customers will be able to use Fable 5.1 with zero data retention until EFS is ready, and Fable 5 too. On the other, the September 1 API release note says both 5.1 models are not available under zero data retention unless expressly authorized by Anthropic, and the retention documentation says the same, that zero data retention is therefore not available for any of them unless expressly authorized. Anthropic does not publish the eligibility criteria for the interim offer, and it does not say whether being an eligible customer and being expressly authorized are the same status reached two ways or two different things. A buyer reading only the news pages and a buyer reading only the docs will build different plans. Put the question in writing, name both sentences in the email, and get the answer into the contract rather than into a slide.

Anthropic’s direct terms and Microsoft’s subprocessor terms for Fable 5.1 are not the same

The conflict above is internal to Anthropic. This one is not, and it is the reason two accurate readings of the same model can reach opposite answers. Everything above about retention describes Anthropic’s direct terms: Fable 5.1 is a Covered Model, 30-day retention is required, and the September 1 API release note says the model is not available under zero data retention unless expressly authorized by Anthropic. The sibling Microsoft 365 Copilot guide reads Microsoft’s living page “Anthropic models in Microsoft Online Services,” stamped September 10, 2026, and finds a different split on the Microsoft route. There the additional admin-center controls and the off-by-default gate attach to Fable 5.0, while for Fable 5.1 the page says that “for certain organizations, Fable 5.1 is available with Anthropic acting as a subprocessor and Microsoft’s Product Terms and Data Protection Addendum (DPA) apply,” and that under that path “Anthropic doesn’t retain customer content (including prompts and responses).” Both pages are accurate about what they describe, which is exactly the problem. Do not carry Microsoft’s sentence over as Anthropic’s, and do not assume the subprocessor path is open to your organization. Settle the prior question first: establish which set of terms governs your deployment, direct Anthropic contracting or a Microsoft commercial agreement with Anthropic as a subprocessor, before you treat either retention answer as final. On this point the route you buy through is doing more work than the model id.

Sonnet 5 is still $2 / $10, and the scheduled rise did not happen

Anthropic canceled the Sonnet 5 increase to $3 / $15 that had been scheduled for September 1, 2026. September 1 has now passed and the price did not move. Anthropic’s wording is unambiguous: “The $2/$10 per million input/output token pricing for Claude Sonnet 5, announced at launch as introductory pricing through August 31, 2026, is now the standard price. The previously scheduled increase to $3/$15 per million input/output tokens on September 1, 2026 will not occur.” Anyone who rebuilt a Sonnet 5 budget at $3 / $15 can rebuild it back down and stop hedging. One honest qualifier survives the good news: a price that has stopped being introductory is an ordinary list price with no committed floor, which is a weaker guarantee than a contract.

Sept 1

Claude Fable 5.1 and Claude Mythos 5.1 released, with 1M token context by default, 128k maximum output and always-on adaptive thinking

$0.25

Cache hit per million tokens on Fable 5.1 and Mythos 5.1: 0.025x base input, where every other Claude model charges 0.1x

$2 / $10

Sonnet 5 per MTok, now confirmed as the standard price; the increase to $3 / $15 scheduled for September 1, 2026 did not occur

30 days

Minimum retention on the four Covered Models. Anthropic’s wording is “at least 30 days,” with safety-investigation and legal carve-outs, and a zero data retention organization can enable it for a single Console workspace

EFS

Enterprise Frontier Safeguards, announced September 1, 2026: privacy Anthropic equates with zero data retention, with customer-hosted data, rolling out in phases starting this fall and with no dated milestone published

50%

Batch discount on both input and output: Fable 5.1 at $5 / $25, Opus 5 at $2.50 / $12.50, Sonnet 5 at $1 / $5 and Haiku 4.5 at $0.50 / $2.50

Three

Active Claude models now carrying a tentative retirement floor before November 30, 2026: the dated Sonnet 4.5, Haiku 4.5 and Opus 4.5 snapshots

400

The status code a 5.1 request returns when it uses tool_choice type any or tool, both of which are no longer supported

Model	Standard base input	5m cache write	1h cache write	Cache hit	Standard output
Claude Fable 5.1	$10	$12.50	$20	$0.25	$50
Claude Mythos 5.1, limited availability	$10	$12.50	$20	$0.25	$50
Claude Fable 5	$10	$12.50	$20	$1	$50
Claude Mythos 5, limited availability	$10	$12.50	$20	$1	$50
Claude Opus 5	$5	$6.25	$10	$0.50	$25
Claude Opus 4.8	$5	$6.25	$10	$0.50	$25
Claude Sonnet 5	$2	$2.50	$4	$0.20	$10
Claude Haiku 4.5	$1	$1.25	$2	$0.10	$5

Same headline price, four times cheaper on cache reads

Read the cache-hit column, not the headline. Fable 5.1 charges $10 per million input tokens and $50 per million output tokens, exactly what Fable 5 charges. But a cache hit on Fable 5.1 costs $0.25 per million tokens against $1 on Fable 5. Every other model in the table prices a cache hit at one tenth of base input; 5.1 prices it at one fortieth. For a workload that re-reads a large stable context on every call, which is most agent work, most document assistants and most long-running coding sessions, that single column is the difference between the two generations. Model your own cache-hit ratio before you assume the saving, because a workload that never reuses a prefix will not see any of it. And note the shape of the change: Anthropic did not cut the sticker price, it cut the price of the thing high-volume customers actually buy the most of.

Model	Where it sits	Status on September 18, 2026
Claude Fable 5.1	Claude API, Amazon Bedrock, Claude Platform on AWS, Google Cloud and Microsoft Foundry	Active, with a tentative retirement floor of “Not sooner than September 1, 2027”; the only Fable model on the models overview
Claude Mythos 5.1	Not generally available	Limited availability through three differently described routes; no retirement floor is published for it at all
Claude Fable 5	Still selectable and still priced at $10 / $50	Listed under the “Legacy models” heading on the pricing page, but Active on the deprecations page, not sooner than June 9, 2027
Claude Opus 5	On Pro, Max, Team and Enterprise; no longer the strongest model reachable on Pro	Released July 24, 2026; per-plan availability is now published by family, per-plan defaults by version number are not
Claude Opus 4.8	Still selectable and still priced, at exactly the Opus 5 rate	Legacy on the pricing page, Active on the deprecations page, not sooner than May 28, 2027
Claude Sonnet 5	Claude Free and Claude Pro	$2 / $10 confirmed as the standard price; the September 1 rise did not occur
Claude Haiku 4.5	Not a plan default; the cheapest priced tier	Active, but `claude-haiku-4-5-20251001` carries a tentative retirement floor of “Not sooner than October 15, 2026”
Claude Opus 4.1	Nowhere; the identifier stops serving requests	Retired August 5, 2026; Anthropic’s current migration pointer sends Opus 4.8 and earlier to Opus 5

Plan	Price as published	Model availability	Notable inclusions
Free	No cost	Sonnet and Haiku. Fable is shown as “No”	Family names only; the page never names a version number
Pro	The page gives two annual figures that do not reconcile, $17 per month billed annually and $200 up front, against $20 monthly	Opus, Sonnet and Haiku, with Fable on “Usage credits”	Claude Code; Claude Design, Slides and Docs; Claude in Chrome and Microsoft 365
Max	From $100 per month	Choose 5x or 20x more usage than Pro	Higher output limits for all tasks and priority access at high traffic times
Team	$20 per seat per month billed annually, $25 monthly; premium seats $100 billed annually, $125 monthly	Premium seats carry 5x more usage than standard seats	For teams of 2 to 150; includes Claude Code and Claude Cowork
Enterprise	Seat price plus usage at API rates. The US$20 per seat per month billed annually figure was read on claude.com/pricing on September 18, 2026 and is identical to a standard Team seat, so confirm it with Anthropic before it enters a quote	Not itemized by family on the pricing page	The Compliance API surfaces, including Claude in Chrome transcripts in beta

Which pricing page to read, and what it will not tell you

Anthropic publishes pricing at two hostnames and only one of them answers. anthropic.com/pricing returns a 406 and cannot be fetched at all; claude.com/pricing renders and sets out the plan and seat structure in the detail above. Read the one that works, and do not treat the refusal from the other as evidence that something is unpublished. Keep pinning a model id in anything you evaluate, because the page names model families, Opus, Sonnet, Haiku and Fable, and never version numbers, and the only places it uses the word “default” concern training rather than model selection. “Sonnet 5 is the Free default” and “Opus 5 is the Max default” remain July 24 claims that no current page restates at version granularity. One consequence is worth its own line: Fable on Pro is described elsewhere as 50% of weekly limits, which is a usage allowance rather than an entitlement, and it is still enough that Opus 5 is no longer the strongest model a Pro subscriber can reach. Two cells in the plan table are transcriptions rather than settled prices, and they are marked as such there. The Pro row carries both $17 per month billed annually and $200 up front, and twelve months at $17 is $204, so one of the two figures is rounded, stale or describing something the page does not spell out; both are printed here as the page gives them. The Enterprise row quotes US$20 per seat per month billed annually, which is the same number as a standard Team seat and one fifth of a Team premium seat, on a row whose model availability cell says the page does not itemize Enterprise at all. Both were read on claude.com/pricing on September 18, 2026. Neither is safe to put in a business case without confirming it with Anthropic. The pricing page is a living document; check it on the day.

Claude Fable 5.1 New, September 1

The new top of the generally available lineup, at model id claude-fable-5-1. Anthropic positions Fable 5.1 and Mythos 5.1 together as “the world’s most advanced models for coding and knowledge work,” which is a vendor claim and should be treated as one until your own evaluation says otherwise. The specification changes that matter operationally are the defaults: 1M token context is the default rather than a beta flag, maximum output is 128k, and adaptive thinking is always on, so there is no configuration in which this model is not reasoning before it answers.

1M token context by default; 128k maximum output
Always-on adaptive thinking, not a toggle
Claude API, Amazon Bedrock, Claude Platform on AWS, Google Cloud, Microsoft Foundry
$10 / $50 per million tokens, with cache hits at $0.25
A Covered Model: 30-day retention required, with a per-workspace override for ZDR organizations

Claude Mythos 5.1 Three framings, no lifecycle

The gated tier, at model id claude-mythos-5-1, sharing Fable 5.1’s specification and price. The access question is now framed three ways by three first-party sources, and none supersedes the others. The September 1 announcement names two programs. Of the Life Sciences Verification Program it says the LSVP “is designed so that life sciences professionals can use Claude Mythos 5.1 with safeguards designed for professional research and development activities.” Of the Cyber Verification Program it says the CVP “currently provides access to certain Opus- and Sonnet-class models with reduced cyber safeguards for defensive security work.” Read literally, that second sentence gates Opus- and Sonnet-class models, which is not the same thing as gating Mythos 5.1, so even a clean two-route reading of the announcement is shakier than it looks. The September 1 API release note still describes claude-mythos-5-1 as being for Project Glasswing participants. And on September 17, 2026 Anthropic gave the LSVP its own article with a grant structure: verification “includes a review of their research credentials, security standards, and ethical research oversight,” and there are two grant types, Standard Use renewed annually and High-risk Use renewed every six months, with Standard Use grants applying “to Mythos 5.1, Opus 5, and Sonnet 5 today, and to future models as they launch.” One more gap: neither claude-mythos-5-1 nor claude-mythos-5 appears anywhere in the deprecations table, so a limited-availability model is running with no published lifecycle at all.

Three first-party framings: the Cyber Verification Program, the Life Sciences Verification Program and Project Glasswing
The CVP as described gates Opus- and Sonnet-class models, not Mythos by name
LSVP grants: Standard Use renewed annually, High-risk Use renewed every six months
Same 1M context, 128k output and $10 / $50 standard pricing as Fable 5.1
No retirement floor published for either Mythos model
Not generally available, so it is not a procurement option today

Breaking change: tool_choice 5.1 returns 400

On Fable 5.1 and Mythos 5.1, the tool_choice types any and tool are not supported and return a 400. The types auto and none are unchanged. Anthropic’s stated alternatives are strict tool use or structured outputs. This is the breaking change most likely to be sitting in production code right now, because forcing a tool call has been a standard way to guarantee a parseable response. Search your codebase for the two removed values before you route any traffic to a 5.1 model.

any and tool return 400 on 5.1
auto and none behave as before
Use strict tool use or structured outputs instead
This fails the request; it does not degrade quietly

Breaking change: thinking blocks 5.1 returns 400

Thinking blocks are preserved only for the model that produced them or a newer one. For accounts created on or after August 31, 2026, replaying a thinking block after a changed system prompt, changed tools or a changed message returns a 400. The beta header thinking-binding-controls-2026-08-01 controls this behavior. The practical exposure is any pattern that caches an assistant turn and replays it against an edited prompt, which is common in agent frameworks and in evaluation harnesses that mutate the system prompt between runs. Note the account-age condition: an older account may not see the error, so a team can be split between working and failing on identical code.

Thinking blocks survive only into the producing model or newer
Accounts created on or after August 31, 2026 get a 400 on a changed system prompt, tools or message
Controlled by the beta header thinking-binding-controls-2026-08-01
Older and newer accounts can behave differently on the same code

Claude Opus 5 The no-retention tier

Released July 24, 2026 at $5 / $25 per million tokens with a $0.50 cache hit, 1M context and 128k maximum output. Its role has quietly sharpened with the 5.1 release: Opus 5 is now the most capable Claude model that does not carry a mandatory retention requirement, which makes it the natural home for regulated traffic that cannot sit under a 30-day window. Anthropic places it within 0.5% of Fable 5’s peak score on the benchmark it cites and describes fast mode as 2.5x faster at 2x base price. Both are vendor claims; time them on your own traffic. The July 24 announcement made Opus 5 the Claude Max default and the strongest model on Claude Pro. The second half of that has stopped being true. The pricing page now shows Fable available on Pro on usage credits, and Fable 5.1 sits above Opus 5, so Opus 5 is not the strongest model a Pro subscriber can reach. Per-plan availability by family is now published and readable; per-plan defaults at version granularity are still not. One more comparison a buyer should see: on the standard table Opus 5 and Opus 4.8 are priced identically, line for line.

$5 / $25 per million tokens on the standard tier; cache hit $0.50
No data-retention requirement, unlike the four Covered Models
Priced identically to Opus 4.8 across every column
No longer the strongest model on Pro, because Pro now reaches Fable on usage credits
In Claude Code, Claude Cowork and on Microsoft Foundry

Claude Fable 5 Legacy in marketing, Active in lifecycle

Fable 5 has not been retired and has not been repriced: $10 / $50 per million tokens on the standard tier with a $1 cache hit. What has changed is who is editorializing. The pricing page used to read as the unopinionated inventory beside an opinionated models overview, and it has stopped being either unopinionated or silent. It now carries a heading reading “Legacy models,” and Fable 5 is the first entry under it. Meanwhile the deprecations page still shows Fable 5 as Active, with a tentative retirement of not sooner than June 9, 2027. So the position to give a buyer is three-part rather than two: legacy in marketing, active in lifecycle, still priced and still selectable. Neither page is wrong; they answer different questions, and a procurement policy that keys off the word “legacy” will read the marketing page and stop. Fable 5 is also one of the four Covered Models, designated on June 9, 2026, so it carries the same 30-day retention requirement as the 5.1 pair and the same per-workspace override. Both pages are living documents; check them on the day.

Still active and still priced at $10 / $50 on the standard tier
First entry under the “Legacy models” heading on the pricing page
Active on the deprecations page, not sooner than June 9, 2027
Cache hits cost $1, four times the 5.1 rate
A Covered Model since June 9, 2026, so retention-required

Claude Sonnet 5 Free and Pro

Anthropic’s most agentic Sonnet, at $2 / $10 per million tokens with a $0.20 cache hit. The pricing question that hung over it is now settled by the calendar rather than by a promise: the scheduled September 1 increase did not occur, and $2 / $10 is the standard price. It plans, uses browsers and terminals, and approaches Opus 4.8 performance on some tasks at a lower operating cost, per Anthropic.

$2 / $10 is the standard price, confirmed after September 1 passed; $1 / $5 on the batch tier
The pricing page lists Sonnet on Free and on every paid plan, by family rather than by version
The July 24 claim that it is the Free and Pro default is not restated at version granularity on any current page
Budget renewals and volume forecasts can rest on it

Claude Opus 4.8 Two pages, two labels

Opus 4.8’s legacy label is not on the page a buyer would check for it, and the page that used to carry one rendered its body in JavaScript, so a fetch came back empty. The models overview no longer has a Legacy heading at all: it returns a four-row comparison table of Fable 5.1, Opus 5, Sonnet 5 and Haiku 4.5, and mentions 4.8 only to say that anyone on Claude Opus 4.8 or earlier should read the migration guide to Claude Opus 5. A different first-party page carries the label plainly. Under “Legacy models” on the pricing page, Opus 4.8 sits at $5 / $25, alongside Sonnet 4.6 and Sonnet 4.5 at $3 / $15 and Opus 4.7, 4.6 and 4.5 at $5 / $25. And the deprecations page still shows Opus 4.8 as Active, not sooner than May 28, 2027. The honest three-way position: legacy in marketing, active in lifecycle, still priced, and priced identically to Opus 5 on the standard table at $5 base, $6.25 five-minute cache write, $10 one-hour cache write, $0.50 cache hit and $25 output. A buyer comparing Opus 4.8 with Opus 5 is not choosing between price tiers, and on the other point, Opus 4.8 is not described as a safety fallback anywhere readable.

Legacy on the pricing page, Active on the deprecations page, still priced
Identical standard pricing to Opus 5, column for column
The models overview no longer carries a Legacy heading at all
Anthropic’s figure: four times less likely than Opus 4.7 to let flaws in its own code pass, a vendor claim
Microsoft’s July 24, 2026 announcement says Opus 5 replaces it in the Microsoft 365 Copilot lineup

Where a blocked request goes Two categories, not one

It is tempting to say Fable falls back specifically to Opus 5, and re-read on September 18, 2026, nothing first-party supports that precision. Anthropic’s readable wording is plural and unnamed: “Our safeguards do, however, still redirect several kinds of dual-use cybersecurity tasks...to our Opus models. This includes penetration testing, exploit generation, and binary-based vulnerability scanning.” Say “an Opus model” in runbooks and user-facing notes, not a version number. There is a second routing category on the same page, and it is easy to read past: “Queries related to research and development in the life sciences will still be directed to our Opus models.” If your safety documentation describes routing as a cybersecurity mechanism, it is describing half of it. The same page also restates the safeguard figures for the 5.1 generation and scopes them more narrowly than the August 7 post did. Anthropic says its latest biology safeguards for Fable 5.1 and Fable 5 fire 85% less often for benign requests related to elementary biology and medical questions, and that Claude Code users can expect an average of around 60% fewer interventions per session from the cyber safeguards relative to the previous safeguards on Fable 5. Both are vendor figures. One capability change belongs on a security team’s desk rather than a safety team’s: Anthropic says Fable 5.1 can now be used to discover software vulnerabilities, though not to develop exploits for them. On the API, the fallbacks parameter accepts a "default" mode applying Anthropic’s recommended fallback per refusal category; server-side fallback is in beta.

The destination is “our Opus models,” plural and unnamed
Two routed categories: dual-use cybersecurity work and life sciences research and development
Named cyber examples: penetration testing, exploit generation, binary-based vulnerability scanning
Vendor figures: 85% fewer fires on benign elementary biology and medical questions for the 5.1 generation, and around 60% fewer Claude Code interventions per session
Anthropic says Fable 5.1 can discover software vulnerabilities but not develop exploits for them
The August 7 figure that under 5% of Fable 5 sessions triggered a fallback also remains a vendor figure
Repoint any runbook that names a specific fallback model

Claude in Chrome Pro feature and Compliance API surface

Claude in Chrome sits on both sides of the buying decision, and it is the surface most likely to be missing from an inventory built before it existed. On the pricing page it is a Pro plan feature, listed as “Claude in Chrome and Microsoft 365.” And as of the September 18, 2026 release note it is a Compliance API surface: local session transcript retrieval for Claude in Chrome, with a product_surface value of claude_in_chrome, in beta for Claude Enterprise organizations. Treat it as its own environment rather than as a feature of chat. A browser agent has a different risk shape from a terminal agent or a governed desktop workspace: the untrusted content is the page itself, the credentials are already in the session, and the blast radius is whatever the signed-in browser can reach. The release-note half of this comes from a rolling page; confirm it directly before a retention process depends on it.

A Pro plan feature on the pricing page
Compliance API transcript retrieval in beta for Claude Enterprise, per the September 18 release note
Page content is untrusted input, and the browser already holds the session
Give it its own row in your environment and boundary inventory

Claude Code enterprise controls Changelog, since September 3

This is the cluster of controls that maps most directly onto what procurement actually asks, and it is published where procurement never looks. Anthropic’s Claude Code changelog has shipped a run of managed-settings controls. Organizations can restrict which models Claude Code uses through managed settings, with unauthorized model switches declined and an admin-enablement message shown to the user rather than a silent failure. A new maxEffortLevel managed setting caps reasoning effort across Bedrock, Vertex and Foundry, which is a spend control as much as a behavior control. Unparseable managed settings now fail startup and name the offending setting, closing a real gap in which a policy looked applied and was not enforced at all. Alongside those: managedMcpServers distributes HTTP and SSE MCP servers organization-wide, permissions.blockReadsOutsideWorkingDirectories arrives with a one-time prompt before the first read outside the working directories, plugins can no longer escape their directory through symlinked component paths, and --permission-prompts none supports unattended headless hosts. Fable 5.1 is now the default Fable model in Claude Code. Read the first three of these beside the agent governance card: model restriction, effort caps and policy-lock enforcement are the “what can it spend, what steers it” questions in concrete form. Every item here comes from the changelog, which is rewritten in place and cannot anchor a dated claim, so confirm each against your own installed version before it enters a control description.

Model restriction through managed settings, with unauthorized switches declined
maxEffortLevel caps reasoning effort across Bedrock, Vertex and Foundry
Unparseable managed settings now fail startup by name instead of going unenforced
managedMcpServers and permissions.blockReadsOutsideWorkingDirectories
Changelog only: confirm against your installed version

Agent governance since September 3 Release notes and the ant CLI

Three additions an agent architect should read, all from rolling pages. Managed Agents permission policies gained an auto mode in which, in Anthropic’s words, “the server evaluates each agent or MCP tool call and runs it, denies it, or pauses for your approval,” with a new evaluation field on the agent.tool_use and agent.mcp_tool_use events so the decision is visible in the event stream rather than inferred after the fact. The ant CLI 1.30.0 added ant apply, which declares agents, environments, skills, memory stores and deployments as files with a committed claude-lock.json lockfile. That is infrastructure as code for an agent estate, and it is the natural companion to Skills loading from GitHub: reusable behavior becomes reviewable in source control, and now so does the estate it runs in. The ant CLI 1.32.0 added ant beta:sessions connect, which attaches to a running agent session so an operator can message it and approve or deny in flight rather than waiting for it to finish being wrong. None of these is anchored to a dated article; confirm all three directly.

auto policies evaluate each tool call and can pause for approval
An evaluation field on agent tool-use events makes the decision auditable
ant apply with a committed claude-lock.json lockfile
ant beta:sessions connect for live attach, message and approve or deny

On-demand compaction September 14 beta

The Messages API can now return a signed compaction block summarizing the messages you sent, behind the beta compact-2026-09-04. You replay that block in place of the messages it covers, recent turns are retained verbatim, and on models that preserve thinking, that thinking stays valid. It lands on two things at once. A 1M-token window is capacity, not recall, and compaction is Anthropic’s own admission that a long window is a thing to manage rather than a thing to fill. And the cache-hit column is where the money is, which means compaction cuts both ways: it shortens the prompt and it changes the prefix, so a workload tuned around a stable cached prefix should measure the net effect before assuming a saving. It is a beta on a rolling release note. Pin it, regression-test it, and confirm it directly.

Signed compaction block, replayable in place of the messages it summarizes
Recent turns retained verbatim; preserved thinking stays valid
Beta header compact-2026-09-04
Compaction changes the prefix, so re-measure your cache-hit ratio

Workbench is Playground, with dates Aug 17 and Aug 18

The rename is now dated on both ends. The legacy Workbench at platform.claude.com/workbench sunset on August 17, 2026, and the August 18, 2026 release note names the replacement at platform.claude.com/playground. Playground supports every Messages API parameter and shows the full SDK request beside the API response, so it is a capability gain rather than a like-for-like port. It is also a rename your onboarding material, internal wiki links and screenshots will not have caught. Both dates come from Anthropic’s rolling release notes, which are rewritten in place, so confirm them directly.

Old URL sunset August 17, 2026
New URL documented in the August 18, 2026 release note
Release-note only: confirm directly
Sweep internal links and training material

Python SDK v1.0 is a breaking release August 20

Anthropic shipped Python SDK v1.0 on August 20, 2026 with four changes that will stop a build rather than warn it: the HTTP layer moved from httpx to httpx2, Python 3.10 or later is now required, Text Completions was removed outright, and temperature, top_p and top_k were removed from Messages. Pin your SDK version and read the migration before an unattended upgrade does it for you. This item rests on a rolling release note, which Anthropic rewrites in place, so confirm it directly.

httpx to httpx2
Python 3.10 or later required
Text Completions removed
temperature, top_p and top_k removed from Messages
Release-note only: confirm directly

Compliance and Admin API mature, and grow a header requirement August 26 to September 18

On August 26, 2026 the Compliance API session endpoints left beta for Claude Cowork and Claude Code, and the Admin API became available in the ant CLI and across all SDKs. Two later changes belong beside that. The September 1 release notes state that the Admin API Claude Enterprise endpoints, the Analytics API and the Compliance API guides now show the anthropic-version header and instruct you to send it on every request to these endpoints. An integration built against the August 26 item may not be sending it, and that is the kind of omission that works until it abruptly does not. And on September 18 the Compliance API gained local session transcript retrieval for Claude in Chrome, with a product_surface value of claude_in_chrome, in beta for Claude Enterprise organizations. All three items live on Anthropic’s rolling release notes, which are rewritten in place, so confirm each directly before a records process depends on it.

Compliance API session endpoints out of beta for Cowork and Claude Code
Send anthropic-version on every Admin, Analytics and Compliance API request
September 18: Claude in Chrome transcripts, beta for Claude Enterprise
Re-test coverage before it carries a legal retention obligation

Console keys, Files and Skills August 27

On August 27, 2026 the Console gained personal keys and service account keys, and the Files API and Skills API exited beta across the SDKs. Service account keys are the practical unlock for anyone who has been running production traffic on a named employee’s key and knew it was wrong.

Personal keys and service account keys in Console
Files API out of beta across SDKs
Skills API out of beta across SDKs

Model Hardware Standard August 27 research preview

Anthropic published “Previewing the Model Hardware Standard” on August 27, 2026. It is a research preview rather than a shipped requirement, so nothing in your deployment changes today, but it is the kind of document an infrastructure or procurement team should read early rather than after it becomes a standard.

A research preview, not a commitment
Relevant to infrastructure and hardware planning
Read it before it turns into a purchasing constraint

New betas worth knowing API

Three beta surfaces landed in this window. Per-message effort, behind the header mid-conversation-output-config-2026-07-01, is available on Fable 5.1, Mythos 5.1 and Opus 5, and lets a single conversation move up and down the effort ladder instead of committing at the first turn. Turn-scoped system messages narrow a system instruction to one turn. And thinking.display: "updates" changes how reasoning is surfaced. All three are beta: regression-test before any of them sits in a critical path.

Per-message effort on Fable 5.1, Mythos 5.1 and Opus 5
Turn-scoped system messages
thinking.display: "updates"
Beta surfaces move; pin and re-test

Claude for Teachers Confirm the window

Verified US K-12 educators get free premium access, and the sign-up window was open through June 30, 2027 when the page was last read. Anthropic can move a date like that without an announcement, so check it before quoting it to anyone.

Announced July 14, 2026
Verification required
Confirm the sign-up window on Anthropic’s page
A recruiting channel as much as a product

Self-serve HIPAA Confirm the scope

HIPAA configuration became self-serve for Enterprise and Platform customers on July 14, 2026, removing a manual step from healthcare deployments. Read it beside the Covered Model retention requirement, the per-workspace override and the platform split on where retained data sits before assuming the two compose. Self-serve is a process change, not a scope change.

Configure without a bespoke process
Still requires your own compliance review
Read it beside the Covered Model retention requirement
Confirm scope before moving regulated data

Claude Tag Team beta

A shared Claude teammate inside selected Slack channels, with channel-scoped context, tools, memories, budgets and audit logs.

Team and Enterprise beta
Shared channel continuity
Administrative spend and permission controls

Claude Science Specialist workbench

A customizable scientific environment with tools, packages, compute and auditable artifacts. The Life Sciences Verification Program, given its own dated article on September 17, 2026, sits next to it as an access route to Mythos 5.1 for verified life scientists, with grants reviewed against research credentials, security standards and ethical research oversight.

Research-oriented execution
Transparent artifacts
LSVP verification is a credential and governance review, not a purchase
A sign of domain-specific workbenches to come

Agent governance controls August 7

A cluster of controls shipped on August 7, 2026: session budgets for Claude Managed Agents, advisor configuration, inference geo controls and Skills loading from GitHub. Read together, they answer the three questions procurement asks about agents: what can it spend, what steers it, and where does the inference happen.

Session budgets cap agent spend rather than account spend
Inference geo controls matter for data-residency commitments
Skills loading from GitHub makes reusable behavior reviewable in source control

The text watermark, explained August 14

Anthropic published “How Claude’s text watermark works” on August 14, 2026. If your organization has a policy on AI-generated text disclosure, or contracts that turn on provenance, this is the document to read before the policy is drafted rather than after.

Relevant to disclosure and provenance policy
A vendor description of its own mechanism, not an independent audit
Test what your own downstream tooling preserves

Read the price column, not the model name

A buyer reading the lineup cold will assume Fable is the mid-tier and Opus the premium one, because that is how the names have worked elsewhere. It is the other way round. Fable 5.1 costs $10 / $50 per million tokens, Opus 5 costs $5 / $25, Sonnet 5 costs $2 / $10 and Haiku 4.5 costs $1 / $5. Add the cache column and a second ordering appears: Fable 5.1 reads cached context at $0.25, half what Opus 5 charges at $0.50 and only five cents above Sonnet 5 at $0.20, even though its base input costs five times Sonnet 5’s. For a cache-heavy workload the effective cost gap between the tiers is far narrower than the headline suggests. One more inversion sits inside the lineup itself: Opus 4.8, which the pricing page files under “Legacy models,” costs exactly what Opus 5 costs, $5 / $25 with a $0.50 cache hit, so the word “legacy” buys you nothing here. And every figure in this paragraph is the standard tier; the batch tier halves both input and output. Assign tiers by measuring your own traffic mix, not by reading the names.

The retirement calendar, re-read on September 18, and the model with no calendar at all

No deprecated Claude model is dated to retire before November 30, 2026, which is true and useless as a planning statement. Read on September 18, 2026, three Active snapshots carry tentative retirement dates inside that window, each phrased as a floor rather than a fixed date: claude-sonnet-4-5-20250929, “Not sooner than September 29, 2026”; claude-haiku-4-5-20251001, “Not sooner than October 15, 2026”; and claude-opus-4-5-20251101, “Not sooner than November 24, 2026.” The first of those is now eleven days out with no deprecation notice attached and no firm date, which is worth watching precisely because a floor that arrives unexercised teaches a planning team the wrong lesson. Further out, Fable 5.1 shows “Not sooner than September 1, 2027,” Fable 5 “Not sooner than June 9, 2027” and Opus 4.8 “Not sooner than May 28, 2027,” all with status Active. The gap is the interesting part: neither claude-mythos-5-1 nor claude-mythos-5 appears anywhere in the deprecations table, so the most restricted model in the lineup is also the only one running with no published lifecycle. The one Mythos line in the table concerns Claude Mythos Preview, which is deprecated with a migration pointer to Mythos 5. Treat a floor as the earliest date you could be forced to move, treat an absent row as no commitment at all, and read the deprecations page yourself: it is a living reference, not a dated article, and no summary of it can stand in for it.

Three facts procurement will ask about

First, retention, which is the fact most often stated wrongly. Fable 5.1, Mythos 5.1, Fable 5 and Mythos 5 are the four Covered Models and all four require 30-day data retention. Anthropic says the data is not used for training and is retained for at least 30 days before automatic deletion, with carve-outs for a safety investigation and for legal requirements. The retention itself is mandatory, but it is not an all-or-nothing gate on a zero data retention organization: Anthropic documents enabling 30-day retention for a single Console workspace while every other workspace keeps zero data retention. Opus 5 and below carry no requirement at all, which is still the contrast that decides tiering for regulated work, and Anthropic has separately published an interim zero data retention offer for eligible customers and the EFS program as the intended permanent answer, without publishing the eligibility criteria for either. Second, continuity. Access to Fable 5 and Mythos 5 was suspended on June 12, 2026, the redeployment notice was published June 30 and redeployment began July 1. Build your plan on the assumption that a top tier can be withdrawn. Third, defaults. The July 24 announcement made Opus 5 the Claude Max default and Sonnet 5 the Free and Pro default. Per-plan availability is now published and readable by model family, but no current page restates a default at version granularity, and anthropic.com/pricing still cannot be fetched at all, returning a 406, while claude.com/pricing renders. Pin a model id rather than relying on a plan default nobody has restated since July.

What changed between September 3 and September 18, 2026

The model and pricing spine did not move. No Claude model retired, every standard price in the table above still matches Anthropic’s pricing documentation, Sonnet 5 is still $2 / $10, and the softened safety-routing language still holds. What moved sits around the edges, and most of it is retention. Zero data retention is not a hard gate on 5.1, because Anthropic documents a per-workspace override; “deleted after 30 days” is really “at least 30 days” with a safety-investigation carve-out and a legal carve-out; Fable 5 and Opus 4.8 are filed under “Legacy models” on the pricing page while the deprecations page still lists both as Active; and per-plan model availability is published in detail at claude.com/pricing even though anthropic.com/pricing cannot be fetched at all. Two things Anthropic published on September 1, on the same day as the launch and largely underneath it, belong in the file: Enterprise Frontier Safeguards and an interim zero data retention offer for eligible customers. Newer still: a September 10 threat intelligence report, a September 14 on-demand compaction beta, a September 17 Life Sciences Verification Program article with a grant structure, a September 18 Compliance API surface for Claude in Chrome, and a run of Claude Code managed-settings controls covering model restriction, effort caps and policy-lock enforcement. The action that follows is narrow: take the two or three sentences about Claude retention already sitting in your procurement file and check each one against Anthropic’s retention page yourself.

3 / Product Family

Claude is now a family of work surfaces

Choosing Claude is no longer just choosing a model. The decisive question is where the work should run: conversation, terminal, governed desktop workspace, team channel, design and document surface, the browser, or a business application. The browser is the newest of those and the one most likely to be missing from an existing inventory.

Analyze and write

Claude chat with Sonnet 5, Opus 5 or Fable 5.1

Use the main Claude experience for source-grounded analysis, research, writing and ready-to-use files. The pricing page puts Sonnet and Haiku on Free and shows Fable there as “No”; Pro adds Opus and reaches Fable on usage credits, which is why Opus 5 is no longer the strongest model a Pro subscriber can call; Max buys 5x or 20x more usage than Pro. Fable 5.1 is the most capable tier, and it is a Covered Model, so 30-day retention comes with it.

Attach the complete source pack
Choose effort by consequence
Check the workspace retention setting before you pick 5.1

Surface	Best for	Context advantage	Primary caution
Claude chat	Analysis, writing, research and files	Conversation, uploads, connectors and memory	Large context still needs coverage checks
Claude Code	Repository-scale engineering	Codebase, terminal, tools and project instructions	Command and prompt-injection risk
Claude Cowork	Desktop knowledge work	Selected folders, local apps, skills and plugins	Mount permissions and action review
Claude Tag	Shared team delegation	Channel history, scoped tools and shared memory	Spend, permissions and public mistakes
Claude in Chrome	Work that lives in a web application	The open page, the signed-in session and the browser’s own reach	Page content is untrusted input; transcript retrieval is beta
Claude Design	Rapid visual exploration	Brand inputs, conversation and prototypes	Still a Labs product; professional finishing required
Claude Slides and Claude Docs	Decks and written deliverables	The analysis already in the conversation	A bundled Pro feature, not a separately governed surface
Claude Science	Auditable scientific work	Packages, tools, compute and artifacts	Expert validation remains essential

4 / Inside Work

From source material to finished work without changing context

Anthropic’s direction is clear: Claude should follow the work across files, applications and teams instead of making users rebuild the brief in every surface.

Engineering

Claude Code

Plans, edits, tests and debugs inside the real repository.

Agent teams and subagents
Long-running workflows
Code review and verification

Knowledge work

Claude Cowork

Runs multi-step tasks against selected desktop files and applications.

Read-only or controlled write modes
Reusable skills and plugins
Dispatch from text or voice

Microsoft 365

Excel, PowerPoint and Word

Claude works directly in business documents and carries context between applications. Write tools shipped on July 7, 2026: Claude can draft and send email, act on the calendar and create or update files.

Model, audit and explain spreadsheets
Build decks from live analysis
Review write permissions before granting them

Collaboration

Claude Tag

A shared team agent in Slack with scoped identity and memory.

Channel-specific access
Shared continuation
Logs and budget controls

Browser

Claude in Chrome

Works inside the browser on web applications that never exposed an API. A Pro plan feature, and now a Compliance API surface in beta for Claude Enterprise.

Read and act on the open page
Page content is untrusted input
Scope the signed-in session deliberately

Creative

Claude Design, Slides and Docs

Explores prototypes and polished visual directions through conversation, and produces decks and documents in the same place. The pricing page bundles the three together as a Pro feature.

Software experience ideation
Slides, docs and one-pagers
Canva handoff

Connectors

MCP and apps

Pulls live context from enterprise and creative systems.

Google Drive and Gmail
Market and portfolio data
Blender and specialist tools

Reusable behavior

Skills and plugins

Packages instructions, tools and workflows into repeatable organizational capability. The Skills API left beta across the SDKs on August 27, 2026.

Reduce prompt folklore
Standardize recurring work
Deploy domain templates

Continuity

Memory and context

Carries relevant context across sessions and, increasingly, across applications. On July 10, 2026 memory was reworked into individual categorized entries, replacing the daily memory summary.

Preserve audience and standards
Inspect and remove entries one at a time
Review what should persist

5 / Creative Work

Claude is becoming useful before, during and after the creative tool

Claude does not replace taste. Its advantage is expanding the number of directions a creative team can explore, automating repetitive production and keeping strategy connected to the artifact.

Claude Design Labs

Explore software experiences, prototypes, slides and one-pagers with visual iteration inside the conversation.

Start with audience and desired behavior
Include brand rules and references
Export into professional finishing tools

Creative connectors MCP

Claude can work with tools such as Blender through connectors, scripts and the application’s own APIs.

Analyze and debug scenes
Batch repetitive changes
Create custom tools inside the creative application

Cross-format production

Claude can turn research into a document, document into deck, data into charts and creative brief into a buildable prototype.

Keep one narrative thread
Reduce manual reformatting
Use humans for taste and final selection

Prototype brief

Build three distinct directions for this onboarding flow. Use our brand system, explain the behavioral idea behind each route, and make every prototype testable rather than decorative.

A strong brief defines the decision the prototype should help the team make.

Production automation

Inspect this Blender scene and identify the repetitive production steps. Propose a reversible automation plan, then implement only the steps I approve.

Keep approval between analysis and modification.

Deck from evidence

Turn this research and workbook into a 12-slide executive deck. Use the approved master, keep one claim per slide, and link every chart to its source.

Ask Claude to maintain provenance, not merely visual polish.

Taste-preserving review

Do not redesign this. Review it as a creative director: identify where the execution departs from the stated concept, audience or brand behavior.

Claude is often more useful as a critic than as an automatic stylist.

6 / Behavior Playbook

Eight habits that make Claude substantially better

The model upgrade matters, but workplace behavior determines whether that capability becomes useful, expensive or risky.

Front-load the context

Give Claude the complete source pack and explain how the files relate before asking for conclusions.

Let it inspect the brief

Ask what is missing or inconsistent before it begins production.

Choose effort deliberately

Use lower effort for reversible work and escalate only when consequence justifies it. On 5.1, adaptive thinking is always on, so effort is the only lever left.

Define completion

Specify the artifact, validation criteria and evidence report, not only the topic.

Keep context across stages

Move from analysis to file to deck without rebuilding the narrative in a new chat.

Use an adversarial second pass

Ask Claude to identify the weakest claims, broken assumptions and missing tests in its own work.

Name the audience and stakes

Claude changes depth, tone and caution when it knows who will use the result and what failure costs.

Trust the loop, inspect the result

Avoid micromanaging every step, but review commands, diffs, formulas and finished artifacts.

Outcome delegation

Produce the completed board memo, not an outline. Validate every recommendation against the attached evidence and include a one-page appendix of unresolved assumptions.

Agentic coding

Add pagination to the users API, update the contract and documentation, add edge-case tests, run the full suite, and stop only when CI passes or you can name the blocker precisely.

Self-review

Re-read your work as a skeptical domain expert. Identify the three conclusions most likely to be wrong and show what evidence would change each one.

Long-context coverage

Before synthesizing, create a coverage map of every supplied document. After the answer, list any sources or sections you did not materially use.

7 / Watch Outs

The more Claude can act, the more its environment matters

Agent capability changes the failure mode. A weak chat answer is inconvenient; a tool-using agent with broad file and network access can create real operational damage. Two further categories belong here: a model generation that fails requests your current code makes legally today, and a set of first-party contradictions a control document can inherit without anyone noticing.

Two breaking changes on 5.1, and a third 400 that is a configuration gate

Moving traffic to claude-fable-5-1 or claude-mythos-5-1 can break working code in two ways. tool_choice types any and tool return a 400. And replaying a thinking block after a changed system prompt, tools or message returns a 400 on accounts created on or after August 31, 2026, controlled by the thinking-binding-controls-2026-08-01 beta header. A third failure path is counted separately here, because it is not a code change at all: both models require 30-day retention, so a request from an organization or workspace that does not have retention enabled returns a 400 invalid_request_error rather than a degraded answer. That one is a configuration problem with a documented fix, not a wall: see the workspace override above. Test all three failure paths before you switch a model id in production.

Zero data retention is not a hard gate on 5.1

The 30-day requirement on Fable 5.1 and Mythos 5.1 reads like a wall and is not one. A zero data retention organization can enable 30-day retention for one Console workspace, under Settings then Workspaces and the workspace’s Privacy controls tab, while every other workspace keeps zero data retention. If a procurement file anywhere says that no configuration flag fixes this, that is the sentence to go and fix first. The requirement itself is real and the failure mode is a clean 400, so the work is a configuration decision and a scoping decision, not an exception request.

Nobody publishes who is eligible for interim zero data retention

Anthropic’s September 1 news pages say eligible customers will be able to use Fable 5.1, and Fable 5, with zero data retention until Enterprise Frontier Safeguards is ready. The September 1 API release note and the retention documentation both say the models are not available under zero data retention unless expressly authorized by Anthropic. No page defines eligibility, and no page says whether eligible and expressly authorized are the same status. This is an unresolved conflict between first-party sources rather than a stale claim, and it is published here as one. Do not let either sentence into a control document on its own.

The retention answer changes with the route you buy through

Read this beside the two cards above, because it is a different kind of conflict. Those are Anthropic pages disagreeing with each other. This is Anthropic’s direct terms and Microsoft’s subprocessor terms describing the same model and not describing the same outcome. On Anthropic’s own pages, Fable 5.1 is a Covered Model with a 30-day retention requirement. On Microsoft’s living page “Anthropic models in Microsoft Online Services,” stamped September 10, 2026, the off-by-default retention gate attaches to Fable 5.0, and for certain organizations Fable 5.1 runs with Anthropic as a subprocessor under Microsoft’s Product Terms and Data Protection Addendum, under which Microsoft states that Anthropic does not retain customer content. Neither source resolves the other. Do not move Microsoft’s sentence into an Anthropic control description or the reverse, and do not let a procurement file inherit a retention posture from whichever page it read first. Establish the contracting route, then the retention answer. The same Microsoft page is where the two-year and seven-year retention tails above come from, which is a second reason to read it rather than summarize it.

EFS is a promise with no dated milestone

Enterprise Frontier Safeguards is described as giving privacy equivalent to zero data retention by keeping customer data on the customer’s own cloud infrastructure, across Claude Code, Claude Enterprise, the Claude Platform, Amazon Bedrock, the Claude Platform on AWS, Google’s Agent Platform and Microsoft Foundry. Anthropic says it is rolling out in phases starting this fall. Phases, not dates. A privacy posture that depends on EFS arriving is a plan with no delivery date in it, which is a different risk from a plan with a date that slips.

Prompt injection

Instructions inside email, documents, websites or pasted prompts can redirect an agent. Treat external content as untrusted input.

Mount and write access

Cowork permissions should match the task. Read-only, read-write and no-delete boundaries are meaningful controls.

Higher effort cost

More reasoning can improve hard work but adds latency, token use and the possibility of unnecessary exploration. On 5.1 you cannot turn thinking off at all.

Preview instability

Dynamic workflows, Labs products and beta surfaces require regression testing before production use. Per-message effort, turn-scoped system messages and the updates thinking display are all beta.

Long context omissions

A 1M-token window is capacity, not proof of complete recall or equal attention. That is now the default window on 5.1 rather than an opt-in, so the coverage discipline applies to every call.

No native media generation

Claude can plan, analyze and prototype visuals, but final image, video and audio generation still depends on other tools.

Model-version drift

A model upgrade can change formatting, tool choices and reasoning style. Re-test critical prompts and workflows.

Diplomatic language

Claude may soften criticism unless directness is explicitly requested. Ask for the harshest defensible review.

Opus 4.1 is gone, and its named replacement has itself moved on

On Anthropic’s deprecations page, checked again on September 18, 2026, claude-opus-4-1-20250805 still shows status Retired with the line “This model was retired August 5, 2026.” Anything still pinned to the 4.1 identifier, an integration, a saved evaluation or a scheduled job, is failing now as a request error rather than a warning. If you have not hit the error yet, that tells you the pinned job is not running, which is its own finding. One note on the destination: the deprecations row still names claude-opus-4-8 as the replacement, but Anthropic’s current migration pointer sends anyone on Claude Opus 4.8 or earlier to Claude Opus 5, and Opus 4.8 is itself now filed under “Legacy models” on the pricing page. Migrating 4.1 to 4.8 today buys you one hop into a bucket you will have to leave again.

An empty retirement calendar was never empty

The claim that no Claude retirement is dated before November 30, 2026 held only for deprecated models. Three Active models carry a tentative retirement floor inside that window: claude-sonnet-4-5-20250929 not sooner than September 29, 2026, claude-haiku-4-5-20251001 not sooner than October 15, 2026, and claude-opus-4-5-20251101 not sooner than November 24, 2026. A floor is not an announcement, but it is also not nothing, and reading the deprecations table as a list of dated retirements misses it entirely.

Python SDK v1.0 will stop a build

The August 20, 2026 release moves httpx to httpx2, requires Python 3.10 or later, removes Text Completions and removes temperature, top_p and top_k from Messages. An unpinned dependency will pick this up on its own schedule, which is rarely yours.

The Workbench URL is dead

The legacy Workbench at platform.claude.com/workbench sunset on August 17, 2026 and the replacement Playground at platform.claude.com/playground was documented in the August 18, 2026 release note. Three experimental /v1/experimental/*_prompt endpoints retired on the same August 17 date. Internal links, onboarding docs and screenshots will still point at the old surface.

July 24 breaking changes still apply

Setting thinking to disabled is only valid at effort high or below; combining it with extra high or max returns a 400. Requesting fast speed on Opus 4.7 errors. Both break working integrations silently until a call fails.

Per-plan defaults still move, so pin a model id

The July 24 announcement put Opus 5 as the Claude Max default and left Sonnet 5 on Free and Pro. Neither is restated at version granularity on any current page, so keep pinning a model id in anything you evaluate or your baseline changes underneath the test. Read the hostname that answers. anthropic.com/pricing returns a 406 and cannot be fetched at all, while claude.com/pricing renders per-plan model availability in detail and files Opus 4.8 under “Legacy models” in plain text. A page that refuses a request is not evidence that the information is unpublished.

The safety fallback has no published name, and two categories rather than one

Anthropic says dual-use cybersecurity work is redirected “to our Opus models,” plural and unnamed, naming penetration testing, exploit generation and binary-based vulnerability scanning as examples. It also says queries related to research and development in the life sciences are directed to the Opus models. No first-party page names a specific fallback model. Two things follow. If a runbook, incident playbook or user-facing message in your product names one, it is repeating an inference rather than a documented fact. And if your safety documentation describes routing as a cybersecurity mechanism only, it will surprise a research team the first time a life sciences query behaves differently from the rest.

Stolen API keys and clients spoofing Claude Code

Anthropic’s September 10, 2026 threat intelligence report is a Watch Out rather than background reading. Treat an API key as a production credential, not a convenience: a stolen key lets an attacker run operations at your expense while the activity is attributed to you, which is both a cost exposure and an attribution problem in an incident review. The report also describes credential harvesting through fake Claude resellers and malicious clients that spoof “Claude Code.” That second one is a supply-chain warning about the tool developers install most casually, so it is worth naming in an internal advisory: install Claude Code from Anthropic’s own distribution and nowhere else.

Mythos has no published lifecycle at all

Neither claude-mythos-5-1 nor claude-mythos-5 appears anywhere in the deprecations table. Every other model in the lineup carries at least a tentative floor. A limited-availability model with no published retirement floor is not a scheduling risk today, because you probably cannot get it, but it is a governance gap for anyone who does: there is no date to plan a migration against and no page that would tell you one had been set.

The anthropic-version header on Enterprise endpoints

The September 1 release notes state that the Admin API Claude Enterprise endpoints, the Analytics API and the Compliance API guides now show the anthropic-version header and instruct you to send it on every request to these endpoints. An integration built against the August 26 Compliance API release may not be sending it. This is a rolling release-note item: check what your client actually sends rather than what your notes say it sends.

Four items sit on pages Anthropic rewrites in place

Claude for Teachers and its June 30, 2027 sign-up window, the self-serve HIPAA item, the August 17 Workbench sunset with the August 18 Playground rename, and the Python SDK v1.0 release all rest on rolling release notes or living pages rather than on dated articles. A rolling page keeps no history, so a date read from one can be wrong the next morning with nothing on the page to show that it changed. If one of them is load-bearing for you, that is the one to check first.

Vendor benchmark claims

The positioning of Fable 5.1 and Mythos 5.1 as the world’s most advanced models for coding and knowledge work is Anthropic’s own, as is the 0.5% gap between Opus 5 and Fable 5. Run a task-level evaluation before moving production work up or down a price tier on the strength of it.

Environment	Useful boundary	What still requires review
Claude chat	Ephemeral server-side container	Files produced, connector scope and claims
Claude Code	Local sandbox plus human command approvals	Shell commands, diffs, secrets and outbound access
Claude Cowork	Sealed VM and selected folder mounts	Mounted workspace, MCP servers and completed actions
Claude in Chrome	The browser tab and the signed-in session it already holds	Page content as untrusted input, form submissions, and what the session can reach
Claude Tag	Channel-scoped tools, memory and spend	Public responses, tool actions and cross-team access

8 / Sources

First-party evidence behind this guide

Product status, availability and pricing can change quickly. These are the primary Anthropic pages checked for this edition, verified September 18, 2026, with seven caveats about coverage. First, the whole retention position, including the per-workspace 30-day toggle in Console, rests on Anthropic’s living retention documentation and on the Covered Models support article. Neither carries a publication date, so neither is listed below as a source, and both should be read directly before a control description quotes them. Second, several other load-bearing facts rest on living pages rather than dated articles and are attributed inline rather than cited: the plan and seat structure, the “Legacy models” heading and every legacy price, which come from the Claude pricing page; every deprecation status, retirement floor and the absence of any Mythos row, which come from the model deprecations table; and the four-model comparison and the absent Legacy heading, which come from the models overview. Third, a large group of operational items comes from Anthropic’s rolling API release notes: the September 3, September 10, September 14 and September 18 items, the anthropic-version header requirement, the on-demand compaction beta, the Managed Agents auto permission policies, the Compliance API surface for Claude in Chrome, the August 17 Workbench sunset and August 18 Playground rename, the August 20 Python SDK v1.0 release, the August 26 and August 27 API changes, the per-message effort, turn-scoped system message and thinking-display betas, and the Sonnet 5 pricing confirmation. Every Claude Code item, including model restriction through managed settings, maxEffortLevel, managedMcpServers and the startup enforcement of unparseable settings, comes from the Claude Code changelog, as do the ant CLI items. Release notes and changelogs are rewritten in place and cannot anchor a dated claim, so they are not listed below and a reader relying on any of them should confirm it directly. Fourth, the three Anthropic sources describing access to Mythos 5.1 frame it three different ways, through the Cyber Verification Program, the Life Sciences Verification Program and Project Glasswing, and all three are reported above rather than one being chosen. Fifth, Anthropic’s news pages and its API release notes disagree about whether eligible customers can use Fable 5.1 under zero data retention in the interim, and Anthropic does not publish the eligibility criteria; that disagreement is printed above rather than resolved. Sixth, four items rest on rolling pages rather than dated articles and are labeled as such where they appear: Claude for Teachers, the self-serve HIPAA item, the Workbench sunset and Playground rename, and Python SDK v1.0. Seventh, one comparison does not rest on an Anthropic page at all. The Microsoft subprocessor terms for Fable 5.1, and the two-year and seven-year retention tails attributed to Anthropic, were read on Microsoft’s living page “Anthropic models in Microsoft Online Services,” stamped September 10, 2026, reached through the sibling Microsoft 365 Copilot guide in this set. That page is rewritten in place, and it is Microsoft describing Anthropic rather than Anthropic describing itself, so it is attributed inline rather than listed below. Confirm it directly, and confirm the two durations against Anthropic’s own retention documentation, before either goes near a control document. The Frontier Red Team’s July 30, 2026 cybersecurity write-up and Anthropic’s September 17, 2026 policy paper on measuring the pace of AI development carry no pricing or model implications, so neither is cited here.

Claude Fable 5.1 and Claude Mythos 5.1Anthropic · September 1, 2026 · the launch, the 30-day retention requirement, the interim zero data retention line, the Cyber and Life Sciences Verification Programs, life sciences routing and the 5.1 safeguard figures Developing Enterprise Frontier Safeguards with our customersAnthropic · September 1, 2026 · EFS as privacy equivalent to zero data retention with customer-hosted data, its phased rollout starting this fall, its platform scope, and the interim zero data retention offer for eligible customers on Fable 5 and Fable 5.1 Introducing the Life Sciences Verification ProgramAnthropic · September 17, 2026 · the LSVP grant structure, Standard Use renewed annually and High-risk Use renewed every six months, and the credential, security and ethical oversight review Detecting and countering misuse of AI: September 2026Anthropic · September 10, 2026 · stolen API keys used to run operations at a victim’s expense while evading attribution, credential harvesting through fake Claude resellers, and malicious clients spoofing Claude Code Improving our alignment and security effortsAnthropic · August 31, 2026 · published the day before the 5.1 launch and easy to miss beside it; context for the safeguard changes that shipped with the 5.1 generation Expanding our support for scientistsAnthropic · August 27, 2026 · the scientific program background that sits behind Claude Science and the Life Sciences Verification Program Previewing the Model Hardware StandardAnthropic · August 27, 2026 · research preview Introducing Claude Opus 5Anthropic · July 24, 2026 · pricing, 1M context, fast mode, plan defaults and the Fable 5 comparison Claude Fable 5 and Claude Mythos 5Anthropic · June 9, 2026 · the previous top tier, pricing, safeguards and retention policy Redeploying Claude Fable 5Anthropic · June 30, 2026 · notice of restored access after the June 12 suspension; redeployment began July 1 Improving Fable 5’s biology safeguardsAnthropic · August 7, 2026 · dual-use work redirects to “our Opus models,” plural and unnamed. The roughly 85 percent figure is restated by the September 1 page for the 5.1 generation and scoped there to benign requests about elementary biology and medical questions; both remain vendor figures How Claude’s text watermark worksAnthropic · August 14, 2026 · how the Claude text watermark works Fable safeguards and the jailbreak severity frameworkAnthropic · July 2, 2026 · the Cyber Jailbreak Severity scale, CJS-0 to CJS-4, developed with Amazon, Microsoft and Google Claude for TeachersAnthropic · July 14, 2026 · free premium access for verified US K-12 educators, sign-up open through June 30, 2027, on a page Anthropic can change without notice Introducing Claude Sonnet 5Anthropic · June 30, 2026 Introducing Claude Opus 4.8Anthropic · May 28, 2026 Introducing Claude TagAnthropic · June 23, 2026 Introducing Claude Design by Anthropic LabsAnthropic · April 17, 2026 Claude for Creative WorkAnthropic · April 28, 2026 Agents for financial servicesAnthropic · May 5, 2026 How we contain Claude across productsAnthropic Engineering · June 2026 Claude can now create and use filesAnthropic · September 9, 2025

Evidence note

The workplace positioning and behavior guidance are AI Mindset interpretation. Model names, dates, availability, prices and product capabilities are drawn from the linked first-party material, or, where the fact lives only on a living reference page, are attributed inline with an instruction to confirm directly. Where Anthropic’s own pages disagree, as they currently do on how Mythos 5.1 access is gated and on whether eligible customers can use Fable 5.1 under zero data retention, the disagreement is reported above rather than resolved for you. The same treatment is given to a disagreement that is not Anthropic’s alone: Anthropic’s direct terms and the subprocessor terms Microsoft publishes for Fable 5.1 do not describe the same retention outcome, and the route you buy through changes the answer, which is recorded rather than settled in favor of the version that reads better.

AI Mindset