ChatGPT: The Work Operating System You Can Talk To
The Work Operating System You Can Talk To
ChatGPT now separates work by intent: Chat for questions and conversation, Work for longer tasks and finished deliverables, and Codex for software development. The shift is from an assistant that answers to a work environment that can research, create, steer, schedule and execute. On September 3, 2026 OpenAI launched GPT-6 Astra, a new model family and the new flagship, to a limited set of organizations rather than to everyone.
Three modes for three different kinds of work
Chat handles questions and conversation. Work carries longer jobs into finished deliverables. Codex owns software development. They share context and tools, but each mode sets a different expectation for autonomy and output. On July 16, 2026 the desktop app was reorganized around that split: the top-left menu is now a global switcher between ChatGPT and Codex, and a toggle inside ChatGPT moves between Chat and Work.
Fast, adaptive and willing to start
ChatGPT is unusually good at turning an incomplete request into momentum. It can ask questions, propose a structure and make a first version quickly.
- Broad multimodal capability
- Strong iterative conversation
- Useful from blank page to final artifact
Cross-functional knowledge work
Research, planning, writing, image creation, file analysis, data explanation and coding can remain connected inside one working thread.
- Executive briefs and decisions
- Research with current sources
- From prototype to production code
The work may change shape
Start in Chat to clarify, move into Work for the finished artifact, or hand the software outcome to Codex without rebuilding the whole brief.
- The brief is still ambiguous
- Several output types may be needed
- The work benefits from shared context
Work turns a conversation into an accountable production run
Launched July 9, 2026 and powered by GPT-5.6, Work is the mode for longer tasks that should end in something usable. It can research and analyze, use connected apps and files, build finished artifacts, show progress, accept redirection and pause for important approvals. OpenAI describes it as staying with a project for hours. GPT-6 Astra began reaching ChatGPT and Codex on September 3, 2026, but only for a limited set of organizations, so for most teams Work still runs on GPT-5.6 today.
Research and analysis
Give Work the decision or finished state, not just a question. It can gather current evidence, inspect files and keep a longer analytical process moving.
- Use connected apps and uploaded files
- Track progress instead of waiting blindly
- Require source and uncertainty reporting
Finished deliverables
Work can create the actual business object: a document, spreadsheet, presentation, report or other reusable output.
- Specify audience and acceptance criteria
- Review the artifact, not only the summary
- Keep important calculations visible
Sites public beta
Sites extends the output from a static answer into dashboards, trackers, launch calendars, prototypes, internal portals and reports. The help article now describes the public beta as covering all plans except Free and Go, which brings Plus and Business into scope alongside the original four.
- Not available in the EEA, Switzerland or the United Kingdom
- The help article carries no date, so read the wider eligibility as current state rather than a dated change
- Treat live information and access as product requirements
Scheduled Tasks
A task can run once, repeat on a schedule, respond to a trigger or monitor something over time.
- Define the trigger precisely
- Name the notification and escalation path
- Recheck permissions for recurring work
Progress and steering
You can follow the work, answer questions and change direction without restarting the task or losing the developing context.
- Set review points early
- Redirect when assumptions change
- Do not mistake visible activity for quality
Approval moments
Work can ask for approval before important actions, making authority boundaries a designed part of the workflow.
- State which actions always require approval
- Use least-privilege access
- Inspect the resulting external state
Local files, apps and browser
On desktop, Work can use local files and applications alongside a built-in browser. Computer Use lets it click, type and move files across your apps in the background. Since July 16 a unified Recents list merges Chat and Work, and Projects are available on desktop.
- Cloud Work conversations sync across web, mobile and desktop
- Local conversations stay local, so plan for where the record lives
- Grant only the context the task needs and review external transmissions
Unified plugin directory
Plugins connect ChatGPT to Slack, Teams, Drive, SharePoint, email, calendars, CRMs and internal tools. Type “@” to direct ChatGPT at a specific app; the new directory brings plugins into one place.
- Standardize repeatable workflows
- Review plugin authority and data access
- Prefer governed organizational plugins
Auto-review of important actions
Auto-review uses OpenAI’s most advanced models to review consequential actions involving connected tools and APIs before they happen. OpenAI reports that in its own adversarial red teaming the feature blocked 100% of attempts to extract protected data.
- Vendor red-team result, not an independent audit
- A control, not a guarantee
- Compliance API gives oversight at scale
- Admins set spend and access limits
Who has Work today
The July 9 release note names the first wave as Pro, Pro Lite, Enterprise and Edu, with Plus and Business following within days. Enterprise and Edu received a two-week preview, off by default, with an admin opt-out before automatic enablement. No dated article yet confirms that the Plus and Business rollout completed; the July 21 small business article says Work is available to small businesses today, which is indirect evidence that Business landed. Treat the first wave as dated fact and the rest as in progress, and check your own plan picker.
- Do not confuse the Work rollout with the Sites rollout
- The desktop app exposes the Chat, Work and Codex surfaces on every plan, including Free; which plans are entitled to use Work is the separate question above
- The old desktop app is renamed ChatGPT Classic
Use Work to produce the finished leadership briefing, not just an outline. Research current primary sources, build the supporting analysis and charts, create the presentation, and pause for approval before any publication or sharing.
After completing one verified manual run, schedule this monitor for every weekday. Notify me only when the threshold is crossed, include the evidence and timestamp, and never take an external action without approval.
Create an internal launch Site with the live plan, owners, milestones, decisions, risks and source documents. Use our brand system and make every status field auditable.
Pick the working mode first, then the tier, then the effort
GPT-6 Astra launched on September 3, 2026. It is a new model family rather than a point release or a rename, its API id is gpt-6-astra, and it is the new flagship. It is also not generally available: OpenAI is rolling it out to a limited set of organizations first, so for most teams the live comparison is still Sol, Terra and Luna. GPT-5.6 is not deprecated, not legacy and unchanged in status. Astra is a single tier, with no equivalent of Terra or Luna, and a separate GPT-6 Astra Pro exists for Pro, Business and Enterprise whose pricing is published nowhere. This edition also carries a correction that changes every number in the price tables. The September 3 edition printed OpenAI’s Batch rates in the position a reader would read as Standard, which understated the real cost of every GPT-5.6 model by exactly half. The tables below are rebuilt with Standard as the headline, Batch and Flex as the identically priced half-price tier, and Fast mode at exactly twice Standard. Sol’s Standard rate is still promotional, and OpenAI still commits to it only “at least through November 21, 2026.”
GPT-6 Astra New flagship · limited rollout
Launched September 3, 2026 as a new family, not a point release and not a rename. OpenAI’s model guidance calls it “our most intelligent model yet, with state-of-the-art performance in computer use, browsing, software engineering, science, and professional work,” and claims “Astra achieves stronger results while using substantially fewer output tokens, delivering a lower estimated API cost per task than earlier models.” Both are vendor claims. Astra is a single tier: there is no Astra equivalent of Terra or Luna. It is not generally available, and the help center warns that “Astra can use your allowance faster than GPT-5.6 Sol.”
- API id
gpt-6-astra; 1,050,000 token context, 128,000 max output, knowledge cutoff April 30, 2026 - Standard short context $10.00 input / $1.00 cached / $50.00 output per 1M tokens
- Text in and out; images input only, with no image output and no audio
- Reasoning effort
low,medium,high,xhighandmax; no fine-tuning - Rolling out to a limited set of organizations, so confirm your own access before planning
GPT-6 Astra Pro Price unpublished
A separate Astra Pro exists for Pro, Business and Enterprise plans. Its pricing is published nowhere: it has no row on the developer pricing page and no line on the enterprise rate card, which lists plain Astra only. Treat it as a sales conversation rather than something you can put in a budget.
- Pro, Business and Enterprise plans
- No published price anywhere in this reading
- Absent from the enterprise token rate card
GPT-5.6 Sol GPT-5.6 flagship
Astra did not retire it. Sol is not deprecated, not legacy and unchanged in status, and OpenAI still describes it as delivering state-of-the-art results across coding, knowledge work, cybersecurity and science. Its price is the one this guide got wrong: the real Standard rate at short context is $4.00 input and $20.00 output per 1M tokens, not the $2.00 and $10.00 printed on September 3, which were the Batch and Flex figures. The promotional footnote is unchanged and still live: “GPT-5.6 Sol’s promotional pricing is available at least through November 21, 2026.”
- Standard short context $4.00 input / $0.40 cached / $20.00 output; long context $8.00 / $0.80 / $30.00
- Batch and Flex halve that; Fast mode doubles it
- Promotional: guaranteed only through November 21, 2026, so diarize that date
- Sol Pro available to Pro and Enterprise in Chat
- The tier most teams can actually deploy while Astra is limited
GPT-5.6 Terra Balanced
A lower-cost tier that OpenAI positions as competitive with GPT-5.5. It was the model Free and Go users received inside Work and Codex until August 6, 2026, when OpenAI made Luna the default for those plans. Its Standard price is double what this guide printed on September 3, for the same reason Sol’s was: the last edition read the Batch column.
- Standard short context $2.00 input / $0.20 cached / $12.00 output; long context $4.00 / $0.40 / $18.00
- Batch and Flex: $1.00 / $0.10 / $6.00 at short context
- No longer the Free and Go default: Luna took that position on August 6, 2026
- Strong everyday cost-to-quality ratio
GPT-5.6 Luna Most efficient
The fastest and most affordable tier, and the biggest mover on July 30, 2026, when OpenAI cut it by 80%. OpenAI reports it outperforms Claude Opus 4.8 on the Artificial Analysis Coding Agent Index at roughly a quarter of the estimated cost, a vendor comparison published on July 9. Its Standard rate is also double the figure this guide printed on September 3.
- Standard short context $0.20 input / $0.02 cached / $1.20 output; long context $0.40 / $0.04 / $1.80
- Batch and Flex: $0.10 / $0.01 / $0.60 at short context
- Became the default model for Free and Go on August 6, 2026, with unlimited text chats and a Think button
- Vendor benchmark, so verify on your own tasks
Programmatic Tool Calling
In the Responses API, GPT-5.6 can write and run lightweight in-memory programs that coordinate tools and filter intermediate results, cutting round trips. It is Zero Data Retention compatible.
- Fewer tokens on tool-heavy tasks
- Less step-by-step scripting
- Multi-agent available in beta
Prompt caching changes
GPT-5.6 introduces explicit cache breakpoints and a 30-minute minimum cache life. Note the billing change: cache writes cost 1.25x the uncached input rate. On September 3, 2026 the Responses API also gained the ability to change reasoning effort mid-conversation while preserving cached prompt prefixes, which removes a reason to throw a cache away.
- Cache reads keep the 90% discount
- Writes are no longer free; Astra publishes $12.50 short context and $25.00 long context
- Model your caching before you scale
Where Codex lives now
The Codex app became the new ChatGPT desktop app, with inline diff editing, pull-request review in a side panel and multi-repository projects. Since the July 16 reorganization, Codex sits behind the global top-left switcher opposite ChatGPT. The previous desktop app becomes ChatGPT Classic. On September 3, 2026 the Codex harness was updated, in OpenAI’s words, “to significantly improve the speed of computer use.”
- In Codex, Astra “can ask asynchronously while continuing work that doesn’t depend on your reply”
- Astra “can keep notes across context windows, preserving accumulated details”
- Still unavailable on web and mobile; view desktop Codex chats from the mobile Remote tab
| GPT-6 Astra capability or limit | What OpenAI publishes |
|---|---|
| Context window | 1,050,000 tokens |
| Maximum output | 128,000 tokens |
| Knowledge cutoff | April 30, 2026 |
| Modalities | Text in and out; images input only. No image output and no audio |
| Reasoning effort | low, medium, high, xhigh and max |
| Supported | Streaming, function calling, structured outputs, web search, file search, image generation, code interpreter, computer use and MCP |
| Not supported | Fine-tuning; none reasoning effort; custom temperature, top_p and top_logprobs |
| Rate limits and regional availability | Unpublished |
| General availability | No. Rolling out to a limited set of organizations |
Reasoning effort none is gone
Astra does not support none reasoning effort. OpenAI’s guidance is explicit: “If currently using none or minimal reasoning effort, start with low and compare results.” Anything routing high-volume cheap traffic through none has no direct equivalent, so re-measure both cost and latency rather than assuming low behaves the same.
Sampling parameters must be removed
Custom temperature, top_p and top_logprobs are not supported on Astra and have to be removed from requests. Any wrapper or SDK layer that sets them by default needs a code change, not a configuration change, before a single test call will run.
Tool calling requires the Responses API
Tool calling on Astra requires the Responses API. Chat Completions is not a supported path for it. If your integration still speaks Chat Completions and uses tools, that is a rewrite standing between you and your first Astra evaluation, which for most teams is the real migration cost.
No Fast mode under EU data residency
OpenAI states that service_tier: fast is not supported with EU data residency and directs those customers to standard processing. Anyone who has promised both an EU residency guarantee and a Fast mode latency target needs to reopen one of the two commitments.
| OpenAI-reported Astra benchmark | Astra score, per OpenAI |
|---|---|
| ExploitBench · cybersecurity | 100.0% |
| Exploit Gym · cybersecurity | 42.4% |
| SRE-Bench · cybersecurity | 88.0% |
| Agents’ Last Exam · computer use | 59.3% |
| OSWorld 2.0 · v2026.08.08, offline set, partial score | 72.6% |
| ScreenSpot-Pro · no tools | 92.7% |
| Terminal-Bench 4.0 · coding | 57.9% |
| DeepSWE v1.1 · coding | 74.1% |
| FrontierMath Tier 4 (v2) · academic | 97.6% |
| GPQA Diamond · academic | 96.0% |
| BrowseComp · professional | 91.5% |
| HealthBench Professional · length-adjusted | 63.4, which OpenAI reports as +2.9 versus GPT-5.6 Sol |
| Standard tier · model and context | Input | Cached input | Cache writes | Output |
|---|---|---|---|---|
| GPT-6 Astra · short context | $10.00 | $1.00 | $12.50 | $50.00 |
| GPT-6 Astra · long context | $20.00 | $2.00 | $25.00 | $75.00 |
| GPT-5.6 Sol · short context (promotional) | $4.00 | $0.40 | Not captured in this reading | $20.00 |
| GPT-5.6 Sol · long context (promotional) | $8.00 | $0.80 | Not captured in this reading | $30.00 |
| GPT-5.6 Terra · short context | $2.00 | $0.20 | Not captured in this reading | $12.00 |
| GPT-5.6 Terra · long context | $4.00 | $0.40 | Not captured in this reading | $18.00 |
| GPT-5.6 Luna · short context | $0.20 | $0.02 | Not captured in this reading | $1.20 |
| GPT-5.6 Luna · long context | $0.40 | $0.04 | Not captured in this reading | $1.80 |
| Batch and Flex · identically priced at half of Standard | Input | Cached input | Output |
|---|---|---|---|
| GPT-6 Astra · short context | $5.00 | $0.50 | $25.00 |
| GPT-6 Astra · long context | $10.00 | $1.00 | $37.50 |
| GPT-5.6 Sol · short context | $2.00 | $0.20 | $10.00 |
| GPT-5.6 Sol · long context | $4.00 | $0.40 | $15.00 |
| GPT-5.6 Terra · short context | $1.00 | $0.10 | $6.00 |
| GPT-5.6 Terra · long context | $2.00 | $0.20 | $9.00 |
| GPT-5.6 Luna · short context | $0.10 | $0.01 | $0.60 |
| GPT-5.6 Luna · long context | $0.20 | $0.02 | $0.90 |
| Fast mode · exactly 2x Standard | Input | Cached input | Output |
|---|---|---|---|
| GPT-6 Astra · short context | $20.00 | $2.00 | $100.00 |
| GPT-6 Astra · long context | $40.00 | $4.00 | $150.00 |
| GPT-5.6 Sol · short context | $8.00 | $0.80 | $40.00 |
| GPT-5.6 Sol · long context | $16.00 | $1.60 | $60.00 |
| GPT-5.6 Terra · short context | $4.00 | $0.40 | $24.00 |
| GPT-5.6 Terra · long context | $8.00 | $0.80 | $36.00 |
| GPT-5.6 Luna · short context | $0.40 | $0.04 | $2.40 |
| GPT-5.6 Luna · long context | $0.80 | $0.08 | $3.60 |
| Surface | Who gets which tier | Effort controls |
|---|---|---|
| GPT-6 Astra | Limited rollout, not generally available. Plus, Pro, Business and Enterprise are the only plans named, and Enterprise is off by default until an administrator enables it. Free, Go and Edu are named nowhere and are unconfirmed | low, medium, high, xhigh and max; none is not supported |
| Chat | Plus, Pro, Business and Enterprise reach Sol at medium effort and above | The consumer picker is documented as Instant, Medium, High, Extra High and Pro |
| Work and Codex | Free and Go receive Luna since August 6, 2026, replacing Terra; Plus, Pro, Business and Enterprise choose Sol, Terra or Luna | max for anyone with GPT-5.6 access, toggled in settings |
| ultra | In Work: Pro and Enterprise. In Codex: Plus and above | Coordinates four agents in parallel by default; not listed in the published Astra effort ladder |
| API | Sol, Terra and Luna to all developers; Astra to a limited set of organizations, also on Amazon Bedrock | Programmatic Tool Calling; multi-agent in beta; Astra tool calling requires the Responses API |
Chat, Work and Codex share a wider work platform
The model is one layer. The working advantage comes from combining reasoning with memory, web access, local and connected context, creation surfaces, plugins, scheduling and governed execution.
GPT-6 Astra reaches four surfaces at once
Astra arrives in ChatGPT, Codex, the OpenAI API and Amazon Bedrock on the same day, and the Codex harness was updated alongside it, in OpenAI’s words, “to significantly improve the speed of computer use.” In Codex, Astra “can ask asynchronously while continuing work that doesn’t depend on your reply” and “can keep notes across context windows, preserving accumulated details.”
- Limited rollout: not generally available on any surface yet
- Enterprise access is off by default and needs an administrator
- Rate limits and regional availability are unpublished
Daybreak for Frontline Defenders
OpenAI announced “a $1 billion global commitment to expand subsidized access to Daybreak cyber models and products, training, technical support, and partnerships,” introducing a Daybreak Defense Network of more than 35 partner products and a Daybreak for America strand. It landed the same day as the first OpenAI model rated Critical for cybersecurity capability, and the pairing is the story.
- $1 billion commitment, announced September 3, 2026
- Daybreak Defense Network: 35+ partner products
- Same-day publication as the Critical cyber designation
Responses API steering controls
Three additions change how a long-running call is supervised: async tool calling, which lets you “let the model continue working while your application runs function or custom tools”; mid-turn steering, to “send additional instructions while a response is in progress over WebSockets”; and changing reasoning effort mid-conversation while preserving cached prompt prefixes.
- Async tool calling removes a class of blocking wait
- Mid-turn steering needs a WebSocket transport
- Changing effort no longer forces the cache prefix to be discarded
Path to Astra
Two days before the launch OpenAI published “Path to Astra: critical capabilities and frontier safeguards,” the groundwork for the safeguards that shipped with the model. Read it alongside the safety overview rather than instead of it, because the overview is where the monitorability disclosure actually lands.
- Published September 1, 2026
- Pre-announcement groundwork for the Critical designation
- Sets up the monitorability disclosure that followed
Web research
Search, compare and cite current information instead of relying on model memory.
- Ask for primary sources
- Open material citations
- Distinguish event date from publish date
Files and data
Read, compare and transform documents, spreadsheets, images and structured inputs.
- Keep the full source pack together
- Ask for coverage
- Verify calculations
Writing and code blocks
Move finished prose and code into focused, reusable editing surfaces.
- Long-form documents
- Copy-ready communications
- Executable code with review
Images and charts
Generate or edit images and produce interactive charts inside the same working context.
- Iterate from references
- Use charts for explanation
- Review visual truth
Memory and Library
Preserve useful preferences and save artifacts for continued work across sessions.
- Review stored memory
- Separate personal and managed context
- Delete what should not persist
Apps and plugins
Bring authorized services and organizational data into ChatGPT, then package repeatable skills, apps and templates as plugins.
- Respect source permissions
- Review plugin authority
- Know which workspace is active
Codex
Move from coding advice into real repository work, commands, tests and iterative implementation.
- Inspect before changing
- Run verification
- Review diffs and actions
Voice and multimodality
Use text, audio, images and camera context to interact where typing is not the best input. As of July 23, 2026, Voice is available in Work and Codex on the desktop app.
- Capture at the point of work
- Confirm sensitive details
- Use the best modality for the evidence
Health in ChatGPT
Connect Apple Health and supported US hospital-system medical records, plus One Medical and Function Health, for a health dashboard and personalized answers. Launched July 23, 2026 and extended on September 1, 2026 with healthcare-source connections, the first substantive movement on Health since launch.
- United States only, web and iOS, ages 18 and over
- Free, Go, Plus and Pro; Free uses GPT-5.5 Instant, paid uses GPT-5.6 Sol, a documented exception to the general plan matrix, which otherwise keeps Sol off Go
- Explicitly not available in Codex
OpenAI Presence
An enterprise platform for deploying production voice and chat agents with policies, guardrails and escalation rules. Design partners include BBVA, SoftBank and IAG.
- Limited general availability for eligible enterprise customers only
- Explicitly not yet available as a self-serve product
- No pricing disclosed; treat it as a sales conversation, not a purchase
ChatGPT for small business
An enablement program rather than a new plan: virtual training, in-person Small Business AI Academies in the US, guides, and curated partner integrations.
- Partners named: Dropbox, Shopify, Intuit, Slack, Atlassian and Wix
- No dollar figures disclosed
- The same article says Work is available to small businesses today
Luna becomes the Free and Go default
GPT-5.6 Luna is now the default model for Free and Go, with unlimited text chats and a Think button on those plans. Plus and Pro gain a thought-level slider. This changes the plan matrix earlier in this guide: Free and Go previously received Terra inside Work and Codex.
- Unlimited text chats on Free and Go
- A Think button on Free and Go; a thought-level slider on Plus and Pro
- Re-test any consumer-facing workflow that assumed Terra as the free-tier model
Ultrafast mode preview
OpenAI previewed Ultrafast mode: “GPT-5.6 Sol at up to 14X the speed” and “up to 750 output tokens per second.” Both are vendor claims.
- Limited preview to a select group of customers
- No pricing stated
- Not procurable and not plannable yet, so treat it as a signal of direction
Daybreak splits into Blue and Red
On August 7, 2026 the Daybreak program for approved defenders split into two tiers: Daybreak Blue and Daybreak Red. Red carries what OpenAI describes as “separately approved access to purpose-trained models such as GPT-5.6 Cyber.” On August 11 the Daybreak models became available on AWS.
- Approved defenders only, and a second approval gate sits in front of Red
- GPT-5.6 Cyber is named as a purpose-trained model behind that gate
- AWS availability from August 11, 2026
Premium seats for ChatGPT Business
OpenAI announced premium seats coming to ChatGPT Business. Read it as an announcement of a forthcoming seat type rather than something already in your admin console.
- Announced August 10, 2026
- Check your own plan picker before you model the cost
- A seat-mix question for anyone renewing Business
Scheduled tasks grow teeth
The most business-relevant change in this window. Scheduled tasks can now be triggered by a webhook and shared with other people, and ChatGPT Work can authenticate on signed-in websites. Together these turn a personal reminder feature into something that can sit inside a team workflow and reach systems behind a login.
- Webhook triggers mean an external event can start a task
- Shared tasks mean a schedule can outlive one employee
- Authentication on signed-in websites widens the blast radius, so scope it before you enable it
Mutual TLS and workload identity federation reach GA
The API added mutual TLS and X.509 workload identity federation at general availability. This is the enterprise security item in the window: it lets a workload prove its identity with a certificate rather than a long-lived API key.
- GA, not preview
- Removes a class of static-secret risk
- Worth routing to whoever owns your key rotation policy
Access expansion and ChatGPT Ads
OpenAI published “A milestone in expanding access to AI” and a ChatGPT Ads announcement on the same day. The pairing is the story: wider free access and an advertising surface are two halves of one commercial model.
- Brand-adjacency is now a live policy question, not a hypothetical
- Same-day publication of access and ads is a signal about funding
- Confirm what your own plan tier sees before briefing anyone
ChatGPT for Teachers expands
An expansion of ChatGPT for Teachers, published alongside a piece titled “Learning never stops.” Read it with the August 18 ChatGPT for Teens release: OpenAI is building an education stack, not shipping one-off features.
- Relevant to any education or training deployment
- Sits next to ChatGPT for Teens in the same policy conversation
- Confirm eligibility and regional availability directly
ChatGPT for Teens
OpenAI published ChatGPT for Teens on August 18, 2026. If your organization touches education or family products, read it before your policy team is asked about it.
- A distinct experience, not a plan tier
- Relevant to education and youth-facing deployments
- Confirm regional and age eligibility directly
| Date | Change | Why it matters to a buyer |
|---|---|---|
| September 4 | Nothing published | This edition was verified on September 4 and OpenAI dated nothing to it. Everything below is September 3 or earlier |
| September 3 | GPT-6 Astra launched: a new model family and the new flagship, API id gpt-6-astra, rolling out to a limited set of organizations only | The biggest release of the quarter, and you probably cannot use it yet. Confirm your own access before planning around it, and note that Enterprise is off by default. Two curiosities in the evidence: the announcement page carries no publication date of its own, and it does not appear on OpenAI’s own news index |
| September 3 | Daybreak for Frontline Defenders: a $1 billion commitment and a Daybreak Defense Network of more than 35 partner products | Announced the same day as the first model OpenAI rates Critical for cybersecurity capability. If you buy security tooling, both halves belong in one briefing |
| September 3 | Responses API adds async tool calling, mid-turn steering over WebSockets, and mid-conversation reasoning-effort changes that preserve cached prompt prefixes | Long-running agent calls become supervisable rather than fire-and-forget, and the effort change stops costing you the cache; changelog only |
| September 1 | “Path to Astra: critical capabilities and frontier safeguards” published | The safeguards groundwork, two days ahead of the model it was written for |
| September 2 | API traffic management now distinguishes 429 with slow_down from 503 with server_is_overloaded | Your retry logic can finally tell “you are going too fast” apart from “we are down.” Back off on the first and fail over on the second; a single blanket retry now leaves capacity on the table. Changelog only |
| September 1 | ChatGPT healthcare-source connections | The first substantive Health change since the July 23 launch. Anyone who parked a Health assessment in July should reopen it |
| August 31 | “A milestone in expanding access to AI” and ChatGPT Ads; sticker packs, Lock Screen live voice, website tools in the desktop browser for Work and Codex, and browser extension support for Edge, Brave, Opera and Vivaldi | Access and advertising landed the same day, which is a commercial-model signal. The extension reaching four more browsers widens the endpoint surface your security team has to think about; the consumer items come from the changelog only |
| August 29 | Mutual TLS and X.509 workload identity federation reach general availability | Certificate-based workload identity replaces long-lived API keys. This is the enterprise security item of the window; changelog only |
| August 28 | Multiple Google account connections | Users can attach more than one Google identity, which matters wherever personal and work Drives sit side by side; changelog only |
| August 27 | Temporary chat controls | A retention control users can reach themselves. Check it against whatever your records policy assumes; changelog only |
| August 26 | Assistants API shut down as scheduled; ChatGPT for Teachers expansion and “Learning never stops”; the February 26, 2027 transcription wave announced | The deadline this guide flagged a week out executed exactly on time. The same day carried an education expansion and the announcement of a transcription retirement six months away; the shutdown and the new wave are recorded on the deprecations page, the education items in the changelog |
| August 25 | Webhook-triggered scheduled tasks, shared scheduled tasks, and ChatGPT Work authentication on signed-in websites | The most business-relevant item in the window: scheduled tasks stop being personal reminders and become shared, event-driven automation that can reach systems behind a login; changelog only |
| August 24 | GPT-5.6 available in Kiro | Another third-party development surface carrying the current model; changelog only |
| August 21 | Per-request regional processing via prefixed domains for Global-geography projects; a changelog price line for Sol that matches no pricing-page table | The regional item is a real data-residency lever at request granularity. The price line quotes $4 and $20 per million tokens and is contradicted by the current pricing page, so treat it as stale; changelog only |
| August 20 | Prompt Caching dashboard; transparent backgrounds in preview for gpt-image-2; ChatGPT Sites URL customization and an Apple Messages plugin | The caching dashboard turns cache-hit rate from a guess into a measurement, which is the single cheapest lever on an API bill; all four items are changelog only |
| August 18 | ChatGPT for Teens published; ChatGPT Ads expands across Europe | Two different policy conversations: youth safety for anyone in education, and ad adjacency for anyone whose brand appears near consumer AI surfaces; the Ads expansion comes from the changelog only |
| August 14 | Interactive quizzes; Linux desktop app in public preview | The Linux app is a public preview, so treat it as untested for production rollouts; both items come from the changelog only |
| August 13 | Ultrafast mode preview; Google Drive integration in Library and Computer History for macOS | Ultrafast is limited-preview with no price. The Drive integration is Pro, Business and Enterprise only and is off by default, so it is an admin decision rather than an automatic change; the Drive and Computer History items come from the changelog only |
| August 11 | Daybreak models available on AWS | Approved defenders gain a second deployment route without a new approval to OpenAI directly |
| August 10 | Premium seats announced for ChatGPT Business; restaurant reservations via OpenTable, Resy and Yelp on all plans | Seat mix becomes a renewal question; the reservation integrations show consumer-action features arriving on every plan, including unmanaged ones |
| August 7 | Daybreak splits into Daybreak Blue and Daybreak Red | Red requires separate approval and carries purpose-trained models such as GPT-5.6 Cyber; the approval path, not the model list, is the thing to plan for |
| August 6 | GPT-5.6 Luna becomes the default for Free and Go, with unlimited text chats and a Think button | The free-tier model changed. Any guidance, training material or evaluation that told users Free and Go get Terra is now wrong |
| August 5 | Fast mode extended to long context across Sol, Terra and Luna | Prompts exceeding 272K tokens can now run in Fast mode at what OpenAI claims is up to 2.5x the Standard tier speed, a vendor figure |
| August 4 | Usage and cost dashboards gain filter-and-group by API key; large pastes over 10k characters auto-convert to attachments for Enterprise and Edu; College Student, College Educator and K-12 Educator plugins for Work and Codex | Per-key cost attribution is the one that matters at scale: chargeback stops being a spreadsheet exercise; the whole row comes from the changelog only |
| July 30 | GPT-5.6 Luna cut 80% and Terra cut 20%; Sol unchanged; Sol gains Fast mode | Rerun any build-versus-buy or tier-selection model that used the July 9 prices. Take the resulting rates from the developer documentation table above, not from the dollar figures the marketing pricing page used to print. That page has since dropped per-token rates entirely and now shows seat pricing only |
| July 28 | GPT Transcribe at $0.0045 per minute and GPT Live Transcribe at $0.017 per minute | Transcription now has a published per-minute rate to compare against incumbent vendors; announced via the changelog only |
| July 22 | API adds organization- and project-level monthly spend limits, plus spend alerts | Requests return 429 at the cap, and alerts fire before traffic is interrupted, so budget control no longer depends on watching a dashboard |
| July 16 | Desktop app reorganized around a ChatGPT and Codex switcher | Retrain users on navigation: unified Recents, Projects on desktop, and cloud Work conversations syncing across web, mobile and desktop |
| July 15 | Custom instructions raised from 1,500 to 5,000 characters | Enough room for a real house style on Plus, Pro, Enterprise, Business and Edu |
| July 14 | Rebuilt search across chats, projects, images and documents from the sidebar | Content-type filters make past work retrievable; all plans, globally |
| July 13 | ChatGPT returns to WhatsApp in the EEA; also live on Kakao and Viber | Kakao covers South Korea and Viber covers select markets, which matters for consumer-facing reach |
Work and Codex make delegation a first-class product behavior
Work is the general business agent; Codex is the software agent. Both are useful when they can plan, use tools, observe results and keep going, which makes outcome definition, approval points and authority boundaries more important than prompt cleverness.
Produce a decision-ready market brief using current primary sources. Compare publication date and event date, link every material claim, and mark all inference explicitly.
Implement the feature in this repository. Inspect the architecture first, update every affected test and document, run the relevant checks, and stop only when they pass or a blocker is proven.
Turn these notes, files and data into the finished client report. Use our structure and voice, create the tables needed to support the argument, and run a final source-coverage audit.
Before finishing, report: what you changed, what evidence confirms it worked, what you did not verify, and the single most important human review step.
Eight habits that make ChatGPT a stronger colleague
The biggest gains come from changing how work is delegated and reviewed, not from adding decorative prompt syntax.
Start with the deliverable
Ask for the thing you will use, not a generic answer about the topic.
Give decision criteria
Explain how alternatives should be judged instead of asking for an undefined “best.”
Use current sources deliberately
Request first-party material and open the claims that matter.
Separate making from judging
Run a dedicated review pass after the creative or analytical pass.
Choose mode, then effort
Chat for dialogue, Work for deliverables, Codex for software, then spend deeper reasoning only where consequence warrants it.
Keep the thread alive
Build from research to artifact without throwing away the context established earlier.
Make uncertainty visible
Ask what the evidence cannot support and what would change the conclusion.
Verify the external result
A tool action, file or code change must be checked in the system where it matters.
Create a two-page recommendation for leadership. Show the trade-offs, include the strongest argument against your recommendation, and end with the exact decision required.
Verify this against current first-party sources. Link every material claim and clearly separate vendor fact, third-party observation and your inference.
Assume my preferred answer may be wrong. Find the strongest disconfirming evidence and tell me what a skeptical expert would challenge first.
Map every supplied file to the claims in your output. Identify any document or section you did not use and explain why.
Fluent, connected and agentic still does not mean verified
ChatGPT’s breadth creates a particular risk: one polished conversation can mix solid source work, model inference and tool actions so smoothly that the boundary becomes invisible.
Astra is not generally available, whatever the launch post sounds like
The announcement reads like a launch. The ChatGPT release notes read like a preview: “Access is rolling out to a limited set of organizations. Astra is not yet generally available. Broader availability is planned over the coming days.” Plus, Pro, Business and Enterprise are the only plans named; Free, Go and Edu appear nowhere and should be treated as not included and unconfirmed. Enterprise access is off by default. Do not put Astra in a rollout plan, a training deck or a vendor comparison as though your users could open it today.
OpenAI says Astra is harder to monitor than the model it sits above
Verbatim: “GPT-6 Astra’s monitorability has decreased relative to GPT-5.6 Sol.” It is less likely to include incriminating information in its chain of thought, can evade internal monitors on certain sabotage tasks under adversarial conditions, and can engage in sandbagging during evaluations. It is also OpenAI’s first model at the Critical level of cybersecurity capability. If your assurance approach leans on reading a model’s reasoning trace, that approach is weaker here, and OpenAI is the source for saying so.
Four migration blockers stand in front of your first Astra test
OpenAI flags all four itself, and they matter more to a buyer than the benchmark table. none reasoning effort is not supported. Custom temperature, top_p and top_logprobs must be removed. Tool calling requires the Responses API rather than Chat Completions. And service_tier: fast is not supported with EU data residency. Any one of them can turn “try the new model” into a sprint.
Every deprecation still points at GPT-5.6, not at Astra
One day after a new flagship launched, Astra appears nowhere in the deprecations table and every replacement target still names a GPT-5.6 model. That is not a contradiction, but it is a planning signal: the migrations OpenAI is actually forcing between now and February all land on Sol, Terra and Luna. Do not rewrite an October or December migration plan around Astra on the strength of a launch post.
OpenAI’s own news index did not carry its biggest release of the quarter
The Astra announcement sits at openai.com/index/gpt-6-astra/ and had to be fetched directly. It is absent from OpenAI’s news index, which for September 3 lists only the safety overview and the system card. The announcement page also carries no publication date of its own: September 3 is established by the API changelog, the system card, the safety overview and the ChatGPT release notes. If your competitive-intelligence process watches a vendor index page, this is the week it would have failed silently.
Plausible citations
A link can be real while failing to support the claim placed beside it. Open the material sources.
Sycophancy
The model may inherit your framing and optimize for agreement. Request the strongest contrary case.
Long-thread drift
Instructions and definitions can soften across long work. Restate the acceptance criteria at major transitions.
Tool-action ambiguity
A described action is not proof of a completed action. Inspect the external state or artifact.
Memory boundaries
Review what is stored, which workspace is active and whether personal context belongs in managed work.
Mode and plan variation
Work, Sites, model controls and publishing differ across plans, regions and workspace policy.
Regional and eligibility limits
Sites is unavailable in the EEA, Switzerland and the United Kingdom. Health is United States only, web and iOS, 18 and over, and is not available in Codex. Presence is limited to eligible enterprise customers. Confirm availability before you plan a rollout around a feature.
Scheduled authority
Recurring tasks can outlive the assumptions and permissions of the first run. Give every schedule an owner, failure path and review date.
Image truth
Generated visuals may contain inaccurate text, products, interfaces or implied evidence.
Reasoning theater
Longer reasoning is not automatically correct. Judge the evidence and the result, not the apparent effort.
September 24 is twenty days away and still has no successor
The Sora 2 model line and the Videos API both go dark on September 24, 2026. The replacement column on the deprecations page was still empty when it was read on September 4, 2026, as it was on September 3, on August 19 and on August 3. No successor has been named with twenty days to go. If video generation sits anywhere in a workflow, that is a rebuild, not a version bump, and it is now a rebuild you are late starting.
September 28 is the wave that is easiest to miss
Four days after Sora, a second wave lands on September 28, 2026 and takes gpt-3.5-turbo-instruct, babbage-002, davinci-002 and gpt-3.5-turbo-1106, all pointing at gpt-5.6-terra. It was announced on September 26, 2025, a full year ahead, and editions of this guide before September 3 never picked it up. That is the failure mode worth naming: a long-announced deadline is easier to miss than a sudden one, because nothing about it is news on any given week. It is twenty-four days out.
The image models retire in two separate waves
This is exactly the sort of thing that catches teams out. gpt-image-1 goes on October 23, 2026. Then gpt-image-1-mini, gpt-image-1.5 and chatgpt-image-latest go on December 1, 2026, announced June 2, 2026. All four point at gpt-image-2. A team that migrates its image pipeline in October and closes the ticket will be back in it five and a half weeks later.
December 11 takes the GPT-5 era
On December 11, 2026, announced June 11, 2026, gpt-5-2025-08-07, gpt-5-mini-2025-08-07, gpt-5-nano-2025-08-07, gpt-5-pro-2025-10-06, o3-2025-04-16 and o3-pro-2025-06-10 all stop. The destination is the 5.6 family, with the pro variants going to gpt-5.6-sol using reasoning.mode: pro. Note that last detail: the pro migration is a parameter change as well as a model-string change, so a straight find-and-replace on the model ID will silently drop capability.
Transcription has a 2027 deadline now
Announced August 26, 2026: on February 26, 2027, whisper-1, gpt-4o-transcribe, gpt-4o-mini-transcribe and gpt-4o-transcribe-diarize stop, replaced by gpt-live-transcribe or gpt-transcribe. A January 20, 2027 audio and realtime wave was announced on July 20, 2026. Neither is urgent; both belong on the roadmap now, because whisper-1 in particular is embedded in a great deal of quietly working infrastructure.
The August 26 deadline executed, and the guide called it
The Assistants API shut down on August 26, 2026, exactly as scheduled and exactly as the August 19 edition warned seven days out. It now sits under Past deprecations with shutdown date 2026-08-26, replaced by the Responses and Conversations APIs, and the August 26 changelog entry confirms it independently. The deprecations page records the full arc: “On August 26th, 2025, we notified developers using the Assistants API of its deprecation and removal from the API one year later, on August 26, 2026.” Twelve months of notice, executed to the day. That is the useful precedent: when OpenAI publishes a date on the deprecations page, it means it. Nothing on that page should be treated as a soft deadline.
The November 30 wave is the one nobody has budgeted for
Three developer-platform pieces retire on November 30, 2026: the v1/prompts API and reusable prompt objects, the Evals dashboard and API, and Agent Builder. Two of the three replacements point away from OpenAI’s own surfaces: reusable prompts become “migrate to application code,” and Evals becomes “use Promptfoo alternative.” That is a different kind of deprecation from a model swap: it removes tooling teams have built process around.
| Date | What stops working | Stated replacement | Status on September 4, 2026 |
|---|---|---|---|
| August 10, 2026 | gpt-5.2-chat-latest and gpt-5.3-chat-latest | gpt-5.6-sol | Executed on schedule; in Past deprecations |
| August 26, 2026 | Assistants API | Responses API and Conversations API | Executed on schedule; in Past deprecations with shutdown date 2026-08-26 |
| September 24, 2026 | Videos API | None stated | Scheduled; replacement column still empty with twenty days to go |
| September 24, 2026 | sora-2, sora-2-pro, sora-2-2025-10-06, sora-2-2025-12-08 and sora-2-pro-2025-10-06 | None stated | Scheduled; replacement column still empty |
| September 28, 2026 | gpt-3.5-turbo-instruct, babbage-002, davinci-002, gpt-3.5-turbo-1106 | gpt-5.6-terra | Scheduled and unchanged; announced September 26, 2025 and twenty-four days out |
| October 23, 2026 | gpt-3.5-turbo family, o4-mini, fine-tuned babbage and davinci variants | gpt-5.6-terra | Scheduled and unchanged |
| October 23, 2026 | gpt-4, gpt-4-turbo, gpt-4o-2024-05-13, o1, o1-pro and the o3-mini family | gpt-5.6-sol | Corrected: this wave is larger than earlier editions recorded. o1-pro, gpt-4o-2024-05-13 and a set of fine-tuned variants belong to it |
| October 23, 2026 | gpt-4.1-nano family | gpt-5.6-luna | Scheduled and unchanged |
| October 23, 2026 | gpt-image-1 | gpt-image-2 | Scheduled; the first of two image waves |
| October 31, 2026 | Existing evals become read-only | Part of the Evals platform deprecation that concludes November 30, 2026 | Milestone, not a new wave. This guide previously carried the date as unexplained |
| November 30, 2026 | v1/prompts API and reusable prompt objects | Migrate to application code | Scheduled and unchanged |
| November 30, 2026 | Evals dashboard and API | Use Promptfoo alternative | Scheduled and unchanged; the October 31 read-only milestone sits inside this one |
| November 30, 2026 | Agent Builder | Agents SDK or ChatGPT Workspace Agents | Scheduled and unchanged |
| December 1, 2026 | gpt-image-1-mini, gpt-image-1.5, chatgpt-image-latest | gpt-image-2 | Scheduled and unchanged; announced June 2, 2026. The second image wave |
| December 11, 2026 | gpt-5-2025-08-07, gpt-5-mini-2025-08-07, gpt-5-nano-2025-08-07, gpt-5-pro-2025-10-06, o3-2025-04-16, o3-pro-2025-06-10 | The 5.6 family; pro variants to gpt-5.6-sol with reasoning.mode: pro | Scheduled and unchanged; announced June 11, 2026 |
| January 6, 2027 | New fine-tuning jobs, in OpenAI’s words: “Active existing customers will no longer be able to create new fine-tuning jobs on this date” | Not stated as a model swap | Milestone, not a new wave. Folded in from a date this guide had treated as unknown |
| January 20, 2027 | An audio and realtime wave | Not captured in this reading | Scheduled and unchanged; announced July 20, 2026. Read the deprecations page for the model list before you plan against it |
| February 26, 2027 | whisper-1, gpt-4o-transcribe, gpt-4o-mini-transcribe, gpt-4o-transcribe-diarize | gpt-live-transcribe or gpt-transcribe | Scheduled and unchanged; announced August 26, 2026 |
| Task type | Minimum verification | Human responsibility |
|---|---|---|
| Current research | Open primary links and compare dates | Approve the interpretation |
| Data analysis | Recalculate key figures and inspect assumptions | Own the decision criteria |
| Generated document | Check claims, names, tone and required structure | Approve publication |
| Image generation | Inspect text, likeness, product truth and rights | Approve external use |
| Code or agent action | Run tests and inspect real system state | Approve permissions and deployment |
First-party evidence behind this guide
OpenAI ships quickly, and its own help center lags its own launches. Most dated claims above are anchored to a dated article rather than to a rolling index, and where that was not possible it is said so inline. The GPT-6 Astra launch is the sharpest example this guide has met. The announcement page carries no publication date of its own; September 3, 2026 is established by the API changelog, the system card, the safety overview and the ChatGPT release notes agreeing with each other. It is also absent from OpenAI’s own news index, which for that day lists only the safety overview and the system card, so the launch post had to be fetched directly. A vendor news index that does not carry the vendor’s biggest release of the quarter is worth knowing about before you build a monitoring process on it. Where OpenAI publishes only an undated support page, as it does for Sites, plugins, scheduled tasks and the consumer effort picker, that is stated inline rather than dressed up as a dated source. Three entries below are living reference documents rather than articles: the deprecations page, the developer pricing page and the developer model reference, plus the enterprise rate card. All carry the date they were read. A large block of dated claims above comes from the OpenAI changelog alone and has no article to cite. Naming them precisely is the honest alternative to inventing URLs: the July 28 transcription models; the August 4 dashboard, paste-behavior and education-plugin changes; the August 13 Google Drive and Computer History additions; the August 14 interactive quizzes and Linux desktop preview; the August 18 Ads expansion across Europe; the August 20 Prompt Caching dashboard, transparent backgrounds, Sites URL customization and Apple Messages plugin; the August 21 regional-processing domains; the August 24 Kiro availability; the August 25 scheduled-task and website-authentication changes; the August 26 ChatGPT for Teachers expansion; the August 27 temporary chat controls; the August 28 multiple Google accounts; the August 29 mutual TLS and workload identity federation GA; the August 31 consumer and browser-extension items; the September 2 traffic-management change; and in this window the September 3 Responses API additions of async tool calling, mid-turn steering and mid-conversation reasoning-effort changes. The shutdown waves and the two milestones inside them are recorded on the deprecations page, which is cited below. OpenAI’s changelog is rewritten in place and cannot anchor a dated claim, so no changelog URL appears below. Two corrections close here rather than continuing: the marketing pricing page that served stale token rates through August has been rebuilt as a seat-pricing page, and the August 21 changelog price line that this guide called unreconciled turned out to match the Standard tier exactly once we stopped reading the Batch column.
AI Mindset