The Google-Native Agent That Can See and Act
Google prints three different prices for the same model in one table. Quote the wrong one and your forecast is out by 2x. So two habits matter more here than any model choice. Name the pricing tier in the same breath as the number, because batch is exactly half of standard. And build on an explicit model ID rather than the gemini-flash-latest alias. Now the lineup. Gemini 3.8 Flash reached general availability on September 2, 2026. It is the newest Flash model in the API, and Google recommends it for computer use. Gemini 3.1 Pro is the shipping Pro model and the one sold inside every consumer AI plan, still in preview seven months after launch. Gemini 3.8 Live and Live Extended Thinking went generally available on September 15. That is also the voice answer for Workspace.
One intelligence layer across Google’s working surfaces
Gemini isn’t one chat product. It’s a family of models and experiences: Search, the Gemini app, Workspace, Android, developer tools, voice and creative systems. Sold through consumer plans as well as the API.
Fast, multimodal and connected
Gemini is strongest when the work mixes current information, visual or voice or audio context, and a Google service that can help finish the job.
- Search and Maps grounding
- Text, image, voice and video
- Consumer, enterprise and developer surfaces
Information-rich action
Use Gemini for current research, Workspace creation, Android assistance, live voice agents, proactive monitoring and agents that need to see and operate interfaces.
- Research and generative UI
- Gmail, Docs, Slides and Sheets
- Browser, mobile and desktop automation
The context already lives with Google
Gemini has a built-in head start when the task begins in Search, Gmail, Drive, Android, Maps, YouTube or a Google developer environment.
- Less context transfer
- Live information
- One model family across many surfaces
A published cutoff, a Pro model stuck in preview, and a deadline nobody confirmed
Name the tier or the number means nothing. Every rate quoted here is paid-tier standard. Batch is exactly half of it. The free tier is free of charge for the Flash text models. Pricing itself hasn’t moved. The lineup has. On September 2, 2026 Google took Gemini 3.8 Flash to general availability, twenty days after 3.7 Flash reached the same status, and published a model card that gives it a March 2026 knowledge cutoff. Gemini 3.8 Live and 3.8 Live Extended Thinking went generally available on September 15. The Pro story is two slipped timelines rather than one. Gemini 3.1 Pro is the shipping Pro model and has been in preview since February 19, 2026. Gemini 3.5 Pro is four months past the “next month” Google promised on May 19. The computer-use documentation named only Gemini 3.5 Flash when the capability launched in June. It now names 3.8 Flash as the recommended model, and Gemini 3.6 Flash is still absent from that list across a fresh edit of the page. And two shutdown dates are behaving badly in opposite directions. The withdrawn Gemini 2.5 shutdown is still withdrawn. The robotics model that was due to stop serving on August 31 has no first-party confirmation, eighteen days later, that it did.
gemini-robotics-er-1.6-preview has been past its August 31, 2026 shutdown date with three Google pages still presenting it as livegemini-2.5-flash-image and the October 2, 2026 shutdown Google has scheduled for that same modelGemini 3.8 Flash GA September 2
The newest Flash model in the API. Its DeepMind model card gives up to 1M tokens of context, 64K tokens of output and a March 2026 knowledge cutoff. Google’s framing is enthusiastic, and it is Google’s. The launch post calls it “our most intelligent workhorse model, delivering significant improvements from 3.7 Flash across software engineering, agentic tasks, and critical, multi-step reasoning.” The docs call it “Our most intelligent Flash model, engineered for long-horizon software engineering, autonomous agents, and complex enterprise workflows.”
- Pin
gemini-3.8-flash - Knowledge cutoff March 2026, published on the model card
- Same promotional pricing as 3.7 and 3.6 Flash, including the January 1, 2027 doubling
- The recommended model for computer use
The knowledge cutoff Google publishes in one place and omits in the other Check both pages
Nobody re-checks a negative. That is how a missing field turns into a false claim. The Gemini 3.8 Flash model card, dated September 2, 2026, publishes a March 2026 knowledge cutoff, with the caveat that in some domains users “may experience the model’s knowledge is limited to January 2025.” Google’s API model page lists the 1,048,576-token input limit, the 65,536-token output limit and a September 2026 update stamp, and it simply has no cutoff field. So read both pages before you report that Google hasn’t published something. The model card is the specification document and it is dated to the launch. The card wins. The omission is an omission, not a denial.
- State March 2026, with the January 2025 caveat attached
- Two Google surfaces, one figure and one blank: quote the card
- Where the vendor contradicts itself, say which document you trusted and why
Gemini 3.1 Pro The shipping Pro model
gemini-3.1-pro-preview is what a buyer actually gets when they ask for Pro. Note the suffix. Google launched it on February 19, 2026 with the words “we are releasing 3.1 Pro in preview today to validate these updates,” and promised general availability soon. Seven months later it’s still a preview endpoint. Google describes it as offering “advanced intelligence, complex problem-solving skills, and powerful agentic and vibe coding capabilities.” It is priced on the paid tier, and it is the Pro model inside the consumer AI plans. Its pricing is context-tiered: paid-tier standard is $2.00 per 1M input and $12.00 per 1M output for prompts of 200k tokens or fewer, with batch at $1.00 and $6.00 on the same tier.
- Pin
gemini-3.1-pro-previewand price the context tier you will actually send - In preview since February 19, 2026, which is a second slipped Pro timeline
- Do not tell a client that Gemini has no Pro model; tell them the Pro model is a preview
Gemini 3.1 Deep Think Listed, not detailed
DeepMind lists Gemini 3.1 Deep Think as “best for modern challenges across science, research and engineering,” and Google sells Deep Think inside the top consumer AI plan. That listing sits on a rolling DeepMind index rather than a dated article. Enough to name the model. Not enough to describe it. Confirm the current availability and plan gating directly before you put it in a proposal.
- Positioned for research and engineering work
- Bundled with the highest consumer tier
- Rolling-page evidence only: confirm before you quote it
Gemini 3.8 Live and 3.8 Live Extended Thinking GA September 15
Two audio-to-audio Live API models, both generally available and both listed as stable. Google positions gemini-3.8-live as the “default Live API model for most low-latency voice agent experiences without reasoning delays,” and gemini-3.8-live-extended-thinking as the “high-reasoning Live API model for voice interactions.” Both are available through the Gemini API and Google AI Studio. The low-latency model is also on Search Live. Enterprise access is a private preview. Google says the pair “deliver the building blocks for reliable, production-ready voice agents,” which is a vendor claim. Paid-tier standard is $0.75 per 1M input and $4.50 per 1M output for text, with the free tier free of charge. Extended Thinking is rolling out to Workspace and Gemini Live subscribers in Docs, Gmail and Keep.
- Choose latency or reasoning, then pin the matching ID
- Google recommends migrating off
gemini-3.1-flash-live-preview, now marked legacy - Enterprise deployment is private preview, so do not promise it
Gemini 3.7 Flash Previous generation
Generally available August 13, superseded September 2. The docs now describe it as “Our previous-generation Flash model for complex coding, agentic workflows, and reliable multi-step execution.” Its model card gives a 1,048,576-token input limit, a 65,536-token output limit and a March 2026 knowledge cutoff. It is still listed for computer use, and its pricing is identical to 3.8 Flash.
- Pin
gemini-3.7-flashif you are staying on it - No forced migration to 3.8 Flash has been announced
- Listed for computer use, unlike 3.6 Flash
- Same promotional rate, same January 1 step-up
The recycled superlative Read this before you quote Google
Three Flash generations. One adjective. At its August 13 launch Google called Gemini 3.7 Flash “our most intelligent workhorse model yet for coding and agents.” Three weeks later near-identical framing belongs to 3.8 Flash. The words were accurate on both days and distinguishing on neither. So treat a vendor superlative as a marketing constant. Take the differences from the model card, the price table and your own traces.
- The same superlative has now covered three Flash generations
- Quote it with its date attached or not at all
- Benchmark the models yourself rather than reading the adjectives
Gemini 3.6 Flash Two generations back
The July 21 default, still listed and still on exactly the same promotional pricing as 3.7 and 3.8 Flash. The models page has re-described it as “Our previous-generation Flash model, balancing speed and multimodal capabilities across general agentic and everyday tasks.” Its distinguishing claim travels with it. Google wrote that “according to the Artificial Analysis Index, it reduces output token usage by 17% compared to 3.5 Flash.” And it’s the one recent Flash model absent from the computer-use list.
- Pin
gemini-3.6-flashif you are staying on it - The 17 percent figure is the Artificial Analysis Index, cited by Google, not a Google benchmark
- Still not listed for computer use as of the September 17, 2026 docs stamp
- Same promotional rate, same January 1 step-up
Gemini 3.1 Flash-Lite The new low-cost floor
The cheapest model Google sells has changed hands, and the model that used to hold the job is still listed at its old price. gemini-3.1-flash-lite is listed as stable, and described as offering “frontier-class performance rivaling larger models at a fraction of the cost.” It is priced on the paid standard tier at $0.25 per 1M input tokens for text, image and video, and $1.50 per 1M output tokens. That undercuts Gemini 3.5 Flash-Lite on both sides of the meter.
- Pin
gemini-3.1-flash-litefor new high-volume work - Cheaper than 3.5 Flash-Lite on input and on output
- The performance claim is Google’s: test it on your own traffic before you reroute
Gemini 3.5 Flash-Lite No longer the cheapest
Still a working high-throughput tier, with Google reporting roughly 350 output tokens per second, and still listed for computer use. List price is $0.30 per 1M input tokens and $2.50 per 1M output tokens on the paid standard tier. It is now the more expensive of the two Flash-Lite models, so route new volume deliberately rather than by habit.
- Pin
gemini-3.5-flash-liteonly if you need it specifically - Listed for computer use, which 3.1 Flash-Lite is not
- Compare against
gemini-3.1-flash-litebefore you scale
Gemini 3.5 Flash Now described as legacy
The May 19 model has survived three newer defaults, but Google’s language about it has changed. The models page now calls it “our legacy Flash model, providing baseline speed and foundational performance for routine, high-throughput workloads.” 3.6 Flash gets “previous-generation.” The computer-use page still lists 3.5 Flash as a previous stable model. And it sits outside the promotion at $1.50 per 1M input and $9.00 per 1M output, with caching at $0.15 and batch at $0.75 and $4.50. So the oldest supported Flash is the expensive one.
- “Legacy” is Google signaling the next deprecation candidate, not a shutdown date
- No forced migration has been announced
- You are paying a premium to stay on it: move deliberately, but do plan the move
Gemini 3.5 Pro Four months overdue
On May 19, 2026 Google said Pro was “already being used internally, and we look forward to rolling it out next month.” That month was June. On September 18, 2026 it’s still absent from the models page and the pricing page. The May post remains the only first-party statement of a date, and on July 21 Google said only that it is testing with partners and will be broadly available when ready. DeepMind’s Gemini page still carries a “3.5 Pro coming soon” line, so the promise is live on one surface and silent on the others.
- Four months past the stated window
- Do not promise availability or a date
- Gemini 3.1 Pro is the Pro model that exists today
Computer use Extended, with one gap
The computer-use documentation, stamped September 17, 2026, names Gemini 3.8 Flash as “The recommended model for computer use, featuring high-accuracy UI interaction and reliable tool calling.” It also lists Gemini 3.7 Flash and Gemini 3.5 Flash as previous stable models, plus Gemini 3.5 Flash-Lite, Gemini 3 Flash Preview and the legacy gemini-2.5-computer-use-preview-10-2025. Gemini 3.6 Flash is still not listed. The omission survived a re-edit of the page, which makes it look deliberate rather than stale.
- 3.8 Flash is the recommended model
- 3.7 Flash, 3.5 Flash and 3.5 Flash-Lite are listed
- 3.6 Flash is still not listed: check the ID you pin
Gemini 3.8 Flash Cyber Fairwind only
The cyber model moved up three generations, from Gemini 3.5 Flash Cyber to Gemini 3.8 Flash Cyber, and changed its access route on September 2, 2026. It is “only available to trusted defenders who require a more comprehensive set of cyber capabilities.” Access now runs through the Fairwind Program, “a limited access program for governments and trusted partners to use our cyber defense tools.” CodeMender remains the tooling layer. Google says “CodeMender with Gemini 3.8 Flash Cyber delivers the specialized reasoning to write and validate code fixes, at a fraction of the operating cost.” Google Cloud customers outside Fairwind can still use CodeMender with publicly available models through the Gemini Enterprise Agent Platform.
- Not a purchasable capability
- The Fairwind page does not mention Gemini 3.5 Flash Cyber at all, so its fate is unconfirmed rather than concluded
- CodeMender is available more widely than the cyber model is
| Model or capability | Paid-tier standard, per 1M tokens | What a buyer needs to know |
|---|---|---|
gemini-3.8-flash, gemini-3.7-flash, gemini-3.6-flash | $0.75 in / $3.75 out through December 31, 2026, then $1.50 / $7.50 | Batch is $0.375 / $1.875 now and $0.75 / $3.75 from January 1. Free tier is free of charge. |
gemini-3.1-pro-preview | $2.00 in / $12.00 out for prompts of 200k tokens or fewer | Context-tiered pricing, so a longer prompt is not on this rate. Batch is $1.00 / $6.00 at the same context tier. |
gemini-3.5-flash | $1.50 in / $9.00 out | Outside the promotion, so the model Google now calls legacy is the expensive one. Caching $0.15, batch $0.75 / $4.50. |
gemini-3.1-flash-lite | $0.25 in / $1.50 out | Cheaper than 3.5 Flash-Lite on both sides. Input rate covers text, image and video. |
gemini-3.5-flash-lite | $0.30 in / $2.50 out | Still the Flash-Lite that is listed for computer use. |
gemini-3.8-live and gemini-3.8-live-extended-thinking | $0.75 in / $4.50 out for text | Free tier free of charge. Enterprise access is private preview. |
gemini-omni-1.1-flash | $1.50 in for text, image, video and audio / $9.00 out for text | No free tier at all, so a free-tier assumption carried over from the Flash text models fails here. |
gemini-3.5-transcribe | $2.00 or $0.003 per audio minute in / $12.00 or $0.002 per minute out | Batch transcription is the cheaper of the two IDs by a wide margin. |
gemini-3.5-transcribe-live | $3.50 or $0.005 per audio minute in / $21.00 or $0.004 per minute out | 75 percent more per 1M tokens than the batch model, and per audio minute 67 percent more on input and 100 percent more on output. Pick the ID on latency need, not habit. |
gemini-3.1-flash-image (Nano Banana 2) | $0.50 in for text and image / $60.00 out for images | Image output is two orders of magnitude above text output, and there is no free tier. |
gemini-3.1-flash-lite-image (Nano Banana 2 Lite) | $0.25 in / $30.00 out for images | The cheap image tier is still expensive next to text. |
gemini-3-pro-image (Nano Banana Pro) | $2.00 in / $120.00 out for images | Model an image pipeline per image, never per token intuition. |
| Search grounding | 5,000 free requests per month, then $14 per 1,000 | The free allowance is shared across all Gemini 3.x models, so it does not multiply when you add one. |
| Shutdown date | Model | Replacement named by Google |
|---|---|---|
| August 31, 2026 (passed, unconfirmed at eighteen days) | gemini-robotics-er-1.6-preview | gemini-robotics-er-2-preview, though three Google pages still present the 1.6 preview as live |
| September 30, 2026 | gemini-omni-flash-preview | gemini-omni-1.1-flash |
| October 2, 2026 | gemini-2.5-flash-image | gemini-3.1-flash-image-preview in the deprecation table, while the models page lists gemini-3.1-flash-image as stable |
Gemini’s advantage is the stack around the model
Each surface brings a different kind of context. Search knows the live web. Workspace knows the work. Android knows the moment. And the API turns the model into a building block. September added voice models, a music model, a new coding agent and two enterprise governance previews, while the consumer Gemini Drop cadence has now missed two consecutive months. This is also where a buyer meets the consumer plans, which decide which model an individual account gets, and how much of it.
Gemini app
Conversation, connected apps, multimodal input, Daily Brief and emerging personal agents. What the account can reach depends on its AI plan.
Search AI Mode
Live web grounding, information agents, Maps and custom generative interfaces.
Google Workspace
Gmail, Docs, Slides, Sheets, Meet and organizational content, with Gemini 3.8 Live Extended Thinking rolling out to Workspace and Gemini Live subscribers in Docs, Gmail and Keep.
Gemini Notebook
Source-grounded notebooks, audio and video overviews, reports and study tools. Renamed from NotebookLM on July 16, 2026, with a secure cloud computer for code execution and sync with the Gemini app.
Gemini API
Model access, built-in tools, computer use, embeddings and application-specific agents.
Antigravity
Agent-first development with artifacts, plans, screenshots and execution evidence. The default agent changed on September 17, 2026.
Android
Camera, voice, notifications, connected apps and contextual assistance close to the moment of work. At Galaxy Unpacked on July 22, 2026 Google said Gemini Intelligence task automation now reaches more than 40 apps, with Gemini Notebook preinstalled on the Galaxy Z Fold8 and Gemini on the Galaxy Watch 9.
Omni, Veo, Lyria and image tools
Image, video and music creation connected to Gemini reasoning and the wider Google ecosystem.
A new Antigravity agent, with a breaking tool contract
Google’s API release notes record antigravity-preview-09-2026 replacing the May 2026 version. Built-in tools moved to PascalCase parameters, and file edits became line-range replacements instead of full rewrites. That is a breaking interface change to the agent ID a managed-agent integration pins, and it lands on an agent that had survived two Flash launches unchanged. This item appears only in the release notes, which are rewritten in place. No dated article on Google’s blogs, DeepMind or the developer blog carries it. So treat it as real but unanchored, and confirm it directly against the release notes before you plan a migration.
Gemini 3.8 Live reached GA
Two audio-to-audio models, gemini-3.8-live for low latency and gemini-3.8-live-extended-thinking for reasoning, generally available through the Gemini API and Google AI Studio, with Search Live for the low-latency model and private preview for enterprises. The same release marks gemini-3.1-flash-live-preview as a legacy Live API preview and recommends moving to 3.8 Live.
Lyria 3.5 reached GA
A full-length song generation model with, in Google’s words, “improved musical coherence, natural vocals, and fine-grained duration and structural control,” generating 44.1 kHz stereo audio from text and image inputs. lyria-3.5 is listed as stable and it shipped into the Gemini app. One caution on the date. September 3 comes from Google’s API release notes, and the announcement page didn’t surface a publication date when it was read. Confirm the date directly rather than quoting it as settled.
Agentic video understanding
Google introduced agentic video understanding for Gemini 3.7 Flash, Gemini 3.6 Flash and Gemini 3.5 Flash-Lite. Google claims up to 88% fewer tokens, 66% lower cost and 7% better accuracy, at no extra charge. Those are Google’s figures from Google’s own announcement, not an independent measurement. Quote them as vendor claims, and test them on your own footage before you rebuild a pipeline around the savings. The flagship 3.8 Flash is still not named, and nothing published between September 3 and September 18 revisits that.
Gemini Omni Flash
The model ID is gemini-omni-1.1-flash. It adds video extension, interpolation and configurable resolution at 360p, 720p, 1080p and 4k. It is priced on the paid standard tier at $1.50 per 1M input and $9.00 per 1M output, with no free tier at all. Google’s own pages give four different answers about its status. The models page lists it under a preview heading. DeepMind lists it as GA. The release notes headline says GA. And the dated announcement says neither, calling it “production-ready” and “rolling out across the Google developer ecosystem.” The preview endpoint it replaces, gemini-omni-flash-preview, is the model scheduled to shut down on September 30, 2026.
Gemini 3.5 Transcribe
Two model IDs, gemini-3.5-transcribe and gemini-3.5-transcribe-live, with support for 85 or more languages, speaker diarization and custom vocabulary biasing. The live model costs 75 percent more per 1M tokens than the batch model, and between 67 and 100 percent more per audio minute depending on direction. So the choice between them is a budget decision as well as a latency one.
Gemini Embedding 2
gemini-embedding-2-preview is Google’s “first multimodal embedding model, mapping text, images, video, audio, and PDFs into a unified embedding space.” DeepMind describes it as generally available while the API models page lists it as a preview, which is the same GA-or-preview split that affects Omni. A unified embedding space across five modalities is a real retrieval capability. A preview ID is still a preview ID.
The August Workspace Feature Drop
Screenshots in Meet notes, automatic organization of meeting artifacts in Drive, quiz generation in Forms, Workspace Studio skills invoked by @mention, and enterprise security controls in Studio Flows. Workspace features arrive by edition and admin policy, so confirm eligibility before you describe any of these as available to your organization. It is still the newest Workspace drop as of September 18, 2026.
Two consecutive months with no Gemini Drop
The July 31, 2026 drop is still the most recent one. Neither an August nor a September edition exists. This check covered two distinct URL patterns, because the drops have migrated between /products-and-platforms/products/gemini/ and /innovation-and-ai/products/gemini-app/. Both patterns 404 for both months, and the Gemini app blog index lists no successor. So the ordinary explanation, that the page simply moved, is ruled out. Android and Pixel drops did ship in September, which makes the missing cadence specific to Gemini rather than a company-wide pause.
New connected apps and services
Google added Granola, Otter.ai, Wix, OpenTable (UK), Ticketmaster, Zocdoc and others to the Gemini app’s connected apps on August 12, 2026. Every connector widens what the assistant may read and act on. So treat the list as a permissions decision, not a feature announcement.
| Surface | Context it owns | Best outcome | Availability caution |
|---|---|---|---|
| Gemini app | Personal conversation and connected apps | Everyday help and proactive assistance | Features vary by plan, language and region; the AI plan decides which Pro model and how much of it the account reaches |
| Search AI Mode | Live web, shopping and Maps | Grounded answer or interactive interface | Generative UI and agents roll out over time |
| Workspace | Organizational mail, files and meetings | Native work artifact | Edition and admin policy matter; the August 26 feature drop is still the current reference point |
| Voice and Live | Spoken conversation in real time | Production voice agents and hands-free help | Generally available through the API and AI Studio since September 15, 2026; enterprise access is private preview |
| Gemini Notebook | Curated user source set | Source-grounded learning and briefing | The cloud computer is on AI Ultra and Workspace Expanded Access; Google’s July 16, 2026 post promised Pro on the web in coming weeks, still not confirmed as shipped as of the September 18, 2026 check |
| API / Antigravity | Developer-defined tools and environment | Custom agent or software workflow | Requires safety and permission design; the default Antigravity agent and its tool contract changed on September 17, 2026 |
| Android | Device, camera and contextual moment | Personal assistance and action | Hardware and rollout vary |
| Plan | What Google lists in it | Price |
|---|---|---|
| Google AI Plus | 400 GB storage, 2x access to Gemini, Gemini 3.1 Pro, Deep Research, image, music and video generation, Gemini Notebook, family sharing | Not readable from a static fetch of Google’s plans page: the amount renders client-side. Confirm in a browser before you quote it. |
| Google AI Pro | 5 TB storage, 4x access to Gemini, expanded Gemini 3.1 Pro, Deep Research, Gemini Spark, Google Flow, $10 a month in Google Cloud credits, Google Home Premium Standard | $19.99 per month, the only consumer figure with first-party support, stated in Google’s August 19, 2026 student-offer post |
| Google AI Ultra | From 20 TB storage, up to 20x access to Gemini, highest Gemini 3.1 Pro limits, Deep Think, Project Genie, YouTube Premium individual, $40 a month in Google Cloud credits, Google Home Premium Advanced | Not readable from a static fetch, and the rendered tiering was ambiguous when checked. Confirm in a browser before you quote it. |
Google is turning Search and the device into persistent workers
The Gemini 3.5 through 3.8 line is built to sustain work, not just answer prompts. Spark, background information agents, voice agents and computer use are Google trying to make intelligence persistent across time and interfaces. Most of those surfaces are still limited, regionally excluded or unconfirmed. September added the first first-party governance layer aimed at watching agents while they run.
Gemini Spark Not in the EEA, Nigeria, Switzerland or the UK
A personal agent meant to help across the digital day. The July Gemini Drop framed the release as “Gemini Spark is going global... now available worldwide.” Google’s support page states the real list: “Available wherever Gemini Apps are supported, except in the European Economic Area, Nigeria, Switzerland and the United Kingdom.” It is also gated to Google AI Pro and above. So read the exclusion list and the plan requirement, not the headline.
- “Worldwide” with four named exclusions, unchanged for seven weeks
- Unavailable to readers in the UK, the EEA, Switzerland and Nigeria
- Inside supported regions it still needs a Pro or Ultra plan
Computer use Now on 3.8 Flash
Announced June 24, 2026 on Gemini 3.5 Flash, and extended since. As of the September 17, 2026 docs stamp, Gemini 3.8 Flash is the recommended model, with 3.7 Flash, 3.5 Flash, 3.5 Flash-Lite, Gemini 3 Flash Preview and a legacy 2.5 preview also listed. Gemini 3.6 Flash is still not on the list. So an agent routed there can’t be assumed to have the capability.
- 3.8 Flash recommended, 3.7 Flash listed
- 3.6 Flash still absent after a further edit of the page
- Verify support against the ID you pin, every time
Managed agents and the new Antigravity default Changed September 17
A model swap shows up in a dashboard. A tool-contract change shows up as failing agents. This is the second kind. Google opened managed agents to free-tier projects on July 28, 2026: “Developers can experiment with agentic workflows using an API key from a project without active billing.” That post named antigravity-preview-05-2026 as running Gemini 3.6 Flash by default. That agent has now been replaced. Google’s API release notes, dated September 17, 2026, record antigravity-preview-09-2026 replacing the May 2026 version, with built-in tools moving to PascalCase parameters and file replacements expressed as line ranges rather than full rewrites. Both are breaking changes for any integration written against the old contract. This entry exists only in the release notes, which Google rewrites in place, so confirm it directly before you schedule the work.
antigravity-preview-09-2026is the current agent- PascalCase tool parameters and line-range file edits are breaking changes
- Release-notes-only evidence: verify on the page before you act
Watching the agent while it runs Private preview
Until this month the safe computer-use workflow above had no first-party Google control plane attached to it. Two posts on Google’s developer blog change that. On September 16, 2026 Google described Agent Anomaly Detection, now in private preview on the Gemini Enterprise Agent Platform. It is an oversight layer that analyzes OpenTelemetry traces to catch behavioral risks without affecting live request performance. On September 15 Google published guidance on building zero-trust AI agents that judge intent rather than syntax, which is runtime governance rather than prompt hygiene. Read both before an agent touches production. Neither is generally available.
- Trace-based oversight rather than prompt-level filtering
- Private preview, so do not scope a program around it yet
- Both posts are dated and cited in the sources section
Voice agents GA September 15
Gemini 3.8 Live and 3.8 Live Extended Thinking make production voice agents a supported build rather than a preview experiment. The low-latency model is aimed at conversational turn-taking. The extended-thinking model is for problems that need reasoning mid-call. Google calls them the building blocks for reliable, production-ready voice agents, which is a vendor claim, and the right one to test with a recorded call set.
- Latency and reasoning are now separate model choices
- Enterprise deployment is private preview
- Design the escalation path to a human before you launch
Agentic video understanding September 1
Google’s agentic video capability applies to Gemini 3.7 Flash, Gemini 3.6 Flash and Gemini 3.5 Flash-Lite, and Google says it costs nothing extra. Google claims up to 88% fewer tokens, 66% lower cost and 7% better accuracy. Note which models are named. The flagship 3.8 Flash isn’t among them, and Google has published nothing since to explain the exclusion.
- Vendor figures, attributed to Google
- Named for 3.7 Flash, 3.6 Flash and 3.5 Flash-Lite
- Validate the savings on your own video before repricing a pipeline
Information agents in Search
Agents can watch an ongoing information need and report what changed, with links for deeper action.
- Repeated research
- Alerts and monitoring
- Source review remains necessary
Daily Brief
The Gemini app can pull timely personal information into one proactive morning view.
- Reduce manual checking
- Depends on connected context
- Review privacy and relevance
Generative interfaces
Search can build a purpose-built interface or simulation instead of forcing every answer into prose. Google promised this for everyone in Search “this summer” on May 19, 2026. No later post confirms that it shipped, and the summer it named runs out on September 22, 2026.
- Promised, not confirmed
- Interactive decision support
- Generated logic still needs testing
Android Halo
A status-bar surface that shows agent activity, announced May 19, 2026 in preview and described as rolling out later this year. It is a visibility surface, not a management console. Nothing further has been published since May 19, as of the September 18, 2026 check.
- Agent activity, not agent control
- Announced in preview
- Rollout status matters
Gemini is joining reasoning and media production
Image output is charged at rates two orders of magnitude above text, and the image models have no free tier at all. That is the least intuitive pricing in the product, so start there. What the creative stack is good for is connecting a source-rich research process straight to image, video, music and presentation outputs, without shrinking creativity down to a single model name.
Nano Banana family Three IDs, three prices
Google’s image stack is three stable models with very different economics, all on the paid tier and none with a free tier. Nano Banana 2 is gemini-3.1-flash-image at $0.50 input for text and image and $60.00 output for images. Nano Banana 2 Lite is gemini-3.1-flash-lite-image at $0.25 and $30.00. Nano Banana Pro is gemini-3-pro-image at $2.00 and $120.00. Note one more thing. The deprecations page names gemini-3.1-flash-image-preview as the October 2 replacement for gemini-2.5-flash-image, while the models page carries gemini-3.1-flash-image as stable. Pin the stable ID.
- Cost image work per image, not by text-token intuition
- No free tier on any of the three
- Check text, likeness and rights in every generated asset
Gemini Omni Flash GA claimed, preview listed
The multimodal creation model is gemini-omni-1.1-flash. It adds video extension, interpolation and configurable resolution at 360p, 720p, 1080p and 4k. It is priced on the paid standard tier at $1.50 per 1M input and $9.00 per 1M output, with no free tier at all. Its status depends which Google page you read. DeepMind and the release notes say GA. The API models page lists it under a preview heading. The August 27, 2026 announcement says only that it is production-ready and rolling out. The preview endpoint it succeeds shuts down on September 30, 2026.
- Pin
gemini-omni-1.1-flash, not the preview - Say “Google describes it as production-ready,” not “it is GA”
- Anything still calling the preview has until September 30, a date Google publishes as the earliest possible one
Lyria 3.5 GA September 3
Google’s full-length song generation model, lyria-3.5, listed as stable and shipped into the Gemini app. Google describes improved musical coherence, natural vocals and fine-grained duration and structural control, generating 44.1 kHz stereo audio from text and image inputs. The September 3 date comes from the API release notes, and the announcement page didn’t surface a publication date when it was read, so treat the date as release-notes-sourced and confirm it directly. Music generation is also listed inside all three consumer AI plans.
- Text and image inputs, 44.1 kHz stereo output
- Rights clearance matters more for music than for most generated media
- Confirm the GA date directly before you cite it
Gemini 3.5 Transcribe GA August 26
Speech-to-text in two forms, gemini-3.5-transcribe and gemini-3.5-transcribe-live, covering 85 or more languages with speaker diarization and custom vocabulary biasing. Vocabulary biasing is the feature that makes internal product names and people’s names survive a transcript. The live model is 75 percent more expensive per 1M tokens than the batch model. Per minute, the gap depends on direction: input is $0.005 against $0.003, which is 67 percent more, and output is $0.004 against $0.002, which is 100 percent more. So the choice between the two IDs has a number attached to it.
- Separate IDs for batch and live, with a real price gap
- Diarization for multi-speaker recordings
- Bias the vocabulary before you judge accuracy
Veo and Flow
Google’s video stack supports cinematic clips and scene-oriented workflows, not just a single generated shot. Google Flow is listed inside the Google AI Pro plan and above.
- Storyboard the intent
- Use references and continuity
- Plan for post-production
Slides and Workspace
Gemini can help move research and narrative into presentation form inside the same organizational environment. The August 26 Workspace Feature Drop adds Meet notes screenshots, automatic meeting-artifact organization in Drive and Forms quiz generation, and remains the most recent drop.
- Use brand templates
- Link claims to sources
- Confirm edition eligibility for each drop feature
Gemini Notebook briefing outputs
Audio, video, reports, quizzes and other outputs turn a curated source set into several learning formats. Google reports more than 30 million users and over 600,000 organizations on the product.
- Source-grounded transformation
- Useful for enablement
- Keep audience needs explicit
Google Vids July 16
Vids gained Gemini Omni clip generation, and personal avatars built from a selfie and a voice sample, watermarked with SynthID. Google lists it for Google AI Pro and Ultra, plus Workspace business editions.
- Check plan eligibility first
- Get written consent for any likeness
- Disclose synthetic presenters
Generative UI
Custom charts, tools and simulations are a new kind of creative output for questions that need interaction. Google promised general Search availability “this summer” on May 19. Treat it as promised, not shipped.
- Explain complex systems
- Let users explore variables
- Test the generated behavior
Using the research notebook and brand system, create three campaign territories. For each, provide the strategic idea, hero image direction, six-second video beat and the evidence that makes the concept credible.
Turn this Gemini Notebook source set into an executive presentation. Keep one claim per slide, cite the exact source behind each claim and use our approved Slides theme.
Build an interactive comparison tool for these options using current Search and Maps data. Let the user change budget, distance and priority, and show how the recommendation changes.
Inspect every generated visual for inaccurate text, impossible product details, misleading data or implied claims that the source material does not support.
Ten habits for using Gemini as an ecosystem
Pick the surface that already owns the context. Pin what you build on. Cost the tier the work will actually run on. And keep the evidence attached as the work moves toward action. That is the whole behavior, right?
Start in the context owner
Use Search, Workspace, Gemini Notebook, Android or the API, depending on where the evidence already lives.
Ask for grounding
Use live Search and Maps when current external reality matters. The free grounding allowance is shared across Gemini 3.x models, so budget it once.
Curate when authority matters
Use Gemini Notebook or a controlled source set instead of the whole web.
Use the actual modality
Give Gemini the image, the video, the voice or the screen state instead of describing it badly. Voice is a first-class production build now, not a demo.
Separate prepare from act
Let Gemini draft first. Require approval before anything consequential happens outside.
Name the surface and the plan
Availability differs across app, Search, Workspace, API, region and subscription, and the consumer plan decides which Pro model and how much of it an account reaches.
Cost the tier, not the headline
Say whether a figure is free tier, paid standard or batch. Batch is exactly half of standard, so a batch workload costed at the standard rate overstates spend by 2x.
Preserve source links
Keep live information attached as it moves into a document, deck or decision.
Discount the superlatives
Google has now called three Flash generations its most intelligent workhorse model. Compare model cards, prices and your own traces instead.
Pin the model, not the alias
Name gemini-3.8-flash or another explicit ID, and check a named replacement against the models page before you adopt it. Aliases move on Google’s schedule, not yours.
An ecosystem this broad is easy to overstate
Three mistakes, in the order people make them. Turning a limited preview into a universal capability statement. Trusting a documentation page to describe the present. And repeating an absence after the vendor has quietly filled it. Gemini features arrive across products, plans and regions at different times, which is what makes all three so easy.
An absence is a claim too
Nobody re-checks a negative. That is what makes it the dangerous kind. Google’s API model page carries no knowledge-cutoff field for Gemini 3.8 Flash at all, which reads like an unpublished specification. The DeepMind model card published March 2026 on the day the model shipped. A statement that something doesn’t exist expires exactly like a statement that it does. So read the model card and the API reference before you report that a specification is missing.
The vendor can contradict itself
March 2026 on the model card. No cutoff field at all on the API model page. GA on DeepMind and in the release notes for Omni, a preview heading on the models page. Two different Nano Banana 2 IDs on two pages. Name both surfaces, say which one you trusted and why, and prefer the dated specification document.
Rollout fragmentation
Plan, region, language, hardware and Workspace edition can each change access. And the consumer plan tier now changes which Pro model an account can reach, not just how often.
Prompt counts are the wrong unit
Google moved the consumer plans from daily prompt limits to a compute-used model, where limits factor in prompt complexity, the features used and chat length. Any plan comparison built on prompts per day is measuring something Google no longer sells.
Quote the tier with the number
Every published Gemini rate belongs to a tier. Free tier is free of charge on the Flash text models, and not available on the image models or Omni. Batch is exactly half of paid standard. A batch workload costed at standard overstates spend by 2x.
The cheapest model moved
gemini-3.1-flash-lite, at a paid-tier standard $0.25 per 1M input and $1.50 per 1M output, undercuts gemini-3.5-flash-lite on both sides. A low-cost tier recommendation made three months ago is a price claim, right? And price claims decay.
“Legacy” is a signal
Google’s models page now calls Gemini 3.5 Flash its legacy Flash model, while 3.6 Flash is merely previous-generation. No shutdown date has been published. But legacy is the word that comes before one, and 3.5 Flash is also the expensive Flash. Plan the move before the notice arrives.
Grounding errors
A cited Search result can still be misunderstood, stale or weaker than the claim.
Computer-use risk
Interfaces can carry malicious instructions, unexpected prompts and sensitive information. Support also varies by model: 3.8 Flash is recommended, 3.6 Flash is not listed at all. And Google’s own runtime oversight tooling, Agent Anomaly Detection, was still private preview as of its September 16, 2026 announcement.
An agent contract can break under you
Google replaced the default Antigravity agent on September 17, 2026, moving built-in tool parameters to PascalCase and file edits to line ranges. A model swap is visible in a dashboard. A tool-contract change shows up as failing agents. So read the agent release notes on the same schedule you read the deprecations table.
Background-agent drift
Long-running monitoring needs a clear objective, notification policy and stop condition.
Gemini now sits inside a competitor’s stack
Apple describes its next-generation Foundation Models as “custom-built in collaboration with Google and its Gemini models.” That places Gemini inside the default model layer of another vendor’s first-party assistant, and the sibling Apple guide records that single sentence as the entire first-party account of the arrangement. It changes nothing in the model and price tables above. But a client may already be reaching Gemini through Apple rather than through Google, so ask where the model runs before you scope a Google agreement.
Personal versus work context
Connected apps and device context can cross boundaries employees or administrators did not intend.
Generated media truth
Visual quality doesn’t guarantee accurate text, products, scenes or implied evidence. Image and music output also carry rights questions that a text answer does not.
Product-name churn
Build guidance around durable workflows, with dated model notes, rather than temporary launch labels. NotebookLM became Gemini Notebook on July 16, 2026.
The superlative is a constant
Google called 3.6 Flash, then 3.7 Flash, then 3.8 Flash a version of “our most intelligent workhorse model.” A phrase reused across three generations can’t distinguish any of them. Take the differences from the model card and the price table, and never let a vendor adjective carry a recommendation.
Announced is not available
Spark, Gemini 3.5 Pro, the cyber models, enterprise Live access and generative UI all have to keep their actual rollout state attached. Spark is the sharpest case. Google’s support page, re-read September 18, 2026, still excludes the European Economic Area, Nigeria, Switzerland and the United Kingdom, and Spark also needs a Pro plan or above.
Earliest possible, not final
The deprecations page says its shutdown dates “indicate the earliest possible dates on which a model might be retired,” with advance notice before actual discontinuation. Plan against the published date, but describe it the way Google does rather than promising a hard deadline Google has not promised.
A shutdown date can be withdrawn
In early August Google’s deprecations page dated the whole Gemini 2.5 family to October 16, 2026, then removed it. Re-checked September 18, across a September 16 revision of the page, all three models still read “No shutdown date announced.” Re-read the table before every publication. A date you recorded once can move in either direction.
A shutdown date can also pass unconfirmed
gemini-robotics-er-1.6-preview was scheduled to shut down on August 31, 2026. Eighteen days later the deprecations page still lists it as a future shutdown. The models page still lists it with no deprecation mark. And the robotics overview, edited on September 4, still says the model “will be shut down at the end of August.” Nothing confirms it executed. So report the passed date and the disagreement, not an outcome.
A migration target can be on a fuse
The rewritten Imagen page tells users to move to gemini-2.5-flash-image. The deprecations page schedules that model to shut down on October 2, 2026, fifteen days after the advice was published. Check every named replacement against the shutdown table, including the ones a fixed page gives you.
Documentation lags its own product
The Imagen page spent weeks describing an August 17 shutdown in the future tense before Google rewrote it on September 17. A reference page describes the moment it was written. When it disagrees with a dated announcement, the announcement wins, and a corrected page is only as current as its correction.
Whose benchmark is it
Google attributes the 17 percent output-token reduction for 3.6 Flash to the Artificial Analysis Index. The 88 percent token, 66 percent cost and 7 percent accuracy figures for agentic video are Google’s own. So are the multipliers advertised on the consumer plans, which are measured against non-AI subscribers. Keep each attribution attached to the party that produced it.
Alias drift
No first-party Google page documents which model gemini-flash-latest serves today. The last documented switch in the release notes remains January 21, 2026, to gemini-3-flash-preview, re-confirmed September 18, and Google’s stated policy is a hot swap with two weeks of email notice. So pin an explicit model ID. The Grok guide records the same failure as completed history: on August 5, 2026 grok-voice-latest rerouted to a model costing 60 percent more per minute of audio, with no action by the customer.
A missing cadence is information
There has been no Gemini Drop for two consecutive months. July 31, 2026 remains the most recent edition, and both candidate URL patterns 404 for August and September. That matters because the drops have migrated between paths before. Android and Pixel drops did ship in September, so this gap is specific to Gemini. Don’t present the drop as a reliable monthly channel.
Slipped timelines, plural
Gemini 3.5 Pro was described on May 19, 2026 as rolling out the next month, and on July 21 as testing with partners. It is now four months past that window and still absent from the models and pricing pages, while DeepMind still shows a “coming soon” line. Gemini 3.1 Pro shipped in preview on February 19, 2026 with general availability promised soon, and it is still a preview endpoint seven months later. Two slipped Pro timelines is a pattern, not an accident.
| Claim type | Before publishing | Safe wording |
|---|---|---|
| Model default | Check the official model release page | “Default in the Gemini app as of…” |
| Model identifier | Check the release notes for the last documented alias change, and check a named replacement against the models page | Name the pinned model ID and its GA date, never an alias |
| A published absence | Re-check the model card before repeating that Google has not published something | Say what the card publishes and on what date you read it |
| Price | Identify the tier: free, paid standard or batch, and the context tier for Pro | Name the tier in the same sentence as the number |
| Consumer plan | Confirm the amount in a rendered browser; the plans page renders prices client-side | Quote only a price a dated Google article states, and say what is unconfirmed |
| Vendor superlative | Check whether the same phrase covered the previous model | Quote it with the date and the model it was written about |
| New agent | Check plan, region, tester status and the agent tool contract | “Rolling out to…” or “available to…” |
| Workspace feature | Check edition and admin controls | Name the Workspace surface and eligibility |
| Computer use | Check the supported-model list, not the original announcement | Name the model IDs the documentation actually lists |
| Creative capability | Check model, product and output availability, and the per-image rate | Separate announced model from shipping user feature |
| Model lifespan | Re-read the deprecations page rather than trusting a date you recorded earlier | Give the shutdown date as the earliest possible date, the named replacement and the date you read the page |
| An executed shutdown | Look for a page that says it happened, not one that says it will | Say the date passed and name the pages that still disagree |
| Regional availability | Read the exclusion list, not the launch framing | Name the excluded regions in the same sentence as the launch |
First-party evidence behind this guide
All but two of these Google announcements carry a full publication date. The Gemini 3.7 Flash model card carries no date at all. The Gemini 3.5 Transcribe announcement is stamped August 2026 with no day. Treat those two as undated first-party pages rather than dated citations. Google’s model lineup moves across the app, Search and the API at different speeds, so a claim that is true of one surface is not automatically true of another. And a large part of what a buyer needs from Gemini lives on pages Google rewrites in place. Those can’t anchor a dated claim, so they are not listed below. Each one is attributed inline with the date it was read, and anyone relying on one should confirm it directly. Here is the list. The paid-tier standard, free-tier and batch prices, and every figure in the price table, from the pricing page. The shutdown dates, the named replacements and the “earliest possible dates” caveat, from the deprecations page, read September 18 with a September 16 stamp. The computer-use model list, read September 18 with a September 17 stamp. The models-page descriptions, including Gemini 3.5 Flash as “legacy,” the stable listings for gemini-3.1-flash-lite, gemini-3.1-flash-image and lyria-3.5, and the preview listings for gemini-omni-1.1-flash and gemini-embedding-2-preview. The rewritten Imagen page and the robotics overview. The Spark availability sentence on Google’s support page. The consumer plan contents on Google’s AI plans page, whose prices render client-side and could not be pinned from a static fetch. The DeepMind model index that still carries “3.5 Pro coming soon” and lists Gemini 3.1 Deep Think. And four items exist only in the API release notes. The September 17, 2026 replacement of the Antigravity agent by antigravity-preview-09-2026. The September 3 general availability of Lyria 3.5. The August general availability dates for gemini-omni-1.1-flash and the Gemini 3.5 Transcribe models. And the January 21, 2026 switch of gemini-flash-latest to gemini-3-flash-preview. One more. The Lyria 3.5 announcement page is named inline rather than listed. It didn’t surface a publication date when it was read, so it can’t serve as a dated citation.
AI Mindset