Google Antigravity Models: List, Limits, Switching (2026)

As of October 8, 2026, Google Antigravity offers nine reasoning models across three families — four Gemini models, four Claude models, one open-weight GPT model — but they are not spread evenly across plans, and three of them carry a removal date. The official model matrix puts a hard November 2, 2026 on Claude Sonnet 4.6, Claude Opus 4.6, and GPT-OSS-120b. Layer on the consumer Gemini shakeout reported for tomorrow, October 9, and "which models can I actually use in Antigravity" has become a genuinely confusing question to answer from search results.

Below: the full model matrix transcribed from Google's official docs, how the five-hour and weekly quota pools work, exact steps for checking and switching models, what the October 9 changes do and don't touch, a troubleshooting list for the "can't change model" reports, and a clear line between what we verified and what we couldn't.

Quick answer

• Which models does Antigravity offer? Four Gemini models (3.8 Flash, 3.7 Flash, 3.6 Flash, 3.1 Pro) on every plan including free; Claude Sonnet 5.5 and Opus 5.5 on Google AI Pro (non-trial) and Ultra; Claude Sonnet 4.6, Opus 4.6, and GPT-OSS-120b on Free, AI Plus, and Pro. Enterprise gets the Gemini lineup only.
• What's being removed, and when? Claude Sonnet 4.6, Claude Opus 4.6, and GPT-OSS-120b — the docs mark all three "will be removed on November 2, 2026."
• How do I check which model I'm using? The model selector under the conversation prompt box shows the current selection; /usage shows your quota pools.
• How do I switch? Pick another model in the same selector. It takes effect on your next turn — a running agent finishes its current steps on the model it started with.
• How do limits work? Two separate pools — one for Gemini models, one for Claude and GPT models — each with a five-hour limit and a weekly limit. Free-tier quota refreshes weekly.
• What happens October 9? Confirmed: the consumer Gemini app drops Flash and Pro from the free tier (Flash-Lite remains) and AI Plus loses Gemini 3.1 Pro. The Antigravity-specific version of this rumor is single-source and contradicted by the official docs as of today — see the dedicated section.
• Are there published prompt counts? No. Google publishes no numeric caps — limits "correlate with the amount of work done by the agent" and are "subject to modification."

Which models Antigravity offers

Transcribed from the official models page at antigravity.google/docs/models, checked October 8, 2026:

Three readings worth spelling out:

• Gemini is the only family on every plan. The plans page names Gemini 3.1 Pro and 3.8 Flash as part of the baseline quota every user gets, alongside unlimited Tab completions. With no paid plan at all, Gemini is what you have.
• Claude splits into two generations with opposite shapes. The newer 5.5 pair serves paid tiers only (and not AI Pro trials), while the older 4.6 pair — the one being removed November 2 — is the only Claude on Free and AI Plus. After that date, those tiers will have no Claude models at all, based on what Google documents today.
• Enterprise is Gemini-only for agents. Every third-party model shows No on Enterprise, with one documented exception: since v2.12.2 (September 3, 2026), Enterprise users can select 3.8 Flash "using ADC" — application default credentials, not the consumer picker.

Image tasks (generating a UI mockup, creating system or architecture diagrams) route to Nano Banana 2, which is not selectable in the model picker. There's also a thinking-effort axis: Gemini models offer Low/Medium/High, the Flash trio shows Medium/Fast options, and GPT-OSS-120b shows Medium — and effort changes how much quota a run burns, since limits "correlate with the amount of work done by the agent."

How to check which model you're using

Two places, both documented:

1. The model selector — the menu directly under the conversation prompt box. It shows the model and effort level the current conversation is set to. The selection is per-conversation, not global.
2. /usage — the slash command for model quotas. It opens a usage panel with "Weekly Limit Remaining" and "Five Hour Limit Remaining" for Gemini models, and a separate pair of the same meters for Claude and GPT models.

That split matters when something stops working: if your Claude pool is empty but Gemini shows headroom, the problem is quota family, not Antigravity as a whole.

How to change models

1. Click the model selector under the conversation prompt box.
2. Pick the model — and the effort level, where offered.
3. Keep talking. The new model handles the next turn.

The documented catch: a mid-run switch doesn't interrupt the agent — per the docs, it "continues to use the previously selected reasoning model until it completes its steps for that turn." If the output still looks like the old model, nothing is broken; finish or stop the current turn and the switch lands.

Models your plan doesn't include simply don't appear as valid options — see the troubleshooting section for the tier gates.

Rate limits: how Antigravity quotas actually work

Google publishes no numbers — no prompt counts, no token caps. What the plans page documents is the shape:

The mechanics around it:

• Two pools, not one. The usage panel tracks Gemini models separately from Claude and GPT models. Draining your Claude quota leaves your Gemini quota untouched, and vice versa.
• Quota scales with work, not prompts. A simple task yields more prompts; a long agentic run consumes quota faster. High thinking effort burns more.
• Everything is "subject to modification." Limits are primarily determined by available capacity and exist to prevent abuse — which is why someone else's "I get X prompts a week" anecdote predicts nothing about your account.
• Overages exist for paid plans. Pro and Ultra users can let runs continue past baseline on AI credits, billed at "standard Gemini Enterprise consumption pricing." The AI Credit Overages setting is a binary — Never or Always. Check it before a long unattended run.

One reference point if you're comparing on price: Gemini 3.8 Flash's API rate is $0.75 per million input tokens and $3.75 per million output tokens (introductory through December 31, 2026, then $1.50/$7.50). Inside Antigravity you spend quota pools instead — a different kind of bill. Our Claude Code pricing analysis does the same math for Anthropic's side if you're weighing a move between tools.

The October 9 changes: what's confirmed and what isn't

Tomorrow's date is doing a lot of work in search results, so let's separate the two claims.

Confirmed — the consumer Gemini app (multi-source: 9to5Google, October 3; Engadget, October 6; also covered by CNET and The Verge). Effective October 9, 2026:

Engadget's report — based on Google's Gemini support page — adds that affected users get an email with their specific effective date, so timing can vary slightly per account.

Not confirmed — the Antigravity version of this story. Community posts (a Reddit thread, relayed by secondary outlets) claim that on October 9 Antigravity drops Gemini 3.6/3.7 Flash for free users, pushes everyone to 3.8 Flash, and adds Claude 5.5. As of October 8, none of that is supported by Google's own surfaces: the official matrix still lists 3.6 and 3.7 Flash for Free and AI Plus, the changelog has no October 9 entry and no Claude addition entry, and Claude 5.5 has been in the matrix since well before this week. We treat the claim as unverified — the section below lists exactly what we checked.

One nuance from the consumer side: a Google support-forum discussion argues free users lose standard access to Flash and Pro rather than every trace of it — occasional openings rather than a hard wall. We can't verify Google's enforcement either way; watch for the email Google said it would send.

Gemini 3.8 Flash vs Claude Opus 5.5 in Antigravity

The two models searchers most often compare sit in different pools, tiers, and price classes:

Practical selection logic: default to 3.8 Flash for everyday agent work — it's in the baseline quota for every plan, and Google's September 2 announcement frames it as doing "extra reasoning steps and iterative tool calls" at 3.7's speed and price. Reach for Claude Opus 5.5 when a task is genuinely stuck — it draws from the smaller, separately-tracked Claude pool that free and AI Plus users can't touch at all. Because the pools bill separately, mixing them is the quota-efficient strategy: burn Gemini on breadth, save Claude for depth.

Note what this table refuses to do: put a benchmark number on Opus 5.5. No official source we could verify publishes comparable figures for it in this context, and we won't borrow one from a chart measuring something else.

Can't change models? Work through this list

Most "can't change model" reports trace to one of these documented behaviors:

1. You switched, but the agent still acts like the old model. Mid-run switches don't apply — the agent "continues to use the previously selected reasoning model until it completes its steps for that turn." Stop the turn or let it finish; the new selection governs the next prompt.
2. Claude Sonnet 5.5 / Opus 5.5 don't appear at all. You're on Free or AI Plus (not in your tier), on an AI Pro trial — Claude 5.5 is restricted to "non-trial subscriptions only" — or on Enterprise, where third-party models are gated off entirely.
3. Claude 4.6 or GPT-OSS-120b vanished (or is about to). Ultra and Enterprise never had the Claude 4.6 pair, and all three starred models are documented for removal on November 2, 2026. If they disappear before that date, check the models page first: availability is "subject to modification" without a changelog entry.
4. The model is there but runs fail or stall. Check /usage. The five-hour pool refills on a rolling basis (Pro and Ultra); the weekly pool does not, and free-tier users live entirely on the weekly refresh. The pools are per family — a drained Claude pool coexists with a full Gemini pool.
5. Options look grayed out mid-session. Selection persists per conversation; a conversation pinned to a model you've since lost access to (tier change, trial ending, removal date passing) won't follow you to the new pick until you change it in the selector.
6. You're on Enterprise and missing 3.8 Flash. Since v2.12.2 (September 3, 2026), Enterprise gets 3.8 Flash via ADC rather than the consumer picker — an auth-setup matter, not a picker bug.

Using Antigravity models outside the IDE

Two routes, one official, one not:

The Antigravity CLI — the documented way. Antigravity ships as a desktop app, a CLI (the agy binary), and IDE extensions for VS Code, Visual Studio, JetBrains, Zed, and Xcode. The CLI shares the same account and quota pools, and the plans page lists it among the features the baseline quota covers. Our Antigravity CLI tutorial walks the flow end to end.

The OpenCode plugin — the unofficial way, with real risk. A community plugin, opencode-antigravity-auth, authenticates OpenCode against Antigravity via Google OAuth (opencode auth login, browser flow on localhost port 51121) and exposes model IDs like antigravity-gemini-3.1-pro and antigravity-claude-opus-4-6-thinking, with multi-account rotation when rate-limited. Two things to know before installing: the README itself warns that "using this plugin (and any proxy for Antigravity) violates Google's Terms of Service," with users reporting banned or shadow-banned Google accounts — and the repository was archived on August 27, 2026, so it no longer tracks the current model catalog. Our OpenCode tutorial covers the supported ways to configure that tool.

What we couldn't verify

• The October 9 Antigravity model change. The claim that free users lose 3.6/3.7 Flash tomorrow and get pushed to 3.8 Flash, with Claude 5.5 newly added, comes from a Reddit thread and secondary retellings. Against it: the official matrix (checked October 8) still lists both Flash models for Free and AI Plus, the changelog contains no such entry, and Claude 5.5 predates the rumor. If it materializes, it will show on the docs pages, not in a Reddit post.
• Any numeric quota. No official page publishes prompt counts or token caps for any tier. Everything quoting specific numbers is anecdote.
• A fallback model on exhaustion. The docs describe the pools and refills but never say what happens the moment a pool empties mid-run — pause, downgrade, or error. Undocumented.
• Which Flash-Lite version survives October 9. Consumer-side coverage says "Flash-Lite" without pinning the generation; we won't guess.
• Whether Antigravity is affected by the consumer change at all. A headline snippet suggests one outlet's paywall piece mentions AI Studio and Antigravity, but the article itself is inaccessible to us, so we can't check what it actually says.

FAQ

What models does Google Antigravity offer?

As of October 8, 2026: Gemini 3.8 Flash, 3.7 Flash, 3.6 Flash, and 3.1 Pro on every plan; Claude Sonnet 5.5 and Opus 5.5 on AI Pro (non-trial) and Ultra; Claude Sonnet 4.6, Opus 4.6, and GPT-OSS-120b on Free/AI Plus/Pro. Image tasks route to Nano Banana 2. The last three are scheduled for removal on November 2, 2026.

Is Gemini 3.8 Flash free in Antigravity?

Yes — it's part of the baseline quota every user gets, and the plans page names it alongside 3.1 Pro as a core agent model. "Free" means quota-limited: a weekly-refreshed pool with no published cap, not unlimited usage.

How do I check which model an Antigravity agent is using?

Look at the model selector under the conversation prompt box — it shows the active model and effort level for that conversation. Run /usage to see both quota pools with their five-hour and weekly meters.

Why can't I change the model mid-task?

By design. A running agent "continues to use the previously selected reasoning model until it completes its steps for that turn." The switch applies from your next prompt. Stop the turn if you need the new model immediately.

Why don't I see Claude Opus 5.5 in the picker?

Three documented gates: Free and AI Plus plans don't include the Claude 5.5 pair; AI Pro trial subscriptions are excluded ("non-trial subscriptions only"); and Enterprise gates off all third-party models. On qualifying plans, the model simply appears in the selector.

Does the October 9 Gemini change affect Antigravity?

The confirmed October 9 change applies to the consumer Gemini app — free users keep only Flash-Lite, AI Plus loses 3.1 Pro. Antigravity has its own model matrix, which as of October 8 still lists the full Gemini lineup for free users. The Antigravity-specific version of the rumor is unverified; the only documented Antigravity removal date is November 2, 2026.

What models are leaving Antigravity, and when?

Claude Sonnet 4.6, Claude Opus 4.6, and GPT-OSS-120b — the official models page marks all three "will be removed on November 2, 2026." After that date, free and AI Plus users will have no Claude or GPT models in the documented matrix, only Gemini.

Related reading

• Antigravity plan mode tutorial — plan before you spend quota on the wrong build
• Antigravity skills tutorial — teach your agents repeatable workflows
• Antigravity YOLO mode tutorial — run agents unattended, safely
• Antigravity CLI tutorial — the same models in your terminal, on the same quota pools
• GPT-5.5 deprecated — the other October retirement wave, and its replacement maps