Which models can your Antigravity plan use?

This is the 2026-10-05 English official Models matrix, including its two footnotes. “Yes” means Google lists that plan as eligible; it does not measure your remaining quota or verify your account’s current selector. “Listed” means the reviewed table does not attach the November 2 removal notice to that row. Official matrix.

Claude 5.5 on Google AI Pro is for non-trial subscriptions only. Do not count a Pro trial as eligible. Ultra lists both 5.5 models; Free/Plus and Enterprise do not. Ultra does not list either 4.6 model. These restrictions are part of the matrix, not a recommendation to upgrade.

The status column repeats the removal footnote so it stays readable on mobile. Check the official page again near the deadline; availability can change.

AgentSkillsHub editorial guidance: use the plan-specific candidate sets below, then test your own workflow. They are eligible choices from the official matrix, not performance rankings or Google-designated replacements. The four Gemini choices listed today are Gemini 3.8 Flash, 3.7 Flash, 3.6 Flash and 3.1 Pro.

• Free / Google AI Plus: investigate those Gemini choices for work currently using 4.6 or GPT-OSS. Claude 5.5 is not listed for this plan group.
• Google AI Pro: the Gemini choices are listed. Non-trial subscriptions additionally list Sonnet 5.5 and Opus 5.5 (thinking). A trial does not meet that 5.5 exception; confirm trial status before relying on it.
• Google AI Ultra: Gemini and both Claude 5.5 choices are listed. GPT-OSS is scheduled to leave; the table already excludes both Claude 4.6 models for Ultra.
• Enterprise: the table lists Gemini, not Claude or GPT-OSS. Follow the actual enterprise account and supported client route, not a consumer subscription assumption.

There is no promised one-to-one replacement from Claude 4.6 to 5.5, nor a named automatic replacement for GPT-OSS in the reviewed guidance. A candidate must preserve the behavior your workflow needs: instructions, tool calls, code changes and output acceptance. Do not infer equivalence from a shared model family.

No plan purchase or upgrade is required by this editorial checklist. Review existing eligible choices first.

Google says the reasoning-model choice persists between messages in a conversation. When you change it during execution, the current turn continues using the previous selection until its steps finish or you cancel that execution. The change is not a mid-turn hot swap. Turn behavior.

Editorial operating rule: let the current turn settle, or deliberately cancel it and inspect work already done. Confirm the new selection before starting the next task. Do not launch duplicate work merely because a dropdown changed.

Post-removal saved-conversation behavior: UNKNOWN. The reviewed official pages do not specify how a saved selection of a removed model will be mapped, rejected or replaced after November 2. Do not assume automatic migration to Claude 5.5 or Gemini. Inventory important saved conversations now and inspect the active choice when reopening them.

For Implementation Plans, Review/Proceed and plan policy, use the separate Antigravity /plan guide. Those controls answer a different question from model eligibility.

The Plans page gives Pro a five-hour refresh until its weekly limit is reached, and Ultra a five-hour refresh plus weekly limits. Other consumer plans refresh weekly. Allowances depend on capacity and work performed and can change; this is not a fixed prompt-count promise.

The Models page illustrates separate “Gemini Models” and “Claude and GPT models” usage groups. That display does not establish every cross-model pooling rule or prove your account has the sample balance. Exact pool allocation and your remaining allowance are UNKNOWN until checked for the account. Do not infer that Claude 5.5 always consumes quota faster than Gemini, or that switching models resets a limit.

For Pro/Ultra, the Plans page describes credits above baseline and AI Credit Overages: Never / Always. Never waits for baseline refresh; Always permits credit use after exhaustion. A listed model can still be quota-limited. Our editorial recommendation is to inspect the current setting before testing, not to enable spending automatically.

The consumer Plans page’s no-BYOK statement concerns extra rate limits. Enterprise has a separately documented connection and billing path; do not merge the two contracts.

In the CLI, Google documents /usage (alias /quota) as refreshing model configuration and quota status from the backend, then opening the quota panel. The usage guide’s example model names are illustrative; the current Models matrix above owns the dated availability list. CLI usage reference.

AgentSkillsHub editorial checks:

1. Confirm the signed-in account and whether its plan is Free/Plus, Pro trial, non-trial Pro, Ultra or Enterprise.
2. Open the current model selector and record the choices actually offered. A model name in an old message is not proof of present access.
3. Refresh usage/configuration where your client supports it; record remaining quota and relevant refresh timing.
4. If the selector differs from the dated table, check the official notice and account/client route before diagnosing it as a product bug. Google notes that app releases roll out gradually; that is not proof of why one particular model is missing.
5. Preserve the observed discrepancy and ask the appropriate administrator or support channel when eligibility remains unclear. Do not invent a fallback or bypass an account restriction.

No account configuration was changed while preparing this guide.

Google documents Enterprise access through either a Cloud project/API with consumption billing or a Gemini Enterprise Standard/Plus license with included quotas and managed overages. The supported route includes Antigravity 2.0, CLI and IDE extensions; the standalone Antigravity IDE is not currently supported for enterprise deployment. Enterprise guide.

Editorial migration check: ask the administrator which account, license/project, client and overage policy your workflow uses. Record the Gemini choices actually available on that route. Do not apply consumer Pro non-trial access or its quota schedule to an Enterprise user by analogy.

No Enterprise account or billing configuration was tested here.

AgentSkillsHub editorial migration checklist — not a Google mandatory procedure. Use the dated removal notice as the trigger and your own acceptance criteria as the decision.

1. Inventory dependencies: find conversations, saved instructions and workflows naming Sonnet 4.6, Opus 4.6 or GPT-OSS-120b. Record the task owner and last known selection.
2. Confirm the account plan: include Pro trial status or Enterprise route. Do not infer it from a teammate’s account.
3. Open the current selector: record eligible destination candidates and any discrepancy against the official table.
4. Choose a test candidate: write down why it is eligible and what result the workflow must preserve; avoid declaring a universal “best model.”
5. Run a representative task: use a safe branch or synthetic input and an approved quota/spending boundary. This is a proposed test, not one we ran.
6. Compare actual results: behavior, tool use, code diff, latency and quota impact where observable. Missing measurements remain unknown. Review the diff and output rather than relying on a success message.
7. Update team instructions: replace removed-model names only after acceptance. Record the tested plan/model, unresolved differences and a fallback available to that same plan.
8. Prepare before the deadline: do not leave high-value workflows until November 2. Recheck Google’s notice and the account selector shortly before the change.

If no candidate passes, keep the failure evidence and arrange a manual path or a separate decision. A removal notice does not authorize a subscription change, broader permissions or automatic production actions.

EDITORIAL EXAMPLE — NOT RUN. A developer reopens a saved conversation whose runbook says to use Sonnet 4.6 for a small bug fix. Their current account is Free, so “switch to Claude 5.5” is not an eligible default response.

They record the branch revision and the bug’s acceptance check, inspect today’s selector, and choose one of the offered Gemini candidates for a synthetic reproduction. They compare the proposed code diff, tests, tool choices and observable quota use. A different answer or latency is evidence to review, not proof that the replacement is better or worse for every task.

Only after the task passes do they update that runbook’s model instruction. They retain an explicit fallback plan and still inspect the selection when reopening old conversations after the removal. Automatic post-removal mapping remains unknown.

This example changed no account, called no model and consumed no quota.

The official Additional models section names Nano Banana 2 for the generative-image tool, including mockups, application images and diagrams. Google describes these additional stack models as not customizable. Keep that auxiliary role separate from the reasoning-model selector and plan matrix.

This guide does not infer more image-model choices, interchangeable quotas or a model-quality ranking. The next useful action is to check your actual plan and selector, then work through the migration checklist.