Claude Opus 5.5 and Sonnet 5.5 Are in Google Antigravity Now — and Free Users Are Losing Claude

Goo​gle has quietly added Claude Opus 5.5 and Claude Sonnet 5.5 to Goo​gle Antigravity, its agent-first coding app, and put a date on retiring the Claude models that came before them. The models themselves are not new: Ant​hropic shipped Claude Opus 5.5 on September 22, 2026, and Claude Sonnet 5.5 on September 28. What changed this week, dated to October 2 and 3, is where you can reach them. Both now appear in Antigravity's reasoning-model picker — but only on Goo​gle AI Pro subscriptions that are past their trial and on Goo​gle AI Ultra. Free, Goo​gle AI Plus and Enterprise accounts get neither. At the same time, Antigravity's model documentation marks Claude Opus 4.6, Claude Sonnet 4.6 and GPT-OSS-120b for removal on November 2, 2026 — which means the free tier's only remaining Claude options are the three rows already scheduled to disappear. Goo​gle has not posted a blog about any of it; the change is visible in the app's model menu and on Goo​gle's own models page, and the app's product lead acknowledged the stale catalogue on X two days before it was fixed.

What actually changed, and when

Antigravity's model documentation lists third-party reasoning models alongside Goo​gle's own, and as of October 4, 2026 the table reads like this:

• Claude Sonnet 5.5 (thinking) — Goo​gle AI Pro, non-trial subscriptions only; Goo​gle AI Ultra

• Claude Opus 5.5 (thinking) — Goo​gle AI Pro, non-trial subscriptions only; Goo​gle AI Ultra

• Claude Sonnet 4.6 (thinking) — Free, Goo​gle AI Plus, Goo​gle AI Pro; removal listed for November 2, 2026

• Claude Opus 4.6 (thinking) — Free, Goo​gle AI Plus, Goo​gle AI Pro; removal listed for November 2, 2026

• GPT-OSS-120b — Free, Goo​gle AI Plus, Goo​gle AI Pro, Goo​gle AI Ultra; removal listed for November 2, 2026

• Gemini 3.8 Flash, Gemini 3.7 Flash, Gemini 3.6 Flash, Gemini 3.1 Pro — all plans, unchanged

The two 5.5 rows are new. An archived copy of the same page from earlier in the week does not list them, Goo​gle's changelog and blog say nothing about the swap, and the release notes stop at Antigravity 2.19.1 on September 30 — whose contents (subagent messaging, PDF export, conversation-only undo) are unrelated. So the documentation moved first and the announcement has not followed, which is the ordinary pattern for third-party model availability in someone else's IDE.

The clearest public thread of the change is a user complaint. On September 30, the Antigravity-tracking account @thtbee_ posted that the model list still carried Sonnet 4.6, Opus 4.6 and GPT-OSS-120B and asked Goo​gle to replace them with Sonnet 5.5, Opus 5.5 and a broader G​PT catalogue. Varun Mohan, who works on Antigravity at Goo​gle DeepMind, replied the next day: "Yes, we will clean this up shortly!" Two days later, the models page carried the 5.5 rows and the removal dates, and users began posting screenshots of the in-app notice. That is a documented request, an acknowledgement from someone in a position to know, and a change in the documentation — which is a stronger chain than most roadmap chatter, but it is still not a launch post, and Goo​gle has published no pricing, quota or date for the swap beyond what the table says.

The plan gate is the actual story

A model appearing in a picker is normally unremarkable. What makes this worth a reader's attention is the shape of the gate, which is narrower than any previous Claude rollout in Antigravity.

Free and Goo​gle AI Plus accounts have no access to either 5.5 model. Neither does Enterprise — Antigravity in Gemini Enterprise lists the Goo​gle models and nothing else. Goo​gle AI Pro gets both, but the page carries a footnote that says "non-trial subscriptions only," which is why the user who started the thread, who describes themselves as a Pro subscriber, still saw an in-app warning about losing third-party models. If you are on a promotional, student or trial Pro plan, the 5.5 rows may simply not appear, and the removal notice may apply to you. Goo​gle AI Ultra, the top individual tier, has both without the footnote.

Support threads from the same morning describe Pro users who could not see Claude Opus 5.5 and received a message that third-party models were being removed; the replies were other users advising patience, with no Goo​gle staff answer. The plausible reading is that trial, promotional and student Pro plans are treated differently from paid Pro, and a slower staged rollout is also possible. What is not in dispute is the documentation: the footnote is Goo​gle's, not an inference.

One consequence is easy to miss. When the 4.6 models and GPT-OSS-120b go on November 2, the plans that lose them are exactly the plans that get nothing in return. Free, Goo​gle AI Plus and Enterprise become Gemini-only in Antigravity. Goo​gle AI Pro keeps a Claude tier, and Ultra keeps it too. Antigravity is becoming a Goo​gle-first product with Ant​hropic access as a paid perk — which is a deliberate commercial position, not an accident of the model list.

November 2 is the deadline you can plan around

Removal dates are rare in this category of documentation, and a dated removal is the single most actionable fact here. If you build inside Antigravity — skills, subagents, rules files, prompts that name a model — a hardcoded reference to Claude Sonnet 4.6, Claude Opus 4.6 or GPT-OSS-120b has a known end date of November 2, 2026. Nothing about the migration is technically hard; the risk is that a workflow written against 4.6 quietly keeps working until it does not.

Two details from the same page are worth knowing before you plan a switch. First, the model choice persists across messages within a conversation, but changing it mid-run does not take effect immediately: the agent finishes the current turn on the previously selected model. Second, quotas for Claude and G​PT models sit in a separate pool from Gemini models, with their own five-hour and weekly limits — the shared-quota change Goo​gle made in May applied to Gemini Flash and Pro against each other, not to third-party models, which stayed on a fixed limit because of capacity. A test that runs clean on a Tuesday afternoon tells you less about the billing than about that five-hour window.

What the three picker options cost when you call them directly

Antigravity does not bill per token. You spend plan quota, and if you exceed it on Pro or Ultra you can nominate AI credits as the overage mechanism — set to Never or Always in the app's settings. There is no bring-your-own-key path; Goo​gle's plans page states plainly that BYOK and bring-your-own-endpoint are not supported. So the Antigravity decision is a subscription decision.

If instead you want these models in your own code, the list prices are public and much lower than most people assume for a frontier tier:

• Claude Opus 5.5 — $4 per million input tokens, $20 per million output, $0.20 per million cache reads, 1M-token context, 128K max output, released September 22, 2026

• Claude Sonnet 5.5 — $2 per million input, $10 per million output, $0.20 cache reads, 1M-token context, 128K max output, released September 28, 2026

• Gemini 3.8 Flash — $0.75 per million input, $3.75 per million output, $0.075 cache reads, 1M-token context, 65,536 max output, released September 2, 2026, with introductory pricing running to the end of 2026

Those are Ant​hropic's and Goo​gle's own rate cards, and Antigravity's plan pricing is a different thing entirely — a $20 Goo​gle AI Pro subscription buys quota, not tokens, and comparing the two directly will mislead you in both directions. What the numbers do tell you is the shape of the workflow the tweet that surfaced this describes: Opus 5.5 for planning, Sonnet 5.5 for frontend, Gemini 3.8 Flash to execute. That is roughly a 5× cost step down from Opus to Sonnet and another 2.7× from Sonnet to Flash on input, which is why the tiering works as a budget strategy and not just as a preference.

We route Claude Opus 5.5 and Claude Sonnet 5.5 at Anthropic's list price with zero markup, and Gemini 3.8 Flash at Goo​gle's — one API key, one endpoint, and the same model IDs the vendors publish. Provider list prices pass through unchanged, so when a vendor moves a number the route moves with it rather than waiting on us. If you want to test whether Opus 5.5 planning is worth its multiple inside your own pipeline before you commit a subscription to it, that is a cheaper way to find out than a plan upgrade.

The Gemini 4 Argon question, kept separate

The thread that started this also argues the swap hints that Gemini 4 Argon is about to roll out widely. That is speculation, and it should be labelled as such. Goo​gle announced Gemini 4 Argon on September 30 and put it into the Fairwind Program — an application-only tier for vetted cyber-defence users, governments and critical-infrastructure operators — with paid API customers and Goo​gle AI Ultra subscribers named as later stages and no date on any of them. The Antigravity models page does not mention Argon at all. Retiring three third-party rows from a picker and opening access to a first-party flagship are separate decisions with separate gates, and nothing in the documentation connects them.

There is one genuine link worth noting, because it is the reverse of the rumour: Artificial Analysis's Coding Agent Index scores the whole agent rather than the model alone, and it carries Gemini 4 Argon running in Antigravity's CLI at 64, with Claude Sonnet 5.5 in Claude Code at 68 ahead of it and Claude Opus 5.5 at 66 — the entry marked as not publicly available. Antigravity's harness is already good enough to place an unreleased Goo​gle model third on that board. That says something about the app's agent layer, and nothing about when Argon becomes callable.

What to do before November 2

Four things are worth checking this week, in this order.

• Confirm which Pro you are on. If your subscription is a trial, promotional or student plan, the 5.5 models may not appear regardless of what the page says, and the removal notice may apply to you. Sign out and back in, update the app, and check the model selector before assuming anything is broken.

• Inventory hardcoded model references. Any skill, subagent, rules file or prompt naming Claude Sonnet 4.6, Claude Opus 4.6 or GPT-OSS-120b has a dated expiry. Move the ones that matter to Claude Sonnet 5.5 or Gemini 3.8 Flash and verify them while both still work.

• Separate the three Gemini Flash tiers in your head. Gemini 3.8 Flash, 3.7 Flash and 3.6 Flash all remain available on every plan, and only 3.8 is the current generation. A workflow pinned to 3.6 because it was stable in July is leaving measurable capability on the table.

• Decide whether you need Antigravity at all for the Claude half. If what you actually want is Opus 5.5 planning calls inside a pipeline you already own, the plan gate is a subscription question and the API is a per-token question, and they are not close in cost at low volume.

Questions the docs leave open

Do the 5.5 models count against the same quota as the 4.6 models did?

The models page shows Claude and G​PT models in one quota pool, separate from Gemini, but does not say whether Opus 5.5 and Sonnet 5.5 sit inside the same fixed limit the 4.6 models occupied or draw more heavily from it. Users in the support threads report tighter limits; Goo​gle has not confirmed that, and prompt cost in Antigravity is work-based rather than token-based, so the comparison is not straightforward. Watch your usage panel for the first fortnight rather than trusting either account.

Is Sonnet 5.5 the right replacement for Sonnet 4.6?

In capability terms, yes — it is a generation newer, and both share the same 1M-token context window. The migration risk is not the model but the effort settings. Sonnet 5.5's reasoning effort levels were recalibrated relative to Sonnet 5, so a configuration tuned against an older Sonnet will not behave the same way after a swap. Re-test prompts you care about instead of carrying settings over.

Will free users ever get Claude back in Antigravity?

Nothing in Goo​gle's own documentation suggests it. The 5.5 footnote ties Pro access to non-trial paid subscriptions, Enterprise carries no Claude rows at all, and the direction of the May 2026 plan changes — consolidating Gemini models into one shared quota while keeping third-party models on a fixed separate limit "because of capacity limitations" — points the other way. Treat Claude in Antigravity as a paid-tier feature going forward.

Bottom line

The models are old news; the gate and the deadline are the news. Claude Opus 5.5 and Claude Sonnet 5.5 arrived in Goo​gle Antigravity on October 2 and 3 with no announcement, restricted to paid, non-trial Goo​gle AI Pro and to Goo​gle AI Ultra — and on November 2 the Claude and G​PT options the free tiers still have are removed, leaving those plans with Gemini alone. If you are on a trial Pro and wondering why Opus 5.5 is missing, the documentation answers you: you are not the customer this tier is for. If you are building on the 4.6 models, you have twenty-nine days and a documented reason to move.