Antigravity's free tier loses every third-party model on 18 November — a date Google already moved once

Google is removing three third-party models from Antigravity, and the entire notice is an asterisk that has already changed date once. Under the reasoning-model availability table in Antigravity’s own docs, a footnote reads: ”* Will be removed on November 18, 2026.” The asterisk is attached to three rows — Claude Sonnet 4.6 (thinking), Claude Opus 4.6 (thinking) and GPT-OSS-120b. There is no blog post, no changelog entry, no reason, no recommended replacement and no grace period.

On 2 November, when we first verified that table, the same footnote read ”* Will be removed on November 2, 2026” — and the models were still selectable. On 3 November it reads November 18, and they are still selectable. Google moved a removal date sixteen days and said nothing, which tells you more about how to plan around this product than the removal itself does.

The story circulating is that this hits the free and Google AI Plus tiers. The table says something broader and more interesting.

The removal is not tier-scoped, and that changes who it hurts

Because the asterisk sits on the model rows rather than on a plan column, it removes the models everywhere they are offered. Here is the availability table as it stands on 3 November 2026, with the departing rows marked:

Model	Free & AI Plus	AI Pro ($19.99)	AI Ultra (from $99.99)	Enterprise
Gemini 4 Argon	—	yes †	yes	yes
Gemini 3.8 Flash	yes	yes	yes	yes
Gemini 3.7 / 3.6 Flash, 3.1 Pro	—	—	—	yes
Claude Sonnet 5.5 (thinking)	—	yes **	yes	yes
Claude Opus 5.5 (thinking)	—	yes **	yes	yes
Claude Sonnet 4.6 (thinking) *	until 18 Nov	until 18 Nov	—	—
Claude Opus 4.6 (thinking) *	until 18 Nov	until 18 Nov	—	—
GPT-OSS-120b *	until 18 Nov	until 18 Nov	until 18 Nov	—

† “Available with paid Google AI Pro plans or on trial with eligible premium devices.” ** “Available on Google AI Pro for non-trial subscriptions only.”

Three consequences fall out of that, in rising order of how much they should bother a buyer.

Paid Pro subscribers lose models, not just free users. Google AI Pro at $19.99/month carries both Claude 4.6 models. Ultra at $99.99 carries GPT-OSS-120b. This is not a free-tier trim; it is a reduction in what two paid subscriptions include, on the same day, with the same footnote.

The free tier goes down to one model. Free and Plus share a single column, and in that column only Gemini 3.8 Flash and the three departures are marked available. Gemini 4 Argon needs a paid Pro plan or an eligible premium device. The Claude 5.5 pair is Pro-and-above. Gemini 3.7 Flash, 3.6 Flash and 3.1 Pro are Enterprise-only. So from 18 November a free or $4.99 Plus user opens the model selector and finds no choice to make. Antigravity’s own pricing page still advertises that tier as “Access to a variety of agent models.”

Enterprise is the only tier that does not shrink. For months the notable quirk of Antigravity’s model table was that Enterprise — the tier most organisations actually buy — was the one place Claude and GPT-OSS were unavailable. On 18 November every other tier moves toward Enterprise rather than the reverse.

What survives, and the condition attached to it

Third-party models are not leaving Antigravity entirely. Claude Sonnet 5.5 (thinking) and Claude Opus 5.5 (thinking) stay on Pro, Ultra and Enterprise. But on Pro they carry their own footnote: “Available on Google AI Pro for non-trial subscriptions only.”

Read the two footnotes together and the product’s position has shifted. Until 18 November you can evaluate Antigravity’s multi-model story on a free seat. After it, third-party model access is a paid, non-trial feature. That is a reasonable commercial decision and a poor one to communicate by asterisk, because the people most affected are exactly the ones who built a pilot on a trial or free seat and will discover the change when the selector no longer offers what their runbook names.

GPT-OSS-120b leaving matters more than its usage probably did

Of the three departures, the open-weight one is the most quietly consequential. GPT-OSS-120b is the only non-proprietary reasoning model in Antigravity’s selector, and the only one available as far up the ladder as Ultra. Its removal leaves every remaining option in the table a hosted, closed model from either Google or Anthropic.

That cuts against the direction Google was advertising only weeks ago. In September it shipped an Antigravity SDK that runs local models through LiteRT, with the explicit pitch that the agent runner is not Google-locked at the model layer. Both things can be true — open weights in the SDK, closed models in the selector — but a buyer who read “not model-locked” as a property of the product should notice that the hosted surface just became more locked, not less. The open-weight path on Antigravity now requires running the model yourself, against a documented floor of more than 24GB of VRAM or unified memory.

The notice quality is the real story

Nothing about dropping a year-old model is unusual. What is unusual is where Google chose to say so:

Not in the changelog. Antigravity’s changelog entries for 27, 28 and 29 October 2026 are all feature releases — Marketplace partner plugins, a conversation navigator, a built-in browser. None mentions the removal.
Not on the developer blog. The three newest posts on Google’s developer blog as of today cover DevFest, A2UI with CopilotKit, and agent-quality evaluation.
Not on a deprecation page. Google’s Gemini API deprecation schedule — which did at least move to a new home at aistudio.google.com — tracks first-party Gemini models. Third-party models inside Antigravity are not on anybody’s retirement calendar, which means there is no page a buyer could have subscribed to in order to see this coming.
Not dated. The docs page carries no last-updated timestamp, so the only way to know the footnote changed is to have read the same table yesterday — which is exactly how the slip from 2 to 18 November surfaced. An undated page is not a source you can monitor; it is one you have to re-read.

Compare that with the vendors tracked on our model retirements calendar: Anthropic commits to at least 60 days’ notice, OpenAI to six months for generally available models, Mistral to six. Google’s Gemini API page states only that listed shutdown dates are “the earliest possible dates” and promises unquantified “advance notice”. For third-party models inside a Google harness, there is no stated policy at all — and a date that moves overnight with no notice demonstrates what that means in practice.

What to do about it

The tier-by-tier version of this, with the cost arithmetic and a pre-18-November checklist, is in Antigravity alternatives before 18 November. The summary:

If you are on free or Plus and Antigravity’s model choice was the reason: you have until 18 November, and then that reason is gone. Gemini 3.8 Flash is a fast, capable default, but it becomes the only default. The harness comparison worth re-reading is Claude Code vs Antigravity, and the broader field is in the best AI coding tools and best AI harnesses.

If you are on Pro at $19.99 and your workflow names Claude Sonnet 4.6 or Opus 4.6: switch to Claude Sonnet 5.5 or Opus 5.5 before 18 November, and confirm your subscription is non-trial before you rely on them. Note what you cannot do instead — Antigravity’s plans docs rule out bring-your-own-key for extra rate limits, so there is no route to paying for the old model with your own API spend. The Antigravity review has the tier-by-tier model table as it now stands.

If the model rather than the harness is what you need, the model is not going anywhere. Claude Sonnet 4.6 is still served by Anthropic — through Claude Code, the API, or another harness. Do not confuse this selector change with Anthropic’s own calendar, which has Claude Opus 4.5 retiring 24 November and Sonnet 4.5 on 30 November; plan choices for that ladder are in which Claude plan to buy.

And if you are evaluating any harness on a free seat, the general lesson is the one this change illustrates cheaply: verify model availability against the tier you will deploy on, not the tier you are testing on. Antigravity’s table has moved repeatedly since September — Gemini 3.7 Flash, 3.6 Flash and 3.1 Pro are now Enterprise-only, the Claude 5.5 pair and Gemini 4 Argon arrived on the paid tiers, three models are leaving, and the date they leave on has itself already changed — and none of it is dated on the page.

Update, 3 November 2026 — the date moved to 18 November, and the correction is the story

This article was first published on 2 November stating that the three models were removed that day, because Antigravity’s docs footnote read “Will be removed on November 2, 2026” when we verified it. Re-checking the same page on 3 November, the footnote reads “Will be removed on November 18, 2026” and all three rows — Claude Sonnet 4.6 (thinking), Claude Opus 4.6 (thinking), GPT-OSS-120b — are still present and still marked available on the tiers they were offered on. Nothing was removed on 2 November. The headline, the table and the dates above have been corrected.

Google published no notice of the slip, which is the same failure as the original announcement rather than a separate one. There is still no changelog entry, no developer-blog post and no deprecation-page row, and the docs page still carries no last-updated date — so the only detectable signal that a date moved was reading the same table on two consecutive days.

Two practical consequences. First, buyers now have fifteen days rather than none: if a Pro runbook names Claude Sonnet 4.6 or Opus 4.6, the switch to Sonnet 5.5 or Opus 5.5 can be tested rather than discovered. Second, 18 November should be read as the earliest credible date, not a commitment. A date that moved once without notice can move again in either direction, and a vendor that does not date its own availability table has not given you anything to monitor. Plan the migration for the date; do not plan on the date holding.

Frequently asked questions

When exactly do these models leave Antigravity?

18 November 2026, as the footnote reads on 3 November. It originally read 'Will be removed on November 2, 2026' — we verified that wording on 2 November — and the following day the same footnote said November 18 instead, with Claude Sonnet 4.6 (thinking), Claude Opus 4.6 (thinking) and GPT-OSS-120b all still present and still marked available. Google announced neither the original date, nor the removal, nor the sixteen-day slip: there is no changelog entry, no developer-blog post and no deprecation-page row. Because the asterisk sits on the model rows rather than on any tier column, the removal is not scoped to a plan — it takes the models out everywhere they currently appear. Treat 18 November as the earliest credible date rather than a commitment.

Does this only affect the free and Google AI Plus tiers?

No, and that is the most widely mis-stated part of the story. Per Antigravity's own table, the Claude 4.6 pair was available on 'Free & Google AI Plus' and on 'Google AI Pro' — so paid Google AI Pro subscribers at $19.99 a month lose those two models as well. GPT-OSS-120b went further up the ladder still, listed on Free and Plus, Pro and Ultra, so Ultra subscribers at $99.99 lose a model too. All three were already marked unavailable on Enterprise, which is the one tier the change does not touch. The asymmetry worth noticing is that Enterprise, long the tier where Antigravity's third-party models were missing, is the only tier whose model list does not shrink on 18 November.

What model choice is left on Antigravity's free tier?

One model: Gemini 3.8 Flash. The free tier and Google AI Plus share a single column in Antigravity's docs, and in that column the only rows marked available are Gemini 3.8 Flash plus the three models leaving on 18 November. Gemini 4 Argon requires 'paid Google AI Pro plans or … trial with eligible premium devices'; Claude Sonnet 5.5 and Opus 5.5 are Pro-and-above; and Gemini 3.7 Flash, 3.6 Flash and 3.1 Pro are listed Enterprise-only. So from 18 November a free or $4.99 Plus user who opens the model selector has no reasoning-model decision left to make, which is a material change to the product's pitch rather than a housekeeping update.

Can paid Antigravity users still run Claude models?

Yes, but with a condition that is easy to miss. Claude Sonnet 5.5 (thinking) and Claude Opus 5.5 (thinking) remain on Google AI Pro, Ultra and Enterprise — and on Pro they carry their own footnote: 'Available on Google AI Pro for non-trial subscriptions only.' So the newer Claude pair is gated behind a converted, paying subscription, not a trial. Combined with the 18 November removal, third-party model access on Antigravity becomes a paid, non-trial feature. Teams that piloted a Claude-in-Antigravity workflow on a trial seat should re-check what the selector offers before committing, because the pilot configuration and the deployed configuration are no longer the same.

Should this change which coding harness you buy?

It should change how much weight you put on model portability as a reason to pick Antigravity. The strongest argument for the product was that the harness and the model are separate purchases — the argument Google itself illustrated in September 2026 by routing its own engineers' Claude Opus 5 access through Antigravity rather than through Claude Code. That argument still holds on Pro and above, where the Claude 5.5 pair is selectable. From 18 November it stops holding on free or Plus, which are left with exactly one model. If multi-model choice is load-bearing for you, treat it as a Pro-tier-and-up feature and verify it against the docs table for the specific tier you will deploy on, because that table has moved more than once since September and carries no last-updated date to warn you.

Is Claude Sonnet 4.6 being retired by Anthropic too?

No — these are two unrelated things, and conflating them leads to the wrong migration. This change is Google removing models from its own harness's selector; the models themselves are unaffected and continue to be served by Anthropic. Anthropic's published retirement calendar is separate, and what it currently shows for the nearby window is Claude Opus 4.5 reaching shutdown on 24 November 2026 and Claude Sonnet 4.5 on 30 November 2026. If you were using Claude Sonnet 4.6 inside Antigravity, the model is still available — through Claude Code, the Claude API, or any other harness that offers it. What you lost is one route to it.

Sources

Related tool reviews

Questions or corrections? Email Pick Right. Want the full list? See all news.