Which alternative you want depends on which tier you are on, and on one tier the answer is “none — just change the model name.” On 18 November 2026 Google removes Claude Sonnet 4.6, Claude Opus 4.6 and GPT-OSS-120b from Antigravity’s selector. On free or $4.99 AI Plus you are left with Gemini 3.8 Flash alone, and the real alternative is a BYO-key agent — Aider, OpenCode or Goose — at $0 plus tokens. On paid AI Pro at $19.99 you keep four models and two vendors: switch Sonnet 4.6 → Sonnet 5.5 and stay. On Ultra, GPT-OSS-120b leaves and nothing open-weight replaces it. You have until 18 November to test the switch.
What each tier actually loses
Verified against Antigravity’s model docs on 3 November 2026, when all three departing models were still selectable:
| Your tier | Models after 18 Nov | What you lost | Do you need to leave? |
|---|---|---|---|
| Free ($0) | Gemini 3.8 Flash | Claude 4.6 pair, GPT-OSS-120b | Yes, if choice mattered |
| Google AI Plus ($4.99) | Gemini 3.8 Flash | Same — a paid tier with the free tier’s list | Yes — Plus buys no Antigravity models |
| Google AI Pro ($19.99) | Gemini 4 Argon, 3.8 Flash, Claude Sonnet 5.5, Opus 5.5 | The Claude 4.6 pair only | No — rename the model |
| Google AI Ultra (from $99.99) | The above, plus Enterprise-grade quota | GPT-OSS-120b | No, unless open weights were the point |
| Enterprise | Unchanged | Nothing | No |
Two things fall out of that table that the circulating version of this story gets wrong. First, this is not a free-tier trim: paid Pro subscribers at $19.99 lose two models and Ultra subscribers at $99.99 lose one. Second, Enterprise is the only tier whose list does not shrink — every other tier moves toward Enterprise rather than the reverse.
The tier that comes off worst in relative terms is Google AI Plus. At $4.99 it is a paid subscription whose Antigravity model list becomes identical to the free one, and Google’s subscriptions page lists no Antigravity benefit against Plus at all. If Plus was bought for Antigravity, it is now buying storage and Gemini app quota.
The real question: who controls your model list?
Here is the part no vendor page will tell you. The reason three models vanished from your tool on a date you did not choose is that the selector belongs to Google, not to you — and Antigravity’s plans docs close the obvious escape hatch, stating that bring-your-own-key and bring-your-own-endpoint for additional rate limits are not supported. There is no route to paying for the old model with your own API spend.
That is a structural property, not a feature gap, and it sorts the alternatives better than any benchmark:
| Tool | Who picks the model list | Can a vendor remove a model from under you? |
|---|---|---|
| Antigravity | Google; no BYOK | Yes — this is what just happened |
| Cursor | Anysphere curates; Plus is Grok-only | Yes — and it is happening on 12 November, six days earlier |
| Claude Code | Anthropic (single vendor by design) | Only via Anthropic’s own 60-day retirement policy |
| OpenAI Codex | OpenAI (single vendor by design) | Only via OpenAI’s own 6-month policy |
| Windsurf | Cognition curates | Yes |
| Aider | You — any API key, any local model | No |
| OpenCode | You — 75+ providers | No |
| Goose | You — BYO model, open source | No |
If what you lost on 18 November was trust that the list would hold, the only durable fix is the bottom three rows. Everything else relocates the risk rather than removing it.
And the obvious escape route is failing in the same week. Cursor is the tool most people name when Antigravity’s model choice disappears — but OpenAI invoked a change-of-control clause after SpaceX acquired Anysphere and moved to terminate model supply, with a proposed shutoff of 12 November 2026. That is six days before the Antigravity deadline, and OpenAI’s position is that future models will not be supplied to Cursor at all. Worse for planning purposes, 12 November is the latest date the contract’s cancellation window permits rather than a chosen deadline, so it can only move earlier — the feature-by-feature survival table is in migrating off Cursor’s OpenAI models.
Two harnesses, two vendors, six days apart, neither decision made by the person paying for the tool. That is the argument for BYO-key stated better than any feature list could put it.
One candidate to rule out explicitly: Gemini CLI is not an alternative. Google shut it down for free, Pro and Ultra individual users on 18 June 2026 with no grace period, and pointed those users at Antigravity. Migrating from Antigravity to Gemini CLI would be migrating backwards into an already-executed sunset — by the same vendor, for the second time in five months.
What a BYO-key setup actually costs
The objection to BYO-key agents is always that metered tokens are scarier than a flat fee. Worth pricing, because the crossover is lower than most people assume.
Claude Sonnet 4.6 bills $3 per million input tokens, $15 per million output, $0.30 per million cached input. Take one agent turn carrying 50,000 tokens of context at a 90% cache hit rate and producing 3,000 tokens of output:
- 45,000 cached input tokens × $0.30/M = $0.0135
- 5,000 fresh input tokens × $3/M = $0.0150
- 3,000 output tokens × $15/M = $0.0450
- ≈ $0.0735 per turn — note that output is 61% of it
Over 22 working days:
| Agent turns per day | Monthly token cost | Versus a $19.99 seat |
|---|---|---|
| 5 | ~$8 | Much cheaper BYO-key |
| 10 | ~$16 | Cheaper |
| ~12 | ~$19.99 | Breakeven |
| 20 | ~$32 | Seat is cheaper |
| 40 | ~$65 | Seat is much cheaper |
| 80 | ~$129 | Seat is much cheaper |
The crossover is roughly 12 agent turns a day. Below it, a free BYO-key agent beats every subscription on this page and cannot have its model list edited by a vendor. Above it, you are buying a subscription for the rate limits, and the honest comparison is which subscription — not whether. Your own numbers will move with context size and cache hit rate; the shape is what transfers. More arithmetic of this kind is in the API cost calculator and what AI automation really costs.
Case 1: you are on free or AI Plus
You lose model choice completely, so the question is whether you were using Antigravity for the choice or for the harness.
- For the choice → go BYO-key. Aider and OpenCode are free and open source, take any provider key, and run Claude Sonnet 4.6 today and whatever replaces it later without anyone’s permission. Goose sits under the Linux Foundation’s Agentic AI Foundation, which is about as far from a vendor-curated selector as this category goes. At under a dozen turns a day this is also the cheaper option outright.
- For the harness → staying free is defensible. Gemini 3.8 Flash is fast and cheap, and the four surfaces (desktop app, CLI, SDK, Managed Agents API) are the genuine product. Just price the tier honestly: after 18 November this is a single-model Gemini tool, whatever the pricing page’s “access to a variety of agent models” says.
- Do not buy AI Plus at $4.99 for Antigravity. It buys no models the free tier lacks.
Case 2: you are on paid AI Pro at $19.99
Do not migrate. Rename the model and verify your subscription is non-trial.
Pro keeps Gemini 4 Argon, Gemini 3.8 Flash, Claude Sonnet 5.5 (thinking) and Claude Opus 5.5 (thinking) — four models across two vendors, which is more multi-model choice than a $20 Claude Pro or ChatGPT Plus seat. Sonnet 4.6 → Sonnet 5.5 and Opus 4.6 → Opus 5.5 are the intended swaps, and Sonnet 5.5 is both newer and cheaper per token than the model it replaces ($2/$10 against $3/$15).
The gotcha is the second footnote: Claude Sonnet 5.5 and Opus 5.5 are “Available on Google AI Pro for non-trial subscriptions only.” Check this before 18 November, not after. A team that piloted a Claude-in-Antigravity workflow on a Pro trial has tested a configuration it does not own, and the day it finds out is the day the fallback models disappear. If you need to convert, Google’s docs route you through a request form rather than a self-serve switch, which is not a thing to start on 18 November.
If you want to leave anyway, the direct comparison is Claude Code vs Antigravity; the wider field is in best AI coding tools and best AI harnesses.
Case 3: you used GPT-OSS-120b
This is the only loss with no in-product replacement at any price. GPT-OSS-120b was the one non-proprietary reasoning model in the selector and the only departing model offered as far up as Ultra. After 18 November every remaining option in Antigravity’s table is a hosted, closed model from Google or Anthropic.
If open weights were the requirement — auditability, no per-token billing, data that does not leave your machine — the replacement is running the weights yourself and pointing a BYO-key agent at them. Antigravity’s own SDK runs local models through LiteRT, so the capability has not left the product, only the convenience: the documented floor is more than 24GB of VRAM or unified memory, and the token cost becomes a hardware cost. Aider, OpenCode and Goose all talk to a local endpoint, so the agent layer is free either way.
What this costs, side by side
| Option | Price | Models you can pick | Vendor can edit the list? |
|---|---|---|---|
| Antigravity free | $0 | Gemini 3.8 Flash only | Yes |
| Antigravity AI Plus | $4.99/mo | Gemini 3.8 Flash only | Yes |
| Antigravity AI Pro | $19.99/mo | Argon, 3.8 Flash, Claude Sonnet/Opus 5.5 ‡ | Yes |
| Antigravity AI Ultra | from $99.99/mo | The above, highest quota | Yes |
| Claude Code Pro | $20/mo ($17 annual) | Anthropic models | Retirement policy only |
| Cursor Plus | $8/mo | Grok only on this tier | Yes |
| Cursor Premium | $30/mo | Third-party models at API list price, billed separately by usage; OpenAI supply ends 12 Nov | Yes |
| Windsurf Pro | $20/mo | Cognition’s curated list | Yes |
| OpenAI Codex | $20/mo with ChatGPT Plus | OpenAI models | Retirement policy only |
| Aider / OpenCode / Goose | $0 + tokens | Anything you hold a key for | No |
‡ “Available on Google AI Pro for non-trial subscriptions only.” Prices verified 3 November 2026 against Google’s subscriptions page, cursor.com/pricing, claude.com/pricing and Cognition’s billing docs; current figures for every tracked vendor are in the pricing tracker.
Note the trap in the middle of that table: Cursor Plus at $8 is the cheapest row that looks like multi-model and is not — its plan card states that only Grok models are available on the plan. Premium at $30 does restore third-party models, but Cursor’s own fine print is that third-party usage “is not included in the monthly prices shown and must be purchased separately” at API list price, so $30 is a floor rather than the bill. The cheapest subscription that genuinely gives you two model vendors with usage included is Antigravity AI Pro at $19.99 — the tier you are already on if you are in Case 2. Leaving it for model choice usually means paying more for less.
Before 18 November: a short checklist
- Grep your configs and runbooks for
claude-sonnet-4-6,claude-opus-4-6andgpt-oss-120b. Anything that names a model by string breaks silently when the selector drops it. - Confirm your Pro subscription is non-trial if you intend to rely on Claude 5.5. This is the step with a lead time.
- Test the replacement now, while both are selectable. After 18 November the comparison is unavailable.
- Re-read the docs table on the day. It has moved three times since September, the removal date already slipped sixteen days without notice, and the page carries no last-updated stamp. Plan the migration for 18 November; do not plan on 18 November holding.
- If the list moving is itself the problem, the fix is BYO-key, not a different vendor’s selector.
Vendors with published retirement policies are tracked on our model retirements calendar — Anthropic commits to 60 days, OpenAI to six months. Third-party models inside somebody else’s harness appear on none of them, which is exactly why this one arrived as a footnote.
What changed
- 3 Nov 2026: Google’s docs footnote moved the removal date from 2 November to 18 November 2026 with no announcement; all three models verified still selectable — see analysis.
- 2 Nov 2026: Removal of Claude Sonnet 4.6, Claude Opus 4.6 and GPT-OSS-120b from Antigravity first surfaced as a single asterisked footnote, with no changelog entry or blog post.
Frequently asked questions
What is the best Antigravity alternative?
It depends which tier you are on, because 18 November takes away something different at each one. On free or $4.99 Google AI Plus you lose model choice entirely — the selector drops to Gemini 3.8 Flash alone — so the alternative is any tool that still offers more than one model, and the cheapest honest answer is a BYO-key terminal agent like Aider, OpenCode or Goose at $0 plus whatever tokens you burn. On paid Google AI Pro at $19.99 you do not need to leave at all: Claude Sonnet 5.5 and Opus 5.5 stay in the selector, and the fix is editing the model name in your runbook — provided your subscription is non-trial, which Google's docs require for those two models. On Ultra you lose GPT-OSS-120b and there is no open-weight replacement anywhere in the selector, so if that was the point, the replacement is running the weights yourself. The one framing to avoid is treating this as a reason to buy a different $20 subscription, because most of them give you fewer models than Antigravity Pro, not more.
What models are left on the Antigravity free tier after 18 November 2026?
One: Gemini 3.8 Flash. Antigravity's docs put free and Google AI Plus in a single combined column, 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 is footnoted 'Available with paid Google AI Pro plans or on trial with eligible premium devices.' Claude Sonnet 5.5 and Opus 5.5 are Pro-and-above. Gemini 3.7 Flash, 3.6 Flash and 3.1 Pro are Enterprise-only. So a free or $4.99 Plus user who opens the model selector on 18 November has no reasoning-model decision left to make. Google AI Plus is the tier that comes off worst in relative terms: it is a paid subscription whose Antigravity model list becomes identical to the free one, and Google's own subscriptions page does not list an Antigravity benefit against Plus at all.
Do I have to migrate off Antigravity if I pay for Google AI Pro?
No, and for most Pro subscribers migrating would be the wrong move. The $19.99 tier keeps Gemini 4 Argon, Gemini 3.8 Flash, Claude Sonnet 5.5 (thinking) and Claude Opus 5.5 (thinking) — that is still four models and two vendors, which is more multi-model choice than a $20 Claude Pro or ChatGPT Plus seat gives you. What you lose is specifically the Claude 4.6 pair, and the migration is a one-line change from Sonnet 4.6 to Sonnet 5.5 in whatever names the model. The catch worth checking before 18 November is the second footnote: the Claude 5.5 models are 'Available on Google AI Pro for non-trial subscriptions only.' If you piloted your workflow on a Pro trial, the configuration you tested is not the configuration you will deploy, and you will discover that on the day the older models disappear rather than before.
Is Claude Sonnet 4.6 being retired, or just removed from Antigravity?
Just removed from Antigravity, and the distinction decides whether you have a migration at all. Anthropic continues to serve Claude Sonnet 4.6 and Claude Opus 4.6; our retirement calendar has Sonnet 4.6 carrying a 'not sooner than' date of 17 February 2027 and Opus 4.6 of 5 February 2027, neither of which is a firm shutdown. What ends on 18 November is one route to those models — Google's selector — not the models themselves. If your work genuinely depends on Sonnet 4.6 specifically, you can keep using it today through Claude Code, the Claude API, or any BYO-key agent, at $3 per million input tokens and $15 per million output. Treating this as a model retirement leads teams to rewrite prompts for a new model when all they needed was a different client.
What does it cost to run Claude Sonnet 4.6 yourself instead of through a subscription?
About $32 a month at a moderate agentic workload, and it breaks even against a $19.99 seat at roughly 12 agent turns a day. The arithmetic, with assumptions stated: Sonnet 4.6 bills $3 per million input tokens, $15 per million output and $0.30 per million cached input. A single agent turn carrying 50,000 tokens of context at a 90% cache hit rate and producing 3,000 tokens of output costs about $0.0735 — roughly $0.014 of cached context, $0.015 of fresh input and $0.045 of output, which is where most of the money goes. Twenty such turns a day over 22 working days is about $32. Five turns a day is about $8, comfortably under any subscription. Eighty turns a day is about $129, comfortably over. Your real numbers will differ with context size and cache behaviour, but the shape holds: light use is cheaper BYO-key, heavy use is cheaper on a seat, and the crossover sits around a dozen turns a day.
Why did Google remove these models without announcing it?
Google has not said, and the absence of a reason is the part worth planning around. The entire notice is a one-line footnote under a docs table: 'Will be removed on November 18, 2026.' There is no changelog entry, no developer-blog post, no deprecation-page row, no stated replacement and no grace period. The page carries no last-updated date, which is why nobody noticed that the same footnote read 'November 2, 2026' on 2 November and 'November 18, 2026' the next day — a sixteen-day slip Google also did not announce. Third-party models inside a vendor's harness sit on nobody's retirement calendar: Anthropic commits to 60 days' notice for its own models and OpenAI to six months for generally available ones, but neither policy covers somebody else's selector. The practical lesson is that an undated availability table is not a source you can subscribe to, only one you have to re-read.