Gemini 4 Carbon Is Public, a Day After Claude Fable 5.2 Took Its Number One Spot. The $2 and $10 Price Lasts Through November 30, and in Google's Apps It Is Ultra Only.
Google made Gemini 4 Carbon generally available on November 10: the Gemini API paid tier in AI Studio, Gemini for business, and Google AI Ultra subscribers in Antigravity, the Gemini app, and AI Mode in Search. The introductory price of $2 per million input tokens and $10 output runs through November 30 and doubles to $4 and $20 on December 1. Carbon held first place on Artificial Analysis's Intelligence Index from October 28 until November 9, when Claude Fable 5.2 arrived at 65; Carbon at high effort now reads 60, fourth of 74, at $7.76 per Index task against Fable 5.2's $26.54, with 66 seconds to first answer token and 205 million output tokens. The default thinking level is medium; AA scored high. Google's three outside-benchmark wins are vendor claims, and one showcase board was co-built with Google DeepMind. We read Google's post, the AI Studio notes, and AA's pages.
By T.W. Ghost
Google opened Gemini 4 Carbon to the public today, November 10, 2026, thirteen days after announcing it for Fairwind cyber defenders. It is on the Gemini API paid tier in AI Studio, in Gemini for business, and for Google AI Ultra subscribers in Antigravity, the Gemini app, and AI Mode in Search. Fairwind keeps a version with fewer guardrails for cyber work; Antigravity gets a multi-agent preview. The model id is gemini-4-carbon; the default thinking level is medium; it is not on the free tier.
The timing is the catch. Carbon held the top of Artificial Analysis's Intelligence Index from October 28 to November 9. Yesterday, Claude Fable 5.2 arrived at 65 and took it. On the live board Carbon at high effort reads 60, fourth of 74, behind Fable 5.2 at max, xhigh, and high, with Opus 5.5 at max fifth at 59. The price is the other half of the story. The introductory rate of $2 per million input and $10 output, which had no published end date at announcement, now has one: November 30. From December 1 it is $4 and $20, with cached input at $0.10 then $0.20 and thinking billed as output. We read Google's availability post, the AI Studio notes, AA's model page and leaderboard, and the Verge and VentureBeat coverage. Here is what held up.
What Changed Today
| Gemini 4 Carbon, generally available (November 10, 2026) | |
|---|---|
| API | Gemini API paid tier in AI Studio; Gemini for business; not on the free tier |
| Consumer | Google AI Ultra subscribers in Antigravity, the Gemini app, AI Mode in Search; no word on lower tiers |
| Fairwind | Keeps "Carbon for cyber" with fewer guardrails |
| Antigravity | A /swarm multi-agent preview |
| Intro price | $2 in / $10 out per 1M through November 30, 2026; cached input $0.10 |
| From December 1 | $4 / $20; cached input $0.20; thinking billed as output |
| Other tiers | Batch and Flex $1 / $5; Priority $3.60 / $18; regional endpoints plus 10% (secondary; not checked on Google's price page) |
| Default thinking level | Medium; AA scores the high setting |
| AA Intelligence Index v5.0.1 | 60 at high, fourth of 74; Fable 5.2 65 at max, 62 at xhigh and high; Opus 5.5 59 at max |
| AA cost per Index task | $7.76 at the $4 / $20 list; Fable 5.2 $26.54 |
| AA run | 205M output tokens (median 72.8M); 26.1 minutes per task; 78 tokens per second; 66.01 seconds to first answer token, on AA's v3 method |
| Vercel AI Gateway | Listed as gemini-4-carbon |
| Ultra pricing (US) | $99.99 Ultra 5x; $199.99 Ultra 20x (press; not rechecked) |
It lost first place before it shipped. Thirteen days as AA's top model, all of them behind a cyber-defender gate, and then a public launch the day after a competitor passed it. Fourth of 74 is still a frontier score; it is not the headline Google wrote on October 28.
Twenty days of intro price. Any cost comparison made this month should print both numbers, because every December invoice is on the second one. At the list price AA's figure is $7.76 per Index task, under a third of Fable 5.2's, and that is the comparison that survives November 30.
Medium is what you get; high is what was scored. The 60 is a high-effort number. An API call that does not set the thinking level runs at medium, and nobody has published what medium scores.
Slow to start, long to finish. Sixty-six seconds to the first answer token and 26 minutes per Index task, on AA's new method that counts thinking time, with 205 million output tokens against a median of 73 million. The per-token price is low; the tokens are many.
The Benchmarks, Sorted by Who Reported Them
| Claim | Number | Who |
|---|---|---|
| Intelligence Index v5.0.1 | 60 at high, fourth of 74 | Independent (Artificial Analysis) |
| Cost, tokens, time per task | $7.76; 205M; 26.1 min; 66 s to first answer | Independent (AA) |
| CursorBench 4.0, Perplexity WANDR, Zapier AutomationBench | First on all three at more than 2x lower cost per task than the next external model | Vendor (Google) |
| CursorBench 4.0 score | 60.4% | X summary only; unverified |
| Multimodal Knowledge Agent | Second, at more than 6x lower cost than first | Vendor; the benchmark is Google DeepMind's own, launched on AA the same day |
| Safety: 99% of red-team trajectories blocked; 0.8% per-attempt breach vs Argon 3.1%; Gray Swan cyberattack success 2.17% vs 9.86%; honesty 92% vs 70% | Vendor (Google's safety report) |
Three honest observations:
- AA's 60 and $7.76 are the two numbers to carry. Fourth on the board at under a third of the leader's cost per task is a real position, and it is the only independent one.
- The showcase board has a Google logo on it. Multimodal Knowledge Agent is a DeepMind benchmark hosted by AA, announced the same day Google cited a second-place finish on it. Disclose it when you quote it.
- The three first-place claims are vendor claims about outside benchmarks. They may hold; the percentages behind them are not on Google's post, and the one circulating for CursorBench came from a social summary.
Verified vs Unconfirmed: The Scorecard
| Claim | Verdict |
|---|---|
| GA Nov 10; surfaces; Ultra-only consumer access; Fairwind variant; Antigravity preview | Verified (Google blog) |
| $2 / $10 through Nov 30; $4 / $20 from Dec 1; cached input; thinking billed as output; not on free tier | Verified (AI Studio notes; Google footnote) |
| Default thinking level medium | Verified (AI Studio notes) |
| AA 60, fourth of 74; Fable 5.2 65; Opus 59 | Verified (AA leaderboard and model pages) |
| $7.76 per task; 205M tokens; 26.1 min; 78 tok/s; 66 s to first answer | Verified (AA model page, v3 method) |
| Batch, Flex, Priority, regional pricing | Secondary (LiteLLM) |
| Three outside-benchmark wins; MKA second | Vendor |
| CursorBench 60.4% | Unverified |
| Safety figures | Vendor |
| Ultra tier prices | Press |
Who Should Use It
Use Carbon through the API this month if fourth place at $7.76 a task beats first place at $26.54 for your work, and set the thinking level yourself, because the default is not the scored setting.
Skip the intro price in any plan that runs past November 30, and skip the number-one framing; that was true for thirteen days and is not true today.
The frontier is elsewhere, by five points and three times the price: Fable 5.2 is first, and the cheaper Gemini 4 you could already call, Argon, is still on the board. Our model comparison has the current standings side by side.
Sources
- Google: Gemini 4 Carbon now available
- Google AI Studio: what's new in Gemini 4 Carbon
- Artificial Analysis: Gemini 4 Carbon model page
- Artificial Analysis: models leaderboard
- The Verge: Google's Gemini 4 Carbon public rollout
- VentureBeat: Google Gemini 4 Carbon is now publicly available and undercuts Claude Fable 5.2 and Opus 5.5 on API price
- Artificial Analysis on X: Multimodal Knowledge Agent benchmark with Google DeepMind
*Not sure whether fourth place at a third of the price is the better deal for your work? Take the free 2-minute quiz and get matched. Then read about the model that took the top row the day before: Claude Fable 5.2.*