On November 10, 2026, Gemini 4 Carbon moved from a closed program for cybersecurity defenders to the Gemini API for every paying developer. It launches at $2 per million input tokens and $10 per million output tokens, the same as Gemini 4 Argon, but only until November 30: from December 1 it costs $4 and $20. On the Artificial Analysis index it lands two points behind Claude Fable 5.2.
The useful question isn't whether it's "the best Gemini yet", but whether it's worth moving traffic to a model whose price doubles in three weeks.
What Carbon brings
Google positions it above Argon, which remains its recommendation for most software engineering work. Carbon is built for harder reasoning: it explores and verifies more before answering, so it burns more tokens than Argon at the same per-token price. It supports three thinking levels (low, medium and high), with medium as the default.
Where you can get it:
- Gemini API in Google AI Studio (paid tier) and Android Studio, as
gemini-4-carbon. - Enterprises: Gemini Enterprise, Google Antigravity and CodeMender.
- Google AI Ultra subscribers: the Gemini app, Antigravity and AI Mode in Search.
The announcement doesn't give a context window, output limit, rate limits or cached-token pricing, and it doesn't mention Vertex AI. To size a deployment you'll need the model docs, not the launch post.
On safety, Google reports a 0.8% per-attempt breach rate in adaptive red-teaming (3.1% for Argon) and 92% on its internal 375-task agentic honesty test (70% for Argon). These are Google's own tests; as for the 0% success rate on indirect prompt injection, The Decoder points out that benchmark may be saturated.
How to try it
The quickest route from a terminal is Simon Willison's llm-gemini plugin, which added Carbon in version 0.36:
llm install -U llm-gemini
llm keys set gemini
git diff main | llm -m gemini-4-carbon -o thinking_level high \
-s 'Review this diff: concurrency bugs and resource leaks'thinking_level controls how much it reasons before answering. To compare cost against Argon, run the same task with -m gemini-4-argon and look at output tokens: per-token pricing is identical until November 30, so the difference on your bill is exactly the difference in tokens.
Carbon vs Claude and GPT-6
Artificial Analysis figures as reported by The Decoder (measured about two weeks before November 10), plus pricing from Google and Simon Willison's comparison:
| Model | Price per million (input / output) | AA Intelligence Index | Cost per solved task (AA) |
|---|---|---|---|
| Gemini 4 Carbon | $2 / $10 until Nov 30; $4 / $20 after | 60 | $7.76 |
| Claude Fable 5.2 (High) | — | 62 | ~$9 |
| Claude Fable 5.2 (Max) | — | — | up to $26.50 |
| Claude Opus 5.5 (Max) | $4 / $20 | — | $10.04 |
| GPT-6 Astra (Max) | — | — | $6.25 |
| Claude Sonnet 5.5 and GPT-6.1 Sol | $2 / $10 | — | — |
On the agentic benchmarks Google published, Carbon ranks first on CursorBench 4.0 (just under 60%, ahead of Fable 5.2 and Argon), on Perplexity's WANDR and on Zapier's AutomationBench, the last two at less than half Fable 5.2's cost per task. On Multimodal Knowledge Agent it comes second, behind Fable 5.2 and ahead of Opus 5.5, at under a sixth of the leader's average cost per task.
Where it wins and where it doesn't
It wins on cost per agentic task. The numbers that matter most are CursorBench and AutomationBench: long coding and automation tasks, exactly where an expensive model shows up on the bill. If you run Fable 5.2 at max effort for agents today, Carbon promises similar results for considerably less.
It isn't the cheapest per task. GPT-6 Astra at max effort comes in cheaper per solved task ($6.25 vs $7.76). And the $7.76 was measured before GA: if it was computed at launch pricing, December's rates would push cost per task towards $15, above Opus 5.5.
It isn't the smartest. Fable 5.2 still leads the overall index (62 vs 60) and MKA. The gap is small, but it's there.
It spends more tokens than Argon. At the same per-token price, every Carbon task costs more than the same task on Argon. For summarization, classification or extraction, Argon (or a smaller model) is still the sensible choice.
The numbers are Google's. Nearly all the agentic benchmarks come from Google or its partners (Cursor, Perplexity, Zapier). The Artificial Analysis index is the only independent figure in the table, and there Carbon is second.
Verdict
Worth it for: teams running coding agents or long automations that currently pay for Fable 5.2 or Opus 5.5 at max effort. Until November 30 you can benchmark Carbon on your own tasks at Argon's price, which is the best window to decide based on your data rather than someone else's table.
Not worth it for: anyone happy with Argon or mid-tier models for short tasks, and anyone who needs Vertex AI with regional guarantees or specific context limits, which the announcement doesn't cover. And if your decision hinges on November's price, run the numbers with December's: $4 and $20 per million is Opus 5.5 pricing.
Sources
- Gemini 4 Carbon is now available, Google, November 10, 2026.
- llm-gemini 0.36, Simon Willison, November 10, 2026.
- Gemini 4 Carbon arrives just in time to feel like yesterday's news, The Decoder, November 10, 2026.
- llm-gemini (README), Simon Willison.