# Gemini 4 Carbon: Google’s new model arrives 15 days after Argon

> What changes with Gemini 4 Carbon versus Argon, Astra and Opus 5.5: Google’s published benchmarks, API pricing and initial access conditions.


Published: 2026-10-28. Original source: [Google · Introducing Gemini 4 Carbon](https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-carbon/).

![Official Google image: Gemini 4 Carbon with a blue crystal and the model name](/news/gemini-carbon/cover.webp)

Official cover from [Google](https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-carbon/). Benchmarks from Google’s published table and methodology.

Google announced Gemini 4 Carbon on October 28, fifteen days after introducing Argon. The company describes it as its strongest model for cybersecurity and long-horizon tasks. The announcement is official; access is beginning with a limited group of cyber defenders.

If you had just settled on Argon, this is a quick follow-up: [Google introduced it on October 13](/en/guides/gemini-4-argon-available-benchmarks/). Sundar Pichai presented Carbon on X; Koray Kavukcuoglu wrote the official blog post. [Google’s announcement](https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-carbon/) · [Sundar Pichai’s post](https://x.com/sundarpichai/status/2115489219804352921).

## What the benchmarks show

Google says Carbon scored 59 on Artificial Analysis Intelligence Index v5.0, the highest figure in the comparison table it published: Argon scored 50, GPT-6 Astra 57, Claude Opus 5.5 58, and Claude Fable 5.1 49. These figures come from Google’s published comparison. Its [methodology](https://deepmind.google/models/evals-methodology/gemini-4-carbon) identifies the index version with tools and explains how each result was obtained.

There is no universal winner in the table. On FrontierSWE 2, GPT-6 Astra scores 65.5%, ahead of Opus 5.5 (62.3%) and Carbon (61.2%). On OSWorld-2.1’s offline subset, Opus 5.5 leads at 87.2%; Carbon scores 81.5%. Google published these results for different tests, with distinct metrics and conditions that should not be collapsed into one ranking.

For cybersecurity evaluations, Google reports 88.2% on SEC-bench Pro for finding vulnerabilities, 54.8% on ExploitGym for exploitation, and 71.0% on CWE-bench for patching. The company says the external evaluations were run without guardrails. The charts identify some conditions—such as per-task time caps for ExploitGym and agent/tool labels for CWE-bench—and the methodology expands on those conditions. Carbon and Argon use the highest thinking settings unless otherwise noted; Google cites the highest available or best published reasoning results for competitors. Google’s table credits the [Gemini 4 Carbon evaluations methodology](https://deepmind.google/models/evals-methodology/gemini-4-carbon).

In ExploitGym, Carbon’s 54.8% uses a ten-hour cap per task; at six hours, Google reports 53.5%. Opus is listed at six hours and Astra without a cap. These are different budgets, not a comparison with identical runtime.

Google also cites internal safety and coding evaluations. These are company-reported results, not independent audits. A 0.0% attack-success figure in the Gray Swan chart applies to the evaluated attacks at the displayed k=15 level; it does not mean the model is immune to attacks. The post also says successful cheating on Terminal-Bench 4.0 was 0.0%, while cheating attempts were 0.4%.

| Benchmark | Gemini 4 Carbon | Gemini 4 Argon | GPT-6 Astra | Claude Opus 5.5 | Claude Fable 5.1 |
|---|---|---|---|---|---|
| AA Index v5.0 | 59 | 50 | 57 | 58 | 49 |
| Vals Index v2.2 | 70.4% | 69.2% | 65% | 70.2% | — |
| SWE-sweep | 9.6% | 4.3% | 9.5% | 3.6% | 2.9% |
| FrontierSWE 2 | 61.2% | 55% | 65.5% | 62.3% | 56.3% |
| Terminal-Bench 4.0 (AA) | 65.2% | 57.1% | 59.1% | 59.6% | 52% |
| Agents’ Last Exam v1.1 — Score | 77.4% | 69% | 73.1% | 71.6% | — |
| OSWorld-2.1 — Offline subset | 81.5% | 74.7% | 80.9% | 87.2% | 76.8% |
| AutomationBench — Zapier | 53.9% | 51.3% | 41.4% | 42.5% | 31.4% |
| Vending-Bench 2 — Bank balance (USD) | 16484 | 13718 | 15515 | 9235 | 5422 |
| Chartography | 72.5% | 71.6% | 71% | 66.3% | 46.2% |
| Blueprint-Bench 2 — Similarity score | 58.6% | 54.4% | 49.7% | 51.2% | 41.9% |
| GraphWalks — 256k to 1M, BFS (F1) | 96.2% | 84.2% | 71.8% | 66.8% | 65% |
| RiemannBench | 80% | 76% | 72% | 69.6% | 65.6% |
| Terminal-Bench-Science 0.1 (AA) | 64.8% | 52.9% | 63.3% | 59% | 43.3% |
| CWE-bench — Pass@1 / Pass@4 tie-break | 71% | 68% | 68% | 67% | 58% |
| ExploitGym — Causal pass rate | 54.8% | 42% | 42.4% | 34.5% | — |
| SEC-bench Pro — Pass@1 | 88.2% | 83.4% | 85.4% | — | — |

Source: Google, Gemini 4 Carbon announcement and comparison graphics: https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-carbon/
Figures reported by Google; unspecified configurations are not inferred. AA Index is the published index, not a computed average. Vending-Bench is bank balance in USD.

### What changes for long-running tasks

Google says its internal teams have used Carbon for days to investigate bugs and complete multi-step engineering work. That is an internal company observation, not a HeyMark test or a duration guarantee for any task. If you use agents, the useful question is whether they finish the job while respecting its constraints; hours of runtime alone do not establish that. [Google’s announcement and examples](https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-carbon/).

## Pricing and access

Google announced introductory pricing of $2 per million input tokens and $10 per million output tokens. Cached input tokens are 95% off the input-token price. After the introductory period, pricing will be $4 and $20 per million, respectively; Google has not said when the initial pricing ends.

Access starts this week with a limited group of cyber defenders. Google plans to expand gradually to trusted partners through its Fairwind Program. Developer, business, and consumer access is expected later, with no specific public date.

| Announced rate | Input USD / 1M tokens | Output USD / 1M tokens |
|---|---:|---:|
| Introductory | 2 | 10 |
| After the introductory period | 4 | 20 |

No public end date is given for introductory pricing. Google announces 95% off cached input, not an additional discount on output. Uncached example: 1,000,000 input tokens and 250,000 output tokens = $4.50 introductory or $9.00 later. This is not a measured bill or confirmation of public access. Source: https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-carbon/

## What is known—and what is not

Carbon follows Argon by two weeks, according to Google. The sequence signals a brisk release cadence, but does not mean Carbon replaces Argon or leads every task. Google’s table shows different strengths: Astra leads the cited FrontierSWE 2 result, while Opus 5.5 leads on OSWorld-2.1 offline.

For now, the available information consists of results reported by Google, announced pricing, and a phased rollout. Google has not given a general availability date. It does publish methodology: OSWorld uses 82 offline tasks and, for Gemini, the highest score across three runs, with one attempt per run; Agents’ Last Exam uses 122 tasks, a five-hour window and the October 17 task version without the October 26 update. Those conditions matter beyond a headline score.

## Sundar Pichai’s original announcement

[Sundar Pichai introduces Gemini 4 Carbon · October 28, 2026](https://x.com/sundarpichai/status/2115489219804352921). The post confirms initial access for cyber defenders and a later expansion to Fairwind partners.

## Plan your next post in HeyMark.

Keep the idea, review the draft with your team, and see how it performed in the accounts you connected.

[Start free](https://app.heymark.ai) · [See how it works](https://heymark.ai/en/#product)

## Developer and agent resources

- [HeyMark MCP documentation](https://heymark.ai/en/mcp/)
- [llms.txt](https://heymark.ai/llms.txt)
- [Full site content for language models](https://heymark.ai/llms-full.txt)
- MCP protocol endpoint: `POST https://mcp.heymark.ai`
- OAuth protected-resource metadata: [/.well-known/oauth-protected-resource](https://mcp.heymark.ai/.well-known/oauth-protected-resource)
- MCP server card: [/server-card](https://mcp.heymark.ai/server-card)