A week after launch, the first independent numbers for Google’s Gemini 4 Carbon are in, and they back up a good part of Google’s claims. On the Vals Index it scores 70.45%, second of 32 models, and it ranks first on Vals’ Finance Index, Legal Research Bench and Tax Agent Bench. Google itself reports a 59 on the Artificial Analysis Intelligence Index, which it says is first place. The catch: almost nobody can use it yet. Carbon is limited to a group of cyber defenders through Google’s Fairwind Program. Here is what the benchmarks say, how Carbon fits next to Gemini 4 Argon, and when you might get access.
- Announced: 28 October 2026, two weeks after Gemini 4 Argon.
- Focus: cybersecurity (vulnerability discovery, patching) and long-running engineering tasks.
- Access: cyber defenders only via the Fairwind Program; trusted partners next. No public date.
- Independent scores (Vals): 70.45% Vals Index (#2/32), #1 Finance Index, #1 Legal Research Bench, #1 Tax Agent Bench, 63.1% Terminal-Bench 4.0 (#4).
- Google’s claim: 59 on the Artificial Analysis Intelligence Index (#1), and more than 2x Argon’s success rate at writing secure code.
- Specs: 1M-token context and output, text/image/video/file input. Pricing is reported differently by different sources (see below).
What Is Gemini 4 Carbon?
Carbon is Google’s new top model, positioned above Gemini 4 Argon, which reached more paying users and developers in mid-October. Google built it for two jobs: cybersecurity, where it claims state-of-the-art results on vulnerability discovery, exploitation and patching, and long-horizon engineering, meaning coding and agent tasks that run for a long time without a human stepping in.
Because of those cyber skills, Google is releasing it carefully. The first users are a subset of cyber defenders in the Fairwind Program. Google says it will gather feedback and keep building guardrails before widening access to more trusted partners, and that it is taking part in the US government’s voluntary pre-release model testing. This is the same “defenders first” pattern other labs have used for their most capable cyber models this year.
The Benchmarks
Vals AI runs its own private test sets, which makes its numbers harder to game than vendor-reported scores. Here is Carbon’s main Vals page:
| Benchmark (Vals) | Score | Rank |
|---|---|---|
| Vals Index | 70.45% | 2 of 32 |
| Vals Finance Index | 67.05% | 1 of 35 |
| Vals Legal Index | 75.93% | 3 of 33 |
| Legal Research Bench | 55.77% | 1 of 89 |
| Tax Agent Bench | 51.78% | 1 of 80 |
| EMB v1.1 | 76.47% | 2 of 37 |
| Finance Agent v2.1 | 67.16% | 5 of 35 |
| Terminal-Bench 4.0 | 63.13% | 4 of 59 |
| Vibe Code Bench v1.1 | 89.06% | 8 of 123 |
A few honest notes. Vals’ own update log from launch day shows slightly different numbers for some rows (for example Tax Agent Bench at 48.86% and third place), so these may still move as runs finish. Carbon is strongest in finance, legal and tax agent work, and good but not leading in coding: fourth on Terminal-Bench 4.0 and eighth on Vibe Code Bench. The Vals run also took almost 48 minutes and about $10.51 per test at high reasoning effort, so this is a slow, expensive thinker.
Google’s own headline claims are a 59 on the Artificial Analysis Intelligence Index (first place) and more than twice Argon’s success rate on its internal secure-code tests. Those are vendor-reported. Third-party tracker Opper lists Carbon with an intelligence score of 60, ahead of OpenAI’s GPT-6.1 Terra at 51.
Artificial Analysis Breakdown
Artificial Analysis (AA) runs its own broad test suite. Its results for Carbon, as republished by Opper, put it at number 1 of 507 models:
| AA measure | Gemini 4 Carbon |
|---|---|
| Intelligence Index | 60.0 (Google’s own figure: 59) |
| Output speed | ~88 tokens/s |
| Terminal-Bench 4.0 / FrontierCode 2.0 | 65% / 41% |
| AutomationBench / GDPval / GDP.pdf | 78% / 62% / 30% |
| Briefcase (Elo) | 1611 |
| Terminal-Bench Science / CritPt (hard) / HLE Diamond | 65% / 69% / 69% |
| Long-context reasoning | 79% |
| MMMU-Pro (multimodal) | 89% |
| Omniscience (knowledge) | 53.5 |
| Domain scores | Finance 62.6, Legal 62.0, Engineering 65.3, Economics 59.6, Strategy and operations 60.9 |
So two independent testers, Vals and AA, both place Carbon at or near the top. The main weakness AA flags is hallucination (see the Argon comparison below).
Gemini 4 Carbon vs Gemini 4 Argon
Argon (released 13 October 2026) is the Gemini 4 model most people can actually use today. Here is how the two compare, using data compiled by Gradually AI from AA and other boards:
| Gemini 4 Carbon | Gemini 4 Argon | |
|---|---|---|
| Status | Limited access (Fairwind) | Generally available |
| AA Intelligence Index | 59.4 | 50.4 |
| HLE Diamond | 69.1% | 56.7% |
| Terminal-Bench | 65.2% | 57.1% |
| Vibe Code Bench v1.1 | 89.1% | 91.9% |
| Hallucination rate (AA Omniscience) | 11.4% | 4.3% |
| Speed | ~100 tokens/s | ~82 tokens/s |
| Context window | 1M tokens | 1.05M tokens |
| Price (per M input / output) | Announced $2 / $10 intro, then $4 / $20 | $2 / $10 until 30 Nov 2026, then $4 / $20 |
This also explains the pricing confusion: $2/$10 is the introductory rate and $4/$20 the standard rate, the same structure Google used for Argon. Carbon is clearly the stronger reasoner and agent; Argon is slightly better at quick “vibe coding” and hallucinates far less, which matters for factual Q&A and research.
Carbon vs Claude, GPT and Open Models
| Model | Price (per M in / out) | Context | Access | Note |
|---|---|---|---|---|
| Gemini 4 Carbon | $2 / $10 intro, $4 / $20 standard | 1M | Cyber defenders only | #1 on AA Intelligence Index (~60) |
| Claude Opus 5.5 | $4 / $20 | 1M, 128K out | Public API | Ranked best overall in Gradually AI’s combined index |
| GPT-6.1 Terra | $1 / $5 | 1.1M | Public API | Opper intelligence score 51 |
| Gemini 4 Argon | $2 / $10 intro | 1.05M | Public | AA ~50.4, lowest hallucination here |
| Tencent Hy4 | $0.90 / $2.70 | Not published | Public API, open weights promised | AA 34.6, strong agentic score for the price |
Rankings depend on the test: AA puts Carbon first, while Gradually AI’s combined index still names Claude Opus 5.5 as the best overall model. For pricing and coding detail on Claude, see our DeepSeek vs Claude comparison, and for Tencent’s new model, our Tencent Hy4 report.
Specs and Pricing
- Context window: 1M tokens, with up to about 1M output tokens.
- Input: text, images, video and files.
- Weights: closed.
- Pricing: sources disagree. Reuters (via Khaleej Times) reported $2 per million input and $10 per million output, matching Argon’s introductory price; Vals lists $4 and $20. Since there is no public API yet, treat both as provisional until Google publishes a pricing page.
For context on how Google’s models compare with Claude and DeepSeek on price and coding, see our DeepSeek vs Claude comparison, which we will update as Carbon opens up.
When Can You Use It?
Google has given no public timeline. The sequence it described is: cyber defenders in Fairwind first, then more trusted partners, then (presumably) developers and paying Gemini users. Based on how Argon rolled out, a wider developer preview within weeks to a couple of months is plausible, but that is my guess, not Google’s statement. If you are a security team, the Fairwind Program is the route in.
What to Use Today Instead
- Need Google’s best you can actually use: Gemini 4 Argon, in the Gemini app for paying users and through the Gemini API and Google AI Studio for developers. It is cheaper during its intro period and hallucinates less than Carbon.
- Long-running coding agents: Claude Opus 5.5 or Sonnet 5.5 remain the safe picks with public APIs.
- Tight budget: GPT-6.1 Terra ($1/$5), DeepSeek’s Flash models or Tencent Hy4.
- Privacy or offline use: an open model on your own GPU. Our best Ollama models guide ranks options by VRAM, and installing Ollama takes about five minutes.
- Security teams: ask Google about the Fairwind Program, which is the only route to Carbon right now.
My Take
The independent Vals results are the real story here. Many “number one” launch claims fall apart on third-party tests; Carbon’s mostly hold up, especially on professional agent tasks in finance, law and tax. For everyday users nothing changes yet, but for anyone choosing an API for long-running business agents, Carbon belongs on the shortlist once it opens. Meanwhile, open models keep closing the gap at a fraction of the price, as our best Ollama models list shows.
Frequently Asked Questions
Is Gemini 4 Carbon available to the public?
No. Access is limited to a subset of cyber defenders through Google’s Fairwind Program, with trusted partners next. There is no public release date.
Is Gemini 4 Carbon better than Gemini 4 Argon?
Google positions Carbon as its top model and says it writes secure code at more than twice Argon’s success rate in internal tests.
What is Gemini 4 Carbon’s context window?
Vals lists a 1M-token context window and up to about 1M output tokens.
How much will Gemini 4 Carbon cost?
The announced pricing is $2/$10 per million input/output tokens as an introductory rate, then $4/$20, the same structure as Gemini 4 Argon. That is why Reuters reported $2/$10 and Vals lists $4/$20. There is no public API yet.
Is Gemini 4 Carbon better than Claude Opus 5.5?
On the Artificial Analysis Intelligence Index, Carbon ranks first. Other combined rankings, such as Gradually AI’s, still put Claude Opus 5.5 on top overall. Carbon also has a higher hallucination rate than Gemini 4 Argon. It depends on the task, and Carbon is not publicly available yet.
How fast is Gemini 4 Carbon?
Artificial Analysis measured about 88 tokens per second, and Gradually AI lists about 100, faster than Argon’s roughly 82.
Should I wait for Carbon or use Argon now?
Use Argon now. It is available, cheaper during its intro period, and hallucinates less. Switch to Carbon for hard reasoning and long agent tasks once Google opens access.
Sources
- Vals AI: Gemini 4 Carbon model page
- Khaleej Times (Reuters): Google launches Gemini 4 Carbon
- Let’s Data Science: Google introduces Gemini 4 Carbon with cyber defenders first
- Opper: AI model release tracker
- Opper: Gemini 4 Carbon (Artificial Analysis scores)
- Gradually AI: Gemini 4 Carbon vs Gemini 4 Argon
