● Pick Gemini 4 Carbon if you need the strongest published cyber benchmarks and you can get through the Fairwind Program.
● Pick Mistral Large 4 if you need the weights on your own hardware, today, with no approval queue.
● The real split Both claim cyber defence. Only one of them will let you actually download it.
Availability is the whole comparison. Gemini 4 Carbon is not generally released - Google is giving it to a subset of cyber defenders first and expanding through the Fairwind Program. Mistral Large 4 weights are on Hugging Face. A benchmark you cannot run is a claim, not a capability.
| Gemini 4 Carbon | Mistral Large 4 |
|---|
| Released | 28 October 2026 | Weights released 29 October 2026 |
| Availability | Subset of cyber defenders, expanding via Fairwind Program | Open weights on Hugging Face, plus Mistral API |
| Self-host | No | Yes, including on-premises |
| Artificial Analysis | #1 on Intelligence Index v5.0, score 59 | Claims top open model on the Coding Agent Index |
| Vulnerability finding | SEC-bench Pro 88.2% | Not published as a benchmark score |
| Exploitation | ExploitGym 54.8%, stated as new state of the art | Not published as a benchmark score |
| Patching | CWE-bench 71.0% | Not published as a benchmark score |
| Prompt injection | Gray Swan 0.0% attack success rate, vs 0.7% for Argon | Not published |
| Web development | WebDev Arena #4, 1768 Elo, up 89 points | Not published |
| API price | Intro $2 / $10 per million in / out, cached input 95% off. Rises to $4 / $20 after the introductory period | Not stated in the weights announcement |
| Efficiency claim | Not published | 47% fewer tokens per task than GLM-5.3 |
Gemini 4 Carbon: the benchmarks are the product
Google describes Carbon as its strongest model with state-of-the-art cybersecurity and long-horizon capabilities, and the release reads like a security datasheet rather than a general model launch. It takes the top of the Artificial Analysis Intelligence Index v5.0 at 59.
The three security numbers are the ones to hold onto: 88.2% on SEC-bench Pro for finding vulnerabilities, 54.8% on ExploitGym for exploitation, which Google calls a new state of the art, and 71.0% on CWE-bench for patching. Google also reports a 0.0% attack success rate on the Gray Swan indirect prompt injection benchmark, against 0.7% for Argon.
The limitation
You probably cannot use it. Carbon is going to a subset of cyber defenders first, widening to trusted partners through the Fairwind Program, with developers, businesses and consumers promised later pending feedback and guardrail work. Introductory pricing of $2 and $10 per million tokens is real, but it applies to whoever gets in - and it rises to $4 and $20 once the introductory period ends.
Mistral Large 4: the weights are the product
Mistral released the weights for its trillion-parameter Large 4, nicknamed Le Chonk, on 29 October through Hugging Face and its own API. The pitch is sovereign cyber defence for enterprises and governments - detect, analyse and respond to threats across code and IT infrastructure, deployed on your own hardware if that is where it needs to live.
Pierre Stock, Mistral’s VP of Science, made the structural argument rather than a benchmark one: with open weights, access "can’t be cut off by definition." Set against a model distributed through an approval programme, that is the entire difference between these two releases.
Mistral claims Large 4 is comparable to GLM-5.3, widely regarded as one of the strongest open-weight models for cyber work, while using 47% fewer tokens per task, and that it outranks all other open models on Artificial Analysis’ Coding Agent Index.
The limitation
Mistral has not published the security benchmark scores Google has. There is no SEC-bench Pro, ExploitGym or CWE-bench figure to compare against Carbon’s 88.2%, 54.8% and 71.0%. You are being asked to take the cyber capability on the company’s description rather than a number - which, to be fair, you can check yourself, because you can download it.
Which should you use?
| If you... | Pick | Why |
|---|
| Need it running this week | Mistral Large 4 | Weights are downloadable now |
| Have data that cannot leave your network | Mistral Large 4 | On-premises deployment is supported |
| Want published, comparable security scores | Gemini 4 Carbon | Mistral has published none |
| Are a national or regulated buyer | Mistral Large 4 | Access cannot be withdrawn by a vendor |
| Can get into the Fairwind Program | Gemini 4 Carbon | Strongest published numbers in the category |
| Are budgeting per token | Gemini 4 Carbon | Only one of the two with a published API price |
The question neither release answers
Both models are sold as cyber defence. Neither announcement addresses the other half of the equation: a model that scores 54.8% on exploitation is a model that is good at exploitation. Google’s staged rollout through the Fairwind Programme is an implicit acknowledgement of this. Mistral’s answer is that the capability should not be a vendor’s to allocate in the first place.
That is a genuine disagreement about how dangerous capability should be distributed, and it was settled differently by two labs in the space of 24 hours. Whichever you pick, you are picking a side of it.
Frequently asked questions
Is Gemini 4 Carbon available to the public?
Not yet. Google says it is going to a subset of cyber defenders first, expanding to other trusted partners through the Fairwind Program, with developers, businesses and consumers to follow pending further feedback and guardrail work.
How much does Gemini 4 Carbon cost?
The introductory price is $2 per million input tokens and $10 per million output, with cached input tokens 95% off the input price. After the introductory period it rises to $4 and $20 per million.
Can I download Mistral Large 4?
Yes. Mistral released the weights on 29 October 2026 through Hugging Face, and the model is also available via Mistral’s API. It can be deployed on-premises. Mistral has not named a licence in the release, so check the model card before commercial use.
Which model is better at cybersecurity?
On published numbers, Gemini 4 Carbon, because it is the only one with published numbers: 88.2% on SEC-bench Pro, 54.8% on ExploitGym and 71.0% on CWE-bench. Mistral describes Large 4 as state-of-the-art sovereign cyber defence but gives no benchmark scores. That is a gap in evidence, not proof of a gap in capability.
What is Le Chonk?
It is Mistral’s nickname for Mistral Large 4, the trillion-parameter model first introduced earlier in October 2026 and whose weights were released on 29 October.
Which is better value?
If you can self-host, Mistral Large 4, because there is no per-token cost once the weights are on your hardware. If you cannot, Gemini 4 Carbon is the only one of the two with a published API price - but you also need access to it.
Related
Sources
Every figure on this page was read from the sources above on 30 October 2026. Where a vendor has not published a number, this page says so rather than repeating a third-party estimate.