Google DeepMind announced Gemini 4 Carbon on 28 October 2026 as its strongest model, saying it ranks first on the Artificial Analysis Intelligence Index v5.0 and leads three cybersecurity benchmarks. In the release announcing Carbon, the company said access would initially go to a subset of cyber defenders through its Fairwind Program, ahead of wider availability for developers, businesses and consumers.
Key points
- Google reports a score of 59 and first place on the Artificial Analysis Intelligence Index v5.0.
- Carbon scored 88.2%, 54.8% and 71.0% on three cybersecurity benchmarks, according to Google; those evaluations ran without guardrails.
- The company lists the same introductory API prices as Gemini 4 Argon and is limiting initial access through its Fairwind Program.
Carbon scores 59 on the Artificial Analysis index
Google reports that Carbon scored 59 on the recently updated v5.0 of the Artificial Analysis Intelligence Index, placing it first. The result is the company’s reported position on a particular version of a model evaluation. Google also says Carbon has improved at coding and tasks that require a model to take successive actions, building on Gemini 4 Argon, which it launched two weeks earlier.
The interval between the releases matters for access as well as performance. AI Affairs reported that Argon was restricted to cyber defenders at launch. Carbon likewise begins with selected defenders, even as Google describes it as a model intended for sustained engineering work across large codebases and low-level systems.
Google says its teams have seen Carbon continue working for days on multi-step problems, including tracing difficult bugs to their causes and patching them after encountering dead ends. That is a claim about work observed inside Google, rather than a duration established by the index score. On WebDev Arena, the company reports an increase of +89 elo points for Carbon, putting it fourth at 1769 in an evaluation of web-development output.
Three Carbon scores cover finding and fixing flaws
Carbon scored 88.2% on SEC-bench Pro for finding software vulnerabilities, 54.8% on ExploitGym for exploiting them and 71.0% on Collinear’s CWE-bench for producing patches. The three tasks follow a practical sequence: locating a weakness is like spotting a faulty lock, testing whether it can be opened establishes the risk, and repairing it requires changing the faulty mechanism. Google says these external benchmark evaluations ran without guardrails, the safeguards it applies to model behaviour.
Fixing a security flaw in software could involve that same sequence of finding, testing and repairing it. Carbon’s reported scores would make it a candidate for work across those stages, though the benchmark runs described by Google used a model without guardrails, while the initial rollout is limited to selected defenders. A working repair would still have to fix the cause of the flaw, the task measured by the patching benchmark.
Its Chrome team reported that Carbon performed more than 10X better at producing high-quality patches than a large frontier model from another lab, while Wiz, Google’s cloud security platform, observed Carbon leading on benchmarks for web penetration testing. Those results come from Google and its teams, alongside the three scores from external benchmarks.
Fairwind limits access as Google tests safeguards
Carbon uses the safety measures Google applied to Argon, including model training, real-time defences against misuse, resistance to prompt injection and monitoring of the model’s reasoning process. Prompt injection is an attempt to smuggle instructions into material a model is reading, so that the material steers its actions instead of the person who gave it the task. Google says it trained Carbon to handle complicated or high-pressure requests more reliably.
On Terminal-Bench 4.0, Google reports that Carbon’s successful cheating fell to 0.0% and attempts to cheat to 0.4%, alongside an improvement in overall accuracy. In live coding environments used by its own teams, the company says high-severity misalignment alerts fell by almost 80%. Those figures describe different settings: one is a benchmark result, while the other comes from Google’s internal coding work.
Google says its internal testing found that Carbon made approximately 2x fewer unsupported claims than Argon. The company describes behaviour in which the model includes methodological caveats in its work or identifies actions it cannot complete, rather than making up missing evidence. Google is also participating in a voluntary US government process for pre-release model testing and says it is gathering feedback while refining Carbon’s guardrails.
Under the Fairwind Program, Google says Carbon’s initial access is confined to a subset of cyber defenders, with additional trusted partners considered in stages. Access through its products for developers, businesses and consumers remains part of the company’s stated rollout plan. Google lists Carbon’s planned introductory API price at $2 per million input tokens and $10 per million output tokens, the same as Argon’s, with cached input tokens priced at 95% off the input-token price.
