Google has unveiled Gemini 4 Carbon, its new frontier model, and in the tables it published the company puts it ahead of nearly every rival. The most striking result is in cybersecurity, but the model is not yet open to the public.
According to Google's table, Carbon ranks first on the Artificial Analysis Intelligence Index v5 with a score of 59, ahead of Claude Opus 5.5 at 58 and GPT-6 Astra at 57. In its official announcement Google calls it "our strongest model with state-of-the-art cybersecurity and long horizon capabilities." On the same table Carbon leads 15 of 17 rows; GPT-6 Astra wins FrontierSWE 2 and Claude Opus 5.5 wins OSWorld-2.1.
Image: Google — https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-carbon/
The security numbers
Security is where the model clearly pulls ahead. Carbon scores 88.2% on SEC-bench Pro, a vulnerability discovery test, 54.8% on ExploitGym, which measures exploitation, and 71% on CWE-bench, which covers security patches. According to Google, all results were obtained without guardrails.
Google backs the gains with three examples: on its internal benchmarks Carbon doubles the rate of secure code generation compared with Argon, its predecessor, while the Chrome team reports high-quality patches generated more than ten times better than another lab's large model. Argon came out just two weeks earlier, and across Google's tests Carbon beats it by narrow margins; many observers suspect it is a post-training upgrade rather than a separate model, though Google presents it as new.
Cost per task is another point of comparison. Per Artificial Analysis, one run of the index costs $7.76, about 23% less than Claude Opus 5.5 ($10.04) but more than GPT-6 Astra ($6.25). Carbon also uses 29% fewer output tokens than Opus 5.5.
Who can use it, and at what price
Access is staged. The model goes first to a subset of cyber defenders, then to more trusted partners through Google's Fairwind program, and only later to developers, businesses and consumers, with no date given. It is also under evaluation by the US government. Unite.AI reports that the introductory API price, via Fairwind, is $2 per million input tokens and $10 per million output tokens, rising to $4 and $20 later. Google's standard listing is $4 and $20 per million tokens, with a 50% discount on initial requests and $0.20 for cached input.
The context window is one million tokens, and the model accepts text, images, video and audio as input, returning text output. Carbon arrives a month after Gemini 4 Argon reached paying subscribers, and access decisions follow the same caution. For Gemini users on a standard plan the effect will be indirect: the strongest models stay reserved for the pricier tiers.
Meanwhile Anthropic is working on Claude Fable 5.5, which by some forecasts could overtake Carbon before it becomes widely available.
No comments yet. Be the first!