Elon Musk doesn't usually hand out compliments to Google. But when Gemini 4 Carbon's benchmark numbers went public this week, Musk called the model "impressive" on X, a notable reaction from the head of a company, xAI, that is racing Google for the same frontier-AI customers.
The praise follows Google's release of Gemini 4 Carbon just two weeks after Gemini 4 Argon, its previous flagship. Carbon now ranks first on the Artificial Analysis Intelligence Index, a widely watched industry scorecard, and second only to Anthropic's Claude Opus 5.5 on the Artificial Analysis Coding Agent Index.
The economics are where Carbon stands out. Google says the model costs $7.76 per Intelligence Index task, about 23 percent less than Claude Opus 5.5's $10.04 at its maximum setting, though still roughly 20 percent more than GPT-6 Astra's $6.25. Carbon also gets there using 123,000 output tokens per task, 29 percent fewer than Claude Opus 5.5 needs, according to Google's own figures.
Security work is where the gap widens further. Carbon scored 88.2 percent on SEC-bench Pro, a vulnerability-discovery benchmark, 54.8 percent on the ExploitGym exploitation test, and 71 percent on CWE-bench, which measures the quality of security patches a model writes. Google says its internal testing shows more than double the success rate of Argon at generating secure code, and that Chrome's engineering team found Carbon produced high-quality patches more than ten times as often as "another lab's large frontier model" did, without naming which one.
None of that is available to ordinary developers yet. Gemini 4 Carbon remains under evaluation by the US government and hasn't shipped as a public product, a gap that matters because the AI race moves fast enough to erase an advantage before it's even monetized. Anthropic is expected to release Claude Fable 5.5 before Carbon reaches wide availability, and that model could reclaim the top spot Carbon currently holds on Artificial Analysis's rankings.
That timing is the real story here, more than the benchmark chart. Google has now put out two frontier models in a month, Argon and Carbon, while still withholding the more capable one from the public it's meant to compete for. Whether that's caution around government review, a staged rollout, or both isn't something Google has explained.
What happens next depends on two things outside Google's control: when US reviewers clear Carbon for broader release, and whether Anthropic's Claude Fable 5.5 lands first. Until one of those resolves, Gemini 4 Carbon's benchmark lead is a number on a chart, not a product anyone outside Google and a few government reviewers can actually use.
ADFiled by The AI Desk
Models, assistants and the companies and chips behind them, reported from what was released and what was claimed, with the difference kept clear.
More from this desk →
Be the first to comment
Join the argument. No password, just your email or a passkey.