Story
Oct 30, 2026
Google Says Gemini 4 Carbon Is No. 1—But Most Users Still Cannot Try It
Google is pitching Gemini 4 Carbon as proof that it has reclaimed the AI frontier, while independent-minded coverage and benchmark caveats point to a narrower reality: a high-scoring model whose safety-led rollout keeps it beyond the reach of most users.
Two weeks after unveiling Gemini 4 Argon, Google introduced Gemini 4 Carbon as the next step in its fast-paced push to the frontier. Chief executive Sundar Pichai called it evidence of “the pace at which we’re pushing the frontier,” saying the model ranked first on the updated Artificial Analysis Intelligence Index.
1
Google’s case rests heavily on coding and cyber performance. Carbon is said to deliver more than twice Argon’s success rate at generating secure code, while Chrome engineers reported more than a tenfold advantage in producing high-quality patches over another frontier model. Google also says Carbon resisted every prompt-injection attempt in its Grey Swan evaluation.2
That is the triumphal reading: Google DeepMind chief Demis Hassabis described Carbon as state-of-the-art across key benchmarks, combining frontier capability with “deep safety & security research.”
3 A separate assessment placed Carbon atop Intelligence Index v5 with a score of 59, ahead of Anthropic’s Claude Opus 5.5 and OpenAI’s GPT-6 Astra; it also reported gains in reasoning, scientific tasks and coding.4
But the victory lap comes with qualifiers. The same assessment found Carbon does not lead every relevant test: on the AA-Briefcase knowledge-work benchmark, it trailed Grok 4.7, Muse Wonder and Kimi K3.1 Pro. It was also judged relatively expensive and notably verbose.4 Another report noted that the gaps separating leading frontier models are only a few benchmark points, a reminder that leaderboard supremacy is not the same as a decisive real-world gulf.2
For now, Google’s safety argument is also its access problem. Carbon is available only to selected users while the company works through voluntary U.S. testing and expands access via its Fairwind cyber-defense program. A broader release has no firm date, though AI Ultra subscribers and Gemini API developers are expected to be first in line.2
Even Elon Musk offered a terse “Impressive” in response to Pichai’s announcement.
5 Yet the central question remains unanswered: whether Carbon’s claimed crown will hold once more developers can test it outside Google’s controlled rollout.