On October 29, Gemini 4 Carbon scored 59 on Artificial Analysis' Intelligence Index v5, surpassing Claude Opus 5.5 and GPT-6 Astra. Yet it remains available only to selected users, with no public release timetable. The contrast highlights a central problem: model capabilities are improving rapidly, but commercial availability remains the bottleneck for deployment.
What this is
Carbon is Google's new flagship model. It supports text, image, video, and voice input, with a 1-million-token context window—the amount of information a model can process in one pass. It took first place in four of nine evaluations, covering complex reasoning, scientific tasks, and knowledge reliability. In coding agents—models that can call tools to complete multistep tasks—it scored 59, just 3 points behind Claude Code paired with Opus 5.5 at 62.
But Carbon is not the leader across the board. Its official pricing is $4 per million input tokens and $20 per million output tokens. Its average cost per task in the evaluation was $7.76, about 23% lower than Opus 5.5 but roughly 24% higher than GPT-6 Astra. More notably, its hallucination rate—the share of plausible-sounding but incorrect content generated by the model—rose from 4% for the previous Argon generation to 11%. Even so, its absolute rate remains the second-lowest among the models tested.
Industry view
We believe Carbon's most important signal is not its No. 1 ranking but the fact that Google has finally entered the first tier in coding capability. At the same price, its Intelligence Index rose from 50 for Argon to 59, indicating that the pace of model iteration is still accelerating. Enterprises need to reassess their model choices for coding assistants, long-document analysis, and automated workflows.
The risks are equally clear: leading a benchmark does not mean being reliable in production. A Hacker News user reported that Argon fabricated entire sections when summarizing project progress. Carbon is not yet broadly available, so its real-world performance, latency, and system stability remain unverified. The mix of tasks used in the evaluation also cannot fully represent an enterprise's own workload. The more practical issue is supply: without a public availability timetable, even a highly capable model cannot translate into procurement contracts or deployment plans.
Impact on regular people
For enterprise IT: If your team is testing coding assistants or knowledge-Q&A products, add Carbon to your watchlist. Once it is formally available, compare accuracy, total cost, and response time using real business data rather than relying only on aggregate rankings.
For workplace professionals: Carbon still has some distance to travel before it becomes a daily tool available to everyone. The more important development right now is that the capability gap between models in coding and knowledge work is narrowing. Teams can begin establishing AI usage policies, with particular attention to verification, permissions, and cost boundaries.
For the consumer market: Smarter models that are more willing to answer may improve writing and learning experiences, but they will also make it easier to produce “plausible” errors. As consumers choose AI products, reliability and verifiability will matter as much as model speed and price.