Google's new Gemini 4 Carbon model invents false information about half as often as its predecessor, but a detailed independent review says it is still not reliable enough for real coding work and trails Anthropic's Opus. Developer Theo Browne, speaking on a podcast, said the model's false-claim rate fell from 36% for the previous Gemini model to 18%. Browne uses the word 'fabricate' for what is often called hallucination: a model stating things as fact when they are not true.
Browne said the 18% figure matched what DeepMind, Google's AI research lab, engineers had privately told him they measured before launch. This is his account of a private conversation, not a published Google benchmark, and it should be read that way. The bigger gap, in his view, is with rival models. He said most competing models, including Anthropic's Opus, fabricate at 0–1%, and a few reach three or four percent.
Fewer false claims, but still far behind
Browne's main complaint is fabrication in codebase analysis and code review. On a review of a Rust codebase, the previous Gemini model made 36 claims, 24 of them false. Carbon made 79 claims, 23 of them false, which he described as more opinions delivered with more confidence rather than fewer errors.
In a second test, an analysis of his own T3 codebase, the earlier model failed 47 of 92 claims, while Carbon failed 23 of 86. Browne linked part of the earlier problem to bugs in the integration layer between his tool and Google's model, which did not pass details such as line numbers the way the official command-line tool does. DeepMind has since fixed those bugs, but Browne said Carbon still invents column names, line numbers for code it did not properly read, and architectural details that are not true.
Browne also argues that official hallucination benchmarks do not measure this failure mode at all. In his view, the trait that makes a coding model untrustworthy in practice is exactly what leaderboards tend to skip.
Speed and pricing
Browne rejects the idea that Carbon is a fast model, arguing that tokens per second is a number users do not actually experience. In Artificial Analysis data, Carbon's average task took 22 minutes on low reasoning effort, the setting that controls how much a model thinks before answering, and 27 minutes on high. Opus ranged from 3 minutes 1 second on low to 25 minutes at max, and Browne called using Carbon 'genuinely obnoxious' and 'miserable to use'.
Carbon launches at $2 per million input tokens and $10 per million output tokens, with cache reads at 10 cents per million. Tokens are the chunks of text a model reads and writes, and a cache stores repeated input so it does not have to be processed again. The Gemini 4 Carbon price is introductory: it doubles on December 1, 2026, to $4 input and $20 output per million tokens, with cache reads rising to 20 cents. Browne said the post-promotion price is Opus-level pricing for a model he finds much less pleasant to use.
Browne said the hidden cost is caching. He contrasted Gemini with OpenAI and Anthropic's Claude, which he said have explicit cache writes, and in Claude's case configurable cache lifetimes. Gemini, he said, caches 'when they feel like it'. In his tests, a 5-second gap between requests carried a 20% chance of a cache miss, a 30-second gap 34%, and longer gaps more still, which could raise the number of requests by up to 50% in the worst case.
His broader argument is that token price is beside the point. What matters, he said, is cost per task, cache efficiency and how many tokens a model burns to finish the job, and he said Carbon 'burns tokens like nothing I've ever seen'.
The one case where Carbon won
Browne reverses course in the final third of his review, over one specific workflow: a security fix. The turning point is a vulnerability report filed against T3 Code by an independent developer named Brandon, who had access to Anthropic's Mythos model through a security programme. Browne said Opus 4.8 produced nothing useful, and Mythos diagnosed the problem correctly but proposed a fix he found unacceptable, which would have required a user to walk to a second machine and approve a link with a six-digit code.
Browne says Carbon beat every rival he tested on that one fix, and it is the single place where his verdict on the model changes.
Aasan alfaaz mein
Google ke naye Gemini 4 Carbon model ne ghalat maloomat dene ki sharah 36% se kam kar ke 18% kar di hai. Lekin developer Theo Browne ke review ke mutabiq ye ab bhi Anthropic ke Opus jaise models se peeche hai, jo 0 se 1 feesad tak ghalat claims dete hain. Is ki introductory qeemat 30 November 2026 tak kam hai aur 1 December 2026 se dugni ho jayegi. Browne ke mutabiq AI tools ka asal kharcha har task ki cost hai, sirf sticker price nahi.
