Gemini 4 Carbon review: Google's claims, independent results and what testers found

Google DeepMind released Gemini 4 Carbon on 10 November 2026, a top-tier model built on Gemini 4 Argon and aimed at long agent runs, multi-step coding and cybersecurity work, priced at $2 per million input tokens and $10 per million output tokens until 1 December. Independent evaluations place Carbon near the top but not first: Artificial Analysis ranks it fourth behind three configurations of Claude Fable 5.2, and reviewers note it is slow, verbose and hallucinates more than Argon. Google's own benchmark claims, including first place on CursorBench 4.0 and Perplexity's WANDR, were published without scores and cannot yet be verified.

Originally published at hamzaahmadaslam.com https://hamzaahmadaslam.com/blog/gemini-4-carbon . How this article was made: drafted with AI assistance from a researched brief, then checked against primary documentation before publishing. Gemini 4 Carbon, often searched as Google Carbon, is Google DeepMind's newest top model. Google announced it on 28 October 2026 for a small group of security teams, then made it generally available on 10 November 2026. It is built on Gemini 4 Argon, which stays on sale beside it. The short version, as of 11 November 2026: Carbon is a strong model at a low launch price, and it leads several independent leaderboards. It does not lead most of them. Artificial Analysis now ranks it fourth, behind three settings of Claude Fable 5.2. It is also slow, it writes a lot of tokens, and it makes up answers more often than Argon does. This review keeps three kinds of evidence apart: what Google says, what outside labs measured, and what the first users wrote after a day with it. Carbon is Google's model for the hardest tasks: long agent runs, multi-step coding and debugging, and scientific research. The model card https://storage.googleapis.com/deepmind-media/Model-Cards/Gemini-4-Carbon-Model-Card.pdf says it uses more thinking tokens than Argon to check its own work. Google's own advice in its launch post https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-4-carbon-now-available/ is that Argon remains the better choice for most software engineering and knowledge work. | Item | Gemini 4 Carbon | |---|---| | Announced | 28 October 2026, for cyber defenders in Google's Fairwind Program only | | Generally available | 10 November 2026 | | Developers | Paid tier of the Gemini API in Google AI Studio, and Gemini for business | | Consumers | Google AI Ultra subscribers, in the Gemini app, AI Mode and Antigravity | | Price until 30 Nov | $2 per million input tokens, $10 per million output tokens | | Price from 1 Dec | $4 input, $20 output | | Context window | Up to 1 million tokens, input and output combined | | Knowledge cutoff | March 2026 | Gemini AI Pro subscribers keep Argon and do not get Carbon in the app. Google has not published regions, rate limits or a Vertex AI date for Carbon. Security teams can apply to the Fairwind Program for a version with safeguards tailored to defensive work, such as checking whether a vulnerability can be exploited and analysing malware. Most of Google's numbers come from its own runs, and its strongest claims are in cybersecurity. VentureBeat https://venturebeat.com/technology/google-unveils-gemini-4-carbon-retaking-the-worlds-no-1-spot-on-updated-harder-benchmarks-but-access-is-limited reported that Google ran these cyber tests without guardrails, through the Fairwind access path. | Benchmark Google's run | Gemini 4 Argon | Gemini 4 Carbon | |---|---|---| | SEC-bench Pro finding vulnerabilities | 83.4% | 88.2% | | ExploitGym | 42.0% | 54.8% | | CWE-bench v1 | 68.0% | 71.0% | | Agentic honesty, 375 tasks | 70% | 92% | Google also says Carbon ranks first on CursorBench 4.0, Perplexity's WANDR and Zapier's AutomationBench at less than half the cost per task of the next best outside model. It published no scores for CursorBench or WANDR, so those two claims cannot be checked yet. Zapier did publish its AutomationBench report https://zapier.com/es/benchmarks/gemini-carbon-4-report : Carbon scored 53.9% at High thinking, for about $2.12 per task, in runs made before the public release. On safety, Google reports that external red-teamers at Trajectory Labs found no new universal jailbreaks in 132 hours and more than 13,000 requests. The same post says resistance to prompt injection "remains an area for continued improvement." Five outside groups had published Carbon results by 11 November. Their numbers mostly agree: Carbon is near the top, and Claude Fable 5.2, released a day before Carbon's general availability, is ahead on the broad indexes. | Evaluator | Carbon result | For comparison | |---|---|---| | Artificial Analysis, Intelligence Index | 60, fourth | Fable 5.2 max 65, Opus 5.5 max 59, GPT-6 Astra 57, Argon 50 | | Artificial Analysis, Coding Agent Index | 59 | Claude Code with Opus 5.5 max 62 | | Vals AI, Vals Index | 70.45%, third of 36 | Fable 5.2 71.31%, first | | Arena, text | First, 1,526 preliminary, 4,282 votes | Argon 1,522, Opus 5.5 High 1,506 | | Arena, web development | Fourth, 1,768 preliminary | Fable 5.2 xHigh 1,836, first | | ARC Prize, ARC-AGI-2 | 94.2% | Fable 5.2 93.8% | | ARC Prize, ARC-AGI-3 | 65.1% | Fable 5.2 81.2% | Artificial Analysis https://artificialanalysis.ai/leaderboards/models ranked Carbon first at launch on 28 October, with 59 points. That ranking is out of date. Fable 5.2 has since scored 65, and Carbon's own score moved to 60. Articles that still call Carbon the number one model are quoting the October result. Vals AI https://www.vals.ai/models/google gemini-4-carbon puts Carbon first on its Finance Index, at 67.05%, and fifth on Terminal-Bench 4.0, at 63.13%. One test of the Vals Index costs $10.51 with Carbon. The Arena text leaderboard https://arena.ai/leaderboard/text has Carbon in first place, but the score is marked preliminary and rests on about 4,300 votes. Its 4-point lead over Argon is smaller than the plus or minus 10 points shown beside Carbon's score. The ARC Prize leaderboard https://arcprize.org/leaderboard shows the widest split. Carbon edges Fable 5.2 on ARC-AGI-2, then trails it by 16 points on the newer ARC-AGI-3. It was cheaper to run there: $7.2K against $21.9K. No METR time-horizon result and no public SWE-bench Verified score existed for Carbon on 11 November. Yes, on the one independent test that measures it. In its launch analysis https://artificialanalysis.ai/articles/gemini-4-carbon-new-number-1 , Artificial Analysis found that Carbon answers more knowledge questions correctly than Argon, but that its hallucination rate rose from 4% to 11%. Artificial Analysis notes that 11% is still the second lowest of any model it has tested, behind Argon. In practice, Carbon knows more and guesses more. If your use case punishes a confident wrong answer, such as legal, medical or financial text, test that rate on your own questions before you switch. Carbon is slow and wordy. Artificial Analysis https://artificialanalysis.ai/models/gemini-4-carbon measured about 80 output tokens per second and ranks it 70th of 73 models for verbosity: it produced 205 million output tokens to complete the index. At High thinking, its leaderboard lists 67 seconds before the first answer token. Per-token prices are low until 30 November, but the bill depends on how many tokens a task uses. Artificial Analysis lists these costs per task on its index, at Carbon's standard price of $4 and $20: | Model and setting | Index score | Cost per task | |---|---|---| | Gemini 4 Carbon high | 60 | $7.76 | | Claude Fable 5.2 max | 65 | $26.54 | | Claude Fable 5.2 high | 62 | $9.08 | | Claude Opus 5.5 max | 59 | $10.04 | | GPT-6 Astra max | 57 | $6.24 | Two things follow. Carbon is cheaper per task than Fable 5.2 at a similar level, but Fable 5.2 at High scores 2 points more for about $1.30 more per task. And the launch price halves every figure in Carbon's row until 30 November, so a December budget built on November invoices will be wrong by a factor of two. Few people had used Carbon for more than a day when this was written, so treat these as first impressions. Simon Willison tested it on 10 November https://simonwillison.net/2026/Nov/10/llm-gemini/ with his usual SVG drawing prompt at three thinking levels, and built a 3D scene in Antigravity that needed a second prompt to become a standalone HTML file. He gave no ranking, and noted the price is "scheduled to double in price to $4/$20 on December 1st." The Hacker News launch thread https://news.ycombinator.com/item?id=50442303 was mixed: Bito, a company that sells an AI code review tool, published its own comparison inside Claude Code https://bito.ai/blog/gemini-4-carbon-vs-gemini-4-argon-vs-gemini-3-1-pro/ : Carbon scored 79.5% against 68.2% for Argon and 85.0% for Opus 5.5, while taking about twice as long as Argon per session. It is one vendor's test, but it matches the pattern above: better than Argon, behind Anthropic's top models on coding, slower. The Safety & Security Report https://storage.googleapis.com/deepmind-media/gemini/gemini 4 carbon safety report.pdf has one detail most launch coverage left out. Google judged that Carbon was not a big enough step up from Argon to need a new full frontier-safety assessment, and it did not run a new round of outside capability testing for this release. It carried Argon's findings forward instead. Those findings matter. Google says it cannot rule out that Carbon, without its outside safeguards, reaches the danger thresholds in its Frontier Safety Framework for cyberattack uplift and for biological information. It says its safeguards bring the remaining risk below those thresholds, and that it strengthened its cyber protections for Carbon because the model's exploit scores went up. That is a defensible process, and Google published it. It also means the safety evidence for Carbon is mostly Argon's evidence plus targeted red-teaming of the safeguards. If you deploy Carbon in a product that browses the web or reads customer documents, treat prompt injection as an open risk, as Google itself does. It depends on what you run now and what the task is. Twenty of your own tasks will tell you more than any table on this page. MODEL SWITCH CHECKLIST: Gemini 4 Carbon Current model: Candidate: Gemini 4 Carbon Thinking level tested: Low / Medium / High 1. TASK SET 20 real tasks from the last month Include 5 the current model got wrong Include 5 where a wrong answer would be costly 2. FOR EACH TASK, RECORD Correct? Yes / Partly / No Invented a fact? Yes / No Input tokens: Output tokens: Cost at the December price $4 / $20 : Time to first token: Time to finish: 3. COMPARE Tasks fixed by Carbon: Tasks broken by Carbon: Cost per correct answer, old vs new: 4. DECIDE Switch all / switch hard tasks only / stay Re-test date after 1 December : Price the test at the December rate. A model that looks cheap in November can cost twice as much a month later. The model matters less than the limits around it. An agent running on Carbon still needs fixed permissions, a person's approval for anything that moves money or data, and a log you can audit. For how the main rival compares, see the Claude Fable 5.2 benchmarks and price https://hamzaahmadaslam.com/blog/claude-fable-5-2 . For running costs, see what AI automation costs a small business https://hamzaahmadaslam.com/blog/ai-automation-cost-small-business . If you want a model like Carbon connected to WordPress, WooCommerce or your own systems with those limits in code, that is the work behind AI integration for business workflows https://hamzaahmadaslam.com/services/ai-integration .