Gemini 4 Carbon

Gemini 4 Carbon is the new #1 on Artificial Analysis's intelligence index, two weeks after Google shipped Gemini 4 Argon, and almost nobody can use it yet. Google DeepMind announced it on October 28, 2026, and the first people to get it are a small group of cyber defenders.

Carbon is Google's new frontier model, pitched as its strongest yet with "state-of-the-art cybersecurity and long horizon capabilities." It builds on Gemini 4 Argon, which went out on September 30 and reached developers in the Gemini API two weeks ago. The rollout copies Argon's: starting this week a subset of cyber defenders get it, more trusted partners follow through Google's Fairwind Program, and developers, businesses and consumers get it "as soon as possible," with no date. The price is set anyway. It launches at Argon's introductory $2 per million input tokens and $10 per million output, cached input is 95% off, and after an unspecified intro period it goes to $4 / $20, which is Claude Opus 5.5's rate card. Artificial Analysis lists a 1M-token context window.

The evidence is a 17-row comparison table against Argon, GPT-6 Astra, Claude Opus 5.5 and Fable 5.1, where Carbon comes first on 15 rows (see ## The benchmarks), plus three cyber charts and a list of internal anecdotes. Google says its teams have watched Carbon run "for days" on multi-step engineering work like root-causing subtle bugs across large codebases, that it writes secure code more than twice as often as Argon on internal tests, and that the Chrome team found it "more than 10X better than another lab's large frontier model" at writing patches. The catches sit in the same materials. Carbon trails GPT-6 Astra and Opus 5.5 on FrontierSWE and Opus 5.5 on computer use, its headline exploit score ran with a 10-hour cap while Opus 5.5's ran with 6, and Artificial Analysis, which Google quotes for the #1 claim, also found Carbon making up answers more often than Argon (11% against 4%). Google is also taking part in the U.S. government's voluntary pre-release testing, and the announcement links Argon's safety report and a five-page evaluation methodology PDF, not a Carbon system card.

What's new

Carbon is a fast follow-up to Argon at the same price, and most of what Google changed is about how the model behaves on long jobs, not only how it scores.

• Trained not to cheat. Google says Carbon's successful cheating on Terminal-Bench 4.0 fell to 0.0%, with attempts down to 0.4%, and that high-severity misalignment alerts in live coding with its own teams dropped by almost 80%.

• Says what it doesn't know (Google's version). On internal tests Carbon cut unsupported claims by about half against Argon, overreach by 26% and under-delivering by 29%, and Google says it flags discrepancies to the user instead of guessing. Artificial Analysis measured the opposite on its knowledge test, where Carbon answers more questions right but hallucinates more often than Argon, so this one needs outside testing.

• Built for cyber defense first. It leads Google's three external cyber charts: finding vulnerabilities (SEC-bench Pro), exploiting them to check which matter (ExploitGym) and patching the root cause (CWE-bench). Wiz, Google's own cloud security company, says it also leads on web penetration testing.

• No successful prompt injections. On Gray Swan's indirect prompt injection test, where hidden instructions in a web page or file try to hijack the model, every attack failed. Argon's 0.7% was already the lowest on the chart.

The benchmarks

Google ran Gemini 4 Carbon and Gemini 4 Argon at the highest thinking setting, pass@1, through the Gemini API, and took GPT-6 Astra, Claude Opus 5.5 and Fable 5.1 at max from the labs' own reports and public leaderboards. A few rows need reading closely: the Terminal-Bench rows are Artificial Analysis's runs, OSWorld is the 82-task offline subset with Gemini's score maxed over three runs, SWE-sweep, Agents' Last Exam, ExploitGym and SEC-bench Pro are Google's own runs for Gemini, and ExploitGym compares Carbon with a 10-hour cap per task against Opus 5.5 with 6 hours (Carbon scores 53.5% at 6 hours). SEC-bench Pro and Gray Swan add rivals from Google's charts.

Google's numbers

• Vals Index v2.2 (real work, weighted by industry): 70.4%. +0.2 over Claude Opus 5.5 (70.2%), +1.2 over Gemini 4 Argon (69.2%), +5.4 over GPT-6 Astra (65.0%).

• SWE-sweep (changes across a big codebase): 9.6%. +0.1 over GPT-6 Astra (9.5%), +5.3 over Gemini 4 Argon (4.3%), +6.0 over Claude Opus 5.5 (3.6%), +6.7 over Fable 5.1 (2.9%).

• FrontierSWE v2 (frontier software engineering): 61.2%. −4.3 under GPT-6 Astra (65.5%), −1.1 under Claude Opus 5.5 (62.3%), +4.9 over Fable 5.1 (56.3%), +6.2 over Gemini 4 Argon (55.0%).

• Terminal-Bench 4.0 (Artificial Analysis harness): 65.2%. +5.6 over Claude Opus 5.5 (59.6%), +6.1 over GPT-6 Astra (59.1%), +8.1 over Gemini 4 Argon (57.1%), +13.2 over Fable 5.1 (52.0%).

• Agents' Last Exam v1.1 (professional tasks, rubric score): 77.4%. +4.3 over GPT-6 Astra (73.1%), +5.8 over Claude Opus 5.5 (71.6%), +8.4 over Gemini 4 Argon (69.0%).

• OSWorld 2.1 (offline subset): 81.5%. −5.7 under Claude Opus 5.5 (87.2%), +0.6 over GPT-6 Astra (80.9%), +4.7 over Fable 5.1 (76.8%), +6.8 over Gemini 4 Argon (74.7%).

Independent numbers

Artificial Analysis tested Gemini 4 Carbon and Gemini 4 Argon at high, Claude Opus 5.5, Claude Sonnet 5.5 and Fable 5.1 at max (with fallback), and GPT-6 Astra and GPT-6.1 Sol at max. Its index moved to v5.0 days before the launch, so these scores don't line up with the older v4 numbers on earlier drops. On LMArena, Carbon debuted #1 in Text Arena and #4 in WebDev, where Google puts it at 1769, up 89 points.

• Artificial Analysis Intelligence Index v5.0: 59. +1 over Claude Opus 5.5 (58), +2 over GPT-6 Astra (57), +4 over GPT-6.1 Sol (55), +5 over Claude Sonnet 5.5 (54), +9 over Gemini 4 Argon (50), +10 over Fable 5.1 (49).

• Artificial Analysis Coding Agent Index (each model in its own coding app): 59. −3 under Claude Opus 5.5 (62), +6 over Gemini 4 Argon (53).

• AA-Omniscience (right answers minus made-up ones): 54. +6 over Claude Opus 5.5 (48).

See how it ranks. Every score above is on the Handy AI benchmarks board, where you can see where Gemini 4 Carbon lands against every model we track, open and closed.

First impressions

The model is a couple of hours old and nobody outside Google's partners has used it, so these are reactions to the numbers, plus a few evaluators who got early access. The Hacker News launch thread passed 100 points and 50 comments inside half an hour.

The positives

Artificial Analysis (@ArtificialAnlys), which runs its own tests on every major model, confirmed Google's headline within half an hour.

Google's #1 claim checks out on the evaluator's own page, by a point over Opus 5.5. The same thread has Carbon second to Opus 5.5 in Claude Code on coding agents, so "best at everything" isn't the claim.

Andon Labs (@andonlabs), which runs Vending-Bench, a test where a model runs a simulated vending machine business, had Google on top for the first time in a year.

A $16,484 ending balance against $9,235 for Opus 5.5 is the biggest gap in Google's table that someone outside Google measured.

Peter Gostev (@petergostev), who posts his own model comparisons, said what a lot of Argon users were thinking.

That matches the benchmark jumps from Argon to Carbon, and it's the kindest possible read of shipping two frontier models in a month.

The negatives

Andon Labs (@andonlabs) followed its own praise with the worst behavior note of the day.

Google's post leans hard on alignment gains. "Less severely" than Argon, and still worse than Astra, is the outside check on that.

kilroy123 on Hacker News had the most-repeated complaint in the thread.

I'm not a big fan of their trend of talking about a new model and not releasing it for a while. (comment)

Argon took two weeks to reach the API after its announcement. Carbon gives no date at all, and developers spent the summer waiting on a Gemini 3.5 Pro that never shipped.

zcw (@zwb44), a general X user, carried Argon's baggage into Carbon.

Bloomberg reported at Argon's launch that Googlers found it weaker at real coding than its scores suggested, and Carbon's own table still has it behind Astra and Opus 5.5 on FrontierSWE. People will believe the coding story when they can run it.

Jake's take

On paper (big caveat, with Google), this is the first Gemini in a long time that beats Opus 5.5 on most tests and at Opus pricing. First on Artificial Analysis, first on 15 of Google's 17 rows, and a Vending-Bench run that ends with almost twice Opus's money. Two weeks ago I said Argon was "great benchmarks, zero tokens." Carbon appears to fix the first half.

But, like with all Google release, it’s weird. It's still a press release for most people. Nobody outside a vetted cyber program can use it, there's no date, and the table has the usual tells. The exploit headline gives Carbon 10 hours a task against Opus's 6, several rows are Google's own runs, and the alignment section brags about halving made-up claims on the same day Artificial Analysis clocks its hallucination rate going from 4% to 11%.

And, the price doubles after the intro period. At $4 / $20 Artificial Analysis already has it at $7.76 a task against $1.27 for GPT-6.1 Sol, four points behind.

The anti-cheating work is good, though. But not good enough to get me off Opus 5.5 (though I’m growing more and more skeptical that anything will anytime soon).