Source reported
Google's Gemini 4 Carbon: 88.2% on SEC-bench Pro, two weeks after Argon
What happened
Google reports a score of 88.2% for Gemini 4 Carbon on SEC-bench Pro, a benchmark for finding software vulnerabilities, with 54.8% on ExploitGym and 71.0% on CWE-bench. The model arrived two weeks after Gemini 4 Argon, the previous flagship. Google says Carbon beats OpenAI's GPT-6 Astra and Anthropic's Claude Opus 5.5 on multiple benchmarks, but every figure so far is Google's own, from internal testing. Carbon opens first to cyber defenders this week, with no date for general availability.
What we know
- Google DeepMind's Koray Kavukcuoglu wrote that the Chrome team saw Carbon perform more than 10 times better than another lab's frontier model at generating high-quality patches, per SiliconANGLE.
- Google lists SEC-bench Pro at 88.2% for finding vulnerabilities, ExploitGym at 54.8% for exploitation and CWE-bench at 71.0% for patching, per 9to5Google.
- Google says Carbon cuts unsupported claims by about 2x versus Argon, and reduced overreach by 26% and under-service by 29%, per 9to5Google.
- Teams using Carbon in live coding environments reported a nearly 80% drop in misalignment alerts, and Carbon attempted to cheat on 0.4% of Terminal-Bench 4.0 questions with no successful attempts, per SiliconANGLE.
- Introductory pricing is $2 per million input tokens and $10 per million output tokens, rising afterwards to $4 and $20, per 9to5Google. Access starts this week through Google's Fairwind Program for organizations such as internet providers, per SiliconANGLE.
What we don't know yet
- Whether independent testers can reproduce the benchmark results. The Artificial Analysis Intelligence Index record was set on a new index version released the same day, with two added benchmarks and saturated questions removed, so it is not directly comparable to earlier scores.
- Which competing model produced the 'more than 10X' patch-quality gap, and under what test conditions. Kavukcuoglu's post refers only to 'another lab's' model.
- When Carbon becomes broadly available. Google says it will refine the guardrails with early-adopter feedback first and gives no date.
- How the offense-oriented results, such as ExploitGym at 54.8%, translate to real-world misuse risk, since the sources report no independent evaluation.
Why it matters
Security teams at organizations in the Fairwind Program will get early use of a model that Google says finds, exploits and patches software flaws. Google's vendor-reported numbers will be tested by independent evaluators before buyers can compare Carbon with GPT-6 Astra and Claude Opus 5.5.
Claims
- ConfirmedGoogle reports Gemini 4 Carbon scores 88.2% on SEC-bench Pro.
- ConfirmedGoogle reports Gemini 4 Carbon scores 54.8% on ExploitGym.
- ConfirmedGoogle reports Gemini 4 Carbon scores 71.0% on CWE-bench.
- ConfirmedGemini 4 Carbon launched two weeks after Gemini 4 Argon.
- ConfirmedIntroductory pricing is $2 per million input tokens and $10 per million output tokens, rising to $4 and $20 afterwards.
- UnconfirmedGoogle says the Chrome team saw Carbon perform more than 10 times better than another lab's frontier model at generating high-quality patches.
- UnconfirmedGoogle says Carbon cuts unsupported claims by about 2x compared with Argon.
- UnconfirmedGoogle says Carbon outperforms GPT-6 Astra and Claude Opus 5.5 across multiple benchmarks.
Related coverage
Why you can trust this story
62%Source map · 2 outlets / 2 articles
- SiliconANGLE
- 9to5Google
- Article 19to5google.comreport
How this credibility score is calculated
- Source reliability
- Corroboration
- Primary source
- Claims checked against the source
- No contradiction
- Settled
- Claim attribution
- AI disclosure
Checked claims were verified against the cited source text. Weights are fixed and explainable.
Human accountability
- 2 independent origins / 2 sources
- Drafted by an automated system; a human editor approves every story before it is published.
Verification ledger · 8
- Claim: Google reports Gemini 4 Carbon scores 88.2% on SEC-bench Pro.Status: ConfirmedVerification trailVerifiedChecked by verification engine
At least one checkable detail was matched against the full text of the cited source.
- Claim: Google reports Gemini 4 Carbon scores 54.8% on ExploitGym.Status: ConfirmedVerification trailVerifiedChecked by verification engine
At least one checkable detail was matched against the full text of the cited source.
- Claim: Google reports Gemini 4 Carbon scores 71.0% on CWE-bench.Status: ConfirmedVerification trailVerifiedChecked by verification engine
At least one checkable detail was matched against the full text of the cited source.
- Claim: Gemini 4 Carbon launched two weeks after Gemini 4 Argon.Status: ConfirmedVerification trailVerifiedChecked by verification engine
At least one checkable detail was matched against the full text of the cited source.
- Claim: Introductory pricing is $2 per million input tokens and $10 per million output tokens, rising to $4 and $20 afterwards.Status: ConfirmedVerification trailVerifiedChecked by verification engine
At least one checkable detail was matched against the full text of the cited source.
- Claim: Google says the Chrome team saw Carbon perform more than 10 times better than another lab's frontier model at generating high-quality patches.Status: UnconfirmedVerification trailUnverifiedChecked by verification engine
No automated check against the source text is available.
- Claim: Google says Carbon cuts unsupported claims by about 2x compared with Argon.Status: UnconfirmedVerification trailUnverifiedChecked by verification engine
No automated check against the source text is available.
- Claim: Google says Carbon outperforms GPT-6 Astra and Claude Opus 5.5 across multiple benchmarks.Status: UnconfirmedVerification trailUnverifiedChecked by verification engine
No automated check against the source text is available.
Update log · 0
AI accelerates. High-risk topics wait for a human editor.