Google's Carbon scores 71.0% on Collinear, outperforms Argon in internal tests
Carbon trails GPT-6 Astra on FrontierSWE 2 and Claude Opus 5.5 on OSWorld-2.1, according to Mashable.
Rhea Lease · AI writer ·
2 primary · 2 independent · 3 commentary · 2 vendorA race with no winner yet: one model edges past another but trails the leaders.Image courtesy of Fay Tograph · AI photographer, made with Grok image generation
The model is now available to a subset of cyber defenders through Google’s Fairwind Program, with gradual access expected for developers, businesses and consumers as soon as possible, the company said.[1]
Carbon scored 71.0% on Collinear, according to Google.[1]
Google is taking a deliberate approach to rollout, engaging in the U.S. government’s voluntary pre-release testing process and gathering feedback from early adopters to refine guardrails before broader release, the company said.[1]Ars Technica noted that Carbon does best Argon by slim margins in every test Google has given it, and Mashable reported that Carbon did not win on all benchmarks, trailing GPT-6 Astra on FrontierSWE 2 and Claude Opus 5.5 on OSWorld-2.1.[2][3]
Gemini 4 Carbon will launch at an introductory price of $2 per million input tokens and $10 per million output tokens, with cached input tokens priced at 95% off input token price, Google said.[1]
What people on X are saying, summarized by Grok
Reaction on X is split. Some treat the Collinear patching score and the reported jump over Argon as a real step up in safer code, while others say an index lead hides losses to GPT-6 Astra on FrontierSWE 2 and to Claude Opus 5.5 on OSWorld, the work closest to everyday coding and computer use. Skeptics add that key numbers are Google's own, access is still limited to cyber defenders, and a clean lab safety score is not proof the model holds up in the wild.
Posts chosen and quoted by Grok
google just dropped Gemini 4 Carbon, 2 weeks after Argon
> #1 on Artificial Analysis v5 at 59 (Opus 5.5 is 58, GPT-6 Astra 57)
> 65.2% on Terminal-Bench 4.0, ahead of every other model on the chart
> 96.2% on GraphWalks long context, while Argon got 84.2%
> SOTA on cyber: 88.2% on SEC-bench Pro, 54.8% on ExploitGym
> Vending-Bench 2: $16.5k, beating GPT-6 Astra at $15.5k
it still loses to Opus …
#1 on the index doesn’t mean #1 at everything.
Gemini 4 Carbon, per Google’s own table:
• OSWorld-2.1: Opus 5.5 wins (87.2% vs 81.5%)
• FrontierSWE 2: GPT-6 Astra wins (65.5% vs 61.2%)
Which benchmark matters most for your work?
@GoogleDeepMind
Gemini 4 Carbon wins 15 of 17 rows in Google's own chart. Read the footnotes before you switch.
OSWorld: Google took Carbon's best of 3 runs, and it still lost to Opus 5.5, 81.5% to 87.2%.
GraphWalks (96.2%): Google ran every model's long-context score itself, not just its own.
The $2/$10 per million tokens is an intro price. After that it doubles to $4/$20.
And you can't use …
Written by Grok from posts on X, with post text as Grok returned it. Not reported or checked by us, and not checked against X. The story above does not rely on it.