02 of 05 · Labs & Models

Gemini 4 Carbon Tops a Big Leaderboard, but Most of Us Can’t Use It Yet

By Willa Ravenwire · · 2 min

WIRE

Why it matters. A strong benchmark story on a model you can't use yet.

Google has announced Gemini 4 Carbon, its new top model, just two weeks after Gemini 4 Argon. The focus is long coding jobs and cybersecurity defense.

What’s confirmed. Access is narrow. Google says Carbon goes first to a subset of cyber defenders this week, then to more trusted partners through its Fairwind Program. Developers, businesses and regular users will get it later, with no date given. Google lists the same launch API price as Argon: $2 for each million input tokens and $10 for each million output tokens. 9to5Google reports the price rises to $4 and $20 after that period.

What Google claims. Carbon ranks first on version 5.0 of the Artificial Analysis Intelligence Index with a score of 59, per Google. It also reports leading security scores: 88.2% on SEC-bench Pro for finding bugs and 71.0% on CWE-bench for patching them. Google says internal teams watched it work on hard bugs for days.

It also claims Carbon makes about half as many unsupported claims as Argon in internal tests. The Verge notes it beats top OpenAI and Anthropic models on some benchmarks.

Our take. The slow rollout is probably the most honest part of this launch. A model that is good at finding security holes is also useful to attackers, so starting with defenders makes sense. But almost no one outside Google can test the claims yet, and the “runs for days” stories are internal and unverified.

What’s real here: a strong benchmark story on a model you can’t use yet.

Share on X

Sources

By Willa Ravenwire · The AI Bandwagon

Next on today’s wire: 03 of 05OpenAI Reportedly Cuts Its Mega-Round to $20 Billion, With $11 Billion Secured

More in Labs & Models

  1. Labs & Models · 2 min

    Mistral Large 4 “Le Chonk” Goes Open Weight, but Bring a Server Rack

    A genuinely open frontier-scale model, mostly useful to people who own data center hardware.

    Nov 3, 2026

  2. Labs & Models · 5 min

    LRMoE Cuts MoE Training Time 21 to 45 Percent by Asking Less of the Backward Pass

    Training a sparse Mixture-of-Experts model is a memory negotiation. Every extra copy of an overloaded expert buys you balance and costs you the buffer space that…

    Oct 31, 2026

  3. Labs & Models · 6 min

    The Enterprise Agent Paper That Mostly Proves Guardrails

    arXiv:2610.08793v1 Announce Type: new Abstract: Enterprise systems increasingly rely on automated workflows, yet many AI-driven solutions remain brittle under non-stationary condit

    Oct 8, 2026

Leave a Reply

Your email address will not be published. Required fields are marked *