Gemini 4 Carbon Is No. 1, So Why Can't Developers Use It Yet?

Google's Gemini 4 Carbon tops the Artificial Analysis index and three security benchmarks, but only vetted cyber defenders get it first. What devs should know.

TheTechLane cover on an ice-white background with the TL logo, a violet and coral carbon hexagon lattice with a coral padlock cell, and the headline "Top of the Charts. Locked to Defenders."

Google has announced Gemini 4 Carbon, its strongest model yet, and it now sits at No. 1 on the Artificial Analysis Intelligence Index. There's a catch: almost nobody can use it. Starting this week, only a small group of vetted cyber defenders get access, while developers wait for Google to finish tuning its guardrails.

Here's what you need to know:

  • Carbon scores 59 on the Artificial Analysis Intelligence Index v5.0, the top spot, ahead of Claude Opus 5.5 and GPT-6 Astra.

  • It leads three security benchmarks: 88.2% on SEC-bench Pro (finding bugs), 54.8% on ExploitGym (exploiting them) and 71.0% on CWE-bench (patching them).

  • First access goes to a subset of cyber defenders through Google's Fairwind Program, which already has more than 650 partners.

  • Introductory API pricing is $2 per million input tokens and $10 per million output tokens, rising to $4 and $20 later.

  • Google hasn't given a date for developer access or for when the introductory price ends.

UML sequence diagram of the Gemini 4 Carbon rollout: Google DeepMind puts Carbon through the US government's voluntary pre-release testing, gives a subset of cyber defenders access this week through the Fairwind Program under background checks, phishing-resistant MFA and no-resale rules, collects guardrail feedback, widens access in phases, and promises developers and the Gemini API access as soon as possible at $2 input and $10 output per million tokens, later $4 and $20

Defenders first, developers later: how Gemini 4 Carbon is rolling out

What Happened

On Wednesday, October 28, Google DeepMind SVP and Chief AI Architect Koray Kavukcuoglu introduced the model in Google's official Gemini 4 Carbon announcement. It lands just two weeks after Gemini 4 Argon, and Google calls it its "strongest model with state-of-the-art cybersecurity and long horizon capabilities."

The pitch covers the whole defence lifecycle. Google says Carbon finds real vulnerabilities, checks which ones can actually be exploited, and writes fixes at the root cause. The company adds that Carbon has more than 2x the success rate of Argon at generating secure code on internal benchmarks, and that the Chrome team saw it do more than 10x better than "another lab's large frontier model" at writing high-quality patches. Google notes the external benchmark scores were measured without guardrails.

It's also built for long jobs. Google says internal teams have watched Carbon "run for days" on multi-step engineering work like root-causing subtle bugs across large codebases. On WebDev Arena it jumped 89 Elo points to No. 4, at 1768.

Safety numbers got a lot of attention. Google says Carbon cut successful cheating on Terminal-Bench 4.0 to 0.0%, reduced unsupported claims by about half compared with Argon, and resisted every attack in Gray Swan's indirect prompt injection benchmark. Ars Technica's report puts that 0% next to Argon's 0.7%, Opus 5.5's 1% and GPT-6 Astra's 8.5%.

Then comes the rollout. Google says it's taking part in the US government's voluntary pre-release testing process and will widen access in phases through the Fairwind Program. According to the Fairwind Program page, partners pass background checks, must use phishing-resistant MFA, can only give access to internal security, incident response or pen-testing teams, and can't share or resell it.

Why It Matters

This is now a pattern. The Verge's coverage notes that Carbon is only available to a select group of cybersecurity defence companies, and Argon followed the same staged path. The most capable models are arriving at defenders first and at the public later.

That follows a busy month for AI and offensive security. Just this week, Anthropic described how an unreleased model chained real bugs into working exploits for Chrome, Safari and Linux. When models can find, validate and patch bugs on their own, labs clearly see the same skills as a risk if anyone can rent them.

Ars Technica also adds some needed scepticism. It points out that Carbon beats Argon by "slim margins" in every test Google ran, that every frontier model scores within a few points of the others, and that, for all anyone knows, Carbon "could be a post-training upgrade of Argon."

What It Means for Developers

You can't call Carbon today unless your organisation is a Fairwind partner. But you can plan for it:

  • Budget for both prices. Plan around $2/$10 per million tokens now, but model your costs at $4/$20 too. Ars reports Google wouldn't say when the discount ends.

  • Lean on caching. Google says cached input tokens are priced 95% below the normal input rate, which matters for long agent runs over big codebases.

  • Treat secure-code claims as a hypothesis. "2x more secure code" comes from Google's internal benchmarks. Run your own evals on your own repos before you trust it in CI.

  • Use what's open now. The Fairwind page says teams that aren't eligible can still use Google's CodeMender security agent with publicly available models.

  • Compare on price, not just rank. If you're choosing a model for coding agents, check how Carbon stacks up against cheaper options like OpenAI's GPT-6.1 Terra once it's actually available to you.

Data card of Gemini 4 Carbon numbers: Google-reported benchmarks without guardrails of 88.2% on SEC-bench Pro, 71.0% on CWE-bench and 54.8% on ExploitGym, 0.0% indirect prompt injection attack success versus Argon's 0.7%, a score of 59 at No. 1 on the Artificial Analysis Intelligence Index v5.0, and API pricing of $2 input and $10 output per million tokens at launch rising to $4 and $20, with cached input 95% off

The Carbon numbers to plan around: benchmarks and API pricing

What's Next

Ars expects Carbon to show up first for Google AI Ultra subscribers and in the Gemini API, possibly within a couple of weeks if it follows Argon. Google itself only promises developer, business and consumer access "as soon as possible."

Open questions remain. Google hasn't said how many organisations are in the first cohort, what guardrails it's still changing, or when the introductory price expires. Argon's discounted rates run through the end of November, according to Ars, but Carbon has no end date yet.

FAQ

Can I use Gemini 4 Carbon in the Gemini API today?

Not unless you're an approved Fairwind partner. Google says broader access for developers, businesses and consumers will come once it's done testing its guardrails, without giving a date.

How much will Gemini 4 Carbon cost?

The introductory price is $2 per million input tokens and $10 per million output tokens, the same as Argon. After that, Google says it rises to $4 and $20.

Is Carbon really better than Claude Opus 5.5 and GPT-6 Astra?

On Google's charts and the Artificial Analysis index, yes, but by a few points. Ars Technica notes that frontier models all score close together, so test on your own workloads.

Sources

Follow TheTechLane

Keep learning AI, development, and modern developer technologies with TheTechLane.

Learn • Build • Share

Follow @thetechlanehq for more developer guides, tutorials, and practical technology content.