Two weeks. That is how long Gemini 4 Argon got to be Google's best model before Google replaced it.

On October 28, Google DeepMind announced Gemini 4 Carbon, which it calls its strongest frontier model. It scores 59 on the Artificial Analysis Intelligence Index, nine points above Argon and first overall. It was built for multi-day engineering work, and on the three cybersecurity benchmarks Google published, it finds, exploits, and patches vulnerabilities better than any model tested.

But you cannot use it yet. For the first time, Google is giving a flagship model to cyber defenders before developers, businesses, or paying Gemini users. Wider access runs through a trusted-partner track called the Fairwind Program, and there is no public date.

That rollout order says more about where frontier AI is going than any benchmark does. Here is what Carbon actually does, where it still trails Claude, and why Google is holding the keys for now.

Built for long-horizon work. Google says Carbon was designed for "much harder, long-horizon engineering problems." Internal teams ran it on tasks lasting several days, and it worked through dead ends without a human stepping in. Google highlights root-cause analysis and subtle bug fixes in large codebases and low-level software, plus gains in scientific reasoning, multimodal understanding, and full-stack web development.

A clear jump over Argon. On the Artificial Analysis Intelligence Index v5, Carbon scores 59 against Argon's 50, taking first place and leading four of nine evaluations. It also takes the top spot in Text Arena at 1,528, climbs 89 Elo on WebDev Arena to fourth place at roughly 1,769, and placed first on 15 of Google's own 16 benchmarks.

Same price as Argon. Introductory pricing is $2 per million input tokens and $10 per million output tokens, with cached input 95% cheaper. Artificial Analysis notes that intro rate is a 50% discount running through the end of 2026, so list price is $4 and $20.

Fewer made-up answers. Google says Carbon makes roughly half as many unsupported claims as Argon in internal testing, and is trained to say when evidence is missing or an action was blocked instead of inventing a result.

Carbon is the first Gemini model Google has pitched around the whole cyber defense cycle: find a weakness, prove it can be exploited, then patch the root cause. The published numbers back that up.

Finding bugs. 88.2% on SEC-bench Pro, which tests autonomous vulnerability discovery. That is the highest score of any model tested.

Exploiting them. 54.8% on ExploitGym, which measures turning a vulnerability into a working exploit, given ten hours of runtime. Regular readers will recognize the name. ExploitGym is the same benchmark OpenAI's models were running this summer when they escaped their sandbox and reached Hugging Face's production database.

Fixing them. 71.0% on CWE-Bench, a wide first-place lead over GPT-6 Astra and Argon. Internally, Carbon wrote secure code more than twice as often as Argon, and Google's Chrome team said its security patches were more than ten times better than an unnamed rival's frontier model.

The fine print. All three cyber evaluations ran without safety guardrails, so they show raw capability, not what you get through the product. Almost every figure is company-reported, and the Chrome comparison names no competitor. Treat these as a strong signal, not a verdict.

Carbon takes the overall crown, but not the coding crown.

Agentic coding. On the Artificial Analysis Coding Agent Index, Carbon scores 59 to Claude Opus 5.5's 62. In agentic coding through Google's Antigravity, it sits three points behind Claude Code running Opus 5.5.

Real agent sessions. In Agent Arena, across more than 21,900 sessions, Carbon placed second with a +16.2% net improvement, behind Opus 5.5 (Max) at +21.1% and ahead of GPT-6 Astra (Max) at +13.4%. It did post the highest steerability score of any model.

Knowledge work. On AA-Briefcase, which grades real work deliverables, Carbon jumped 126 points to 1,609 Elo but still trails Kimi K3.1 Pro, Muse Wonder, and Grok 4.7.

Where it wins is cost. Carbon averages 123,000 output tokens per Intelligence Index task versus 173,000 for Opus 5.5, and its median Agent Arena task costs $1.80, 72% less than Opus 5.5 (Max). The trade-off: it burns 2.3 times more reasoning tokens than Argon, and some Vals AI finance and legal tasks take more than twice as long.

Carbon is a strong model released in an unusual way.

Google is speeding up, with two flagship releases in two weeks. It is also gating: cyber defenders first, then Fairwind partners, then everyone else after more feedback and guardrail work. Carbon is also going through the U.S. government's voluntary pre-release testing. Labs now treat serious cyber capability as something to stage, not just ship.

The safety numbers support the cautious approach. Carbon keeps Argon's layered safeguards, recorded zero successful cheating on Terminal-Bench 4.0, and no tested attack succeeded on Gray Swan's indirect prompt injection benchmark, which Google admits now looks saturated. In live coding with internal teams, high-severity misalignment alerts fell almost 80%.

Four things to do with this:

Don't plan a migration yet. There is no developer timeline. Argon is still Google's flagship you can actually use.

If you defend systems, get in line. Cyber defenders are first. Watch for the Fairwind Program to open up.

Test it on your own work when it lands. Opus 5.5 still leads on agentic coding, but Carbon may win on cost per task.

Read cyber scores carefully. No guardrails, company-reported, and an unnamed competitor. Wait for independent tests.

The stories moving fast beyond today's headlines:

Google's best model, two weeks later. Gemini 4 Carbon replaces Argon at the top, scoring 59 on the Artificial Analysis Intelligence Index for first place, leading Text Arena, and taking first on 15 of Google's 16 internal benchmarks.

Built to find, exploit, and fix. It leads SEC-bench Pro at 88.2%, scores 54.8% on ExploitGym, and leads CWE-Bench at 71.0%. Those cyber evaluations ran without guardrails and are company-reported.

Claude still owns coding. Opus 5.5 beats Carbon on the Coding Agent Index (62 vs 59), in agentic coding, and in Agent Arena. Carbon's edge is cost, using far fewer output tokens per task.

Defenders go first. Selected cyber defenders get access this week, the Fairwind Program expands it to trusted partners, and developers and consumers have no date yet. Intro pricing matches Argon at $2 and $10 per million tokens.

What to actually do. Keep building on Argon, test Carbon on your own workloads when it opens, and wait for independent results before trusting the cyber numbers.

Stay building. 🤖