Two weeks after launching Gemini 4 Argon, Google already has a successor: Gemini 4 Carbon, which the company says tops version 5.0 of the Artificial Analysis Intelligence Index with a score of 59. The striking part is who gets it first. There is no open API and no date for developers: starting the week of October 28, 2026, only a select group of cyber defense teams gets access.
What Google announced
Google calls Carbon its "strongest model", built around two strengths: cybersecurity and long-horizon work. According to the announcement, internal teams have kept it running for days on problems such as root-causing and patching subtle bugs across large codebases and low-level systems.
The numbers Google published, measured without product guardrails:
| Benchmark | What it measures | Gemini 4 Carbon |
|---|---|---|
| Artificial Analysis Intelligence Index v5.0 | Overall index | 59 (#1) |
| SEC-bench Pro | Finding real vulnerabilities | 88.2% |
| ExploitGym | Checking whether they are exploitable | 54.8% |
| CWE-bench (Collinear) | Patching the root cause | 71.0% |
| WebDev Arena | Full-stack web development | 1768 Elo (#4, +89) |
| Gray Swan IPI | Indirect prompt injection success rate | 0.0% (Argon: 0.7%) |
Google also shares internal numbers against Argon: more than twice the success rate at generating secure code, half as many unsupported claims, and almost 80% fewer high-severity misalignment alerts in live coding sessions. On Terminal-Bench 4.0, cheating attempts drop to 0.4% and none of them succeed. The Chrome team reports that Carbon's patches were more than ten times better than those of "another lab's large frontier model", without naming it.
How it stacks up against Argon and Claude
The Artificial Analysis index has just changed version, so older scores are not comparable. For the first time, v5 relies on private test sets for 60% of the index to curb benchmark-targeted training, and the gap between open models and the top of the table widens from 3 to 12 points: Kimi K3.1 Pro falls from fourth to ninth place, behind eight closed models.
Where the numbers do line up is pricing per million tokens:
| Model | Input | Output |
|---|---|---|
| Gemini 4 Carbon (launch price) | $2 | $10 |
| Gemini 4 Carbon (afterwards) | $4 | $20 |
| Gemini 4 Argon (launch price) | $2 | $10 |
| Claude Sonnet 5.5 | $2 | $10 |
| Claude Opus 5.5 | $4 | $20 |
| Claude Fable 5.1 | $10 | $50 |
Cached input on Carbon is 95% off the input price. At launch Google matches the price of Argon and Claude Sonnet 5.5, and once the promotion ends Carbon will cost exactly the same as Claude Opus 5.5. An agent burning 100 million input tokens and 10 million output tokens a month would pay $300 at the launch price and $600 afterwards, the same as with Opus 5.5. Among the closed labs the fight is no longer about price; it is about which model gets more done for the same money.
What it means
Carbon confirms a pattern: labs no longer ship their best model to everyone on day one. Google is routing it through Fairwind, its program for high-priority defenders (governments, critical infrastructure operators, healthcare and telecom providers), which already counts more than 650 partners. Getting in requires an organizational background check, phishing-resistant MFA, and restricting access to security, incident response or penetration testing teams, with usage tracked per employee.
The rationale is defender's advantage: a model that finds, validates and patches vulnerabilities is just as useful to an attacker, so Google wants defenders to get there first. For anyone building with AI, the practical takeaway is different: the model at the top of the leaderboard may stay out of reach for weeks, and what you can actually buy through the API today is Argon.
What we still don't know
The announcement gives no model ID, context window, output limit or date for the API, AI Studio or the Gemini app, and it does not say when the launch price ends. And while the security numbers come from external benchmarks, Google ran them itself: independent breakdowns from Artificial Analysis and other evaluators are still to come.
Sources
- Gemini 4 Carbon: our next frontier model, Google, October 28, 2026.
- Fairwind Program, Google DeepMind.
- New Artificial Analysis index widens the gap between open and closed models, The Decoder, October 28, 2026.
- Models overview, Claude documentation.