Google releases Gemini 4 Carbon two weeks after Argon: 59 and first place on the Artificial Analysis Intelligence Index v5, state-of-the-art cybersecurity and long-horizon coding, with early access for cyber defenders in the Fairwind program

Background and the problem

Google shipped Gemini 4 Argon in mid-October and followed it with Gemini 4 Carbon two weeks later; Sundar Pichai wrote on X that he had said Google would release models periodically. The focus is not chat but two of the hardest jobs: long-horizon engineering tasks that take days and hit dead ends, and the full cyber-defense chain from finding a vulnerability to validating it and patching the root cause. Because its offensive-security capability is strong, Google is not opening an API yet and is giving trusted defenders access first.

Architecture and how it works

• Google has not disclosed parameters or architecture; the blog says Carbon reuses Argon's multi-layered safety stack of model training, real-time misuse defenses, prompt-injection resilience and chain-of-thought monitoring.
• Long-horizon work: Google says internal teams saw Carbon run for days, persisting through dead ends to root-cause and fix bugs across large codebases and low-level systems.
• Alignment training targeted not taking shortcuts under pressure: when a task is ambiguous it retrieves sources and surfaces discrepancies before acting, and when evidence is missing it writes caveats into the deliverable or lists what is still required instead of fabricating.
• Google reports 26% less overreach and 29% less under-service, so multi-step plans run end to end within project constraints more reliably.

Benchmarks

All figures are Google's, with cybersecurity evaluations run without guardrails: 88.2% on SEC-bench Pro for finding real vulnerabilities, 54.8% on ExploitGym for exploitation and 71.0% on Collinear's CWE-bench for turning findings into working fixes, all claimed as leading scores. It scores 59 and ranks first on the Artificial Analysis Intelligence Index v5.0, and gains 89 points to fourth place on WebDev Arena at 1769. On safety, successful cheating on Terminal-Bench 4.0 is 0.0% with a 0.4% attempt rate, and the attack success rate on Gray Swan's indirect prompt injection benchmark is 0.0%, down from Argon's 0.7%. Internally, it more than doubles Argon's success rate at generating secure code, roughly halves unsupported claims and cut high-severity misalignment alerts by almost 80%; the Chrome team says it did more than 10x better at high-quality patches than another, unnamed lab's large frontier model.

Getting started

1. For now only a subset of cyber defenders in the Fairwind program have access; Google says it will widen this to more trusted partners and then bring Carbon to developers, businesses and consumers, with no date given.
2. Pricing is set: the introductory rate matches Argon at $2 per million input tokens and $10 per million output tokens, with cached input 95% off the input price, then $4 and $20 after the introductory period. Use that to estimate the cost of moving from Argon.
3. While you wait, keep building long-horizon agents and security-scanning pipelines on Argon, then swap the model name and compare once Carbon is available.
4. Details are in Google's official blog post.