Google unveiled Gemini 4 Carbon on Wednesday, October 28, two weeks after Gemini 4 Argon reached customers, and it framed the new frontier model squarely around cybersecurity and long-running engineering work. Reuters reported that Google claims state-of-the-art performance on vulnerability discovery, exploitation and software patching, and that Carbon generates secure code at more than twice the success rate of Argon on Google's own internal benchmarks. Google DeepMind executive Koray Kavukcuoglu called it the company's strongest model, with state-of-the-art cybersecurity and long horizon capabilities.
On the numbers Google published, the jump is large. Artificial Analysis lists Carbon at 59 on version 5.0 of its Intelligence Index, one point ahead of Claude Opus 5.5 and two ahead of GPT-6 Astra, while Argon scored 50 on the same index. Google reported 88.2 percent on SEC-bench Pro for vulnerability detection, 71.0 percent on CWE-bench for root-cause patching and 54.8 percent on ExploitGym. Treat those as claims to test, not settled facts: several reports note the cyber scores were evaluated without guardrails, none of the headline numbers have been reproduced outside Google, and the Chrome team's claim of more than ten times better patch quality does not name the rival model it beat.
Almost nobody can test them yet. Carbon is rolling out first to a subset of cybersecurity defenders through Google's Fairwind Program while Google continues testing safeguards, and Reuters gave no date for developer, business or consumer access. The price is set anyway: 2 dollars per million input tokens and 10 dollars per million output tokens at introductory rates matching Argon, with cached input heavily discounted and secondary reporting putting the later step-up at 4 and 20 dollars. A 1 million token context window is listed by Artificial Analysis.
The contrast on the same day was instructive. While Google gated its strongest coder behind a defender program, Anthropic launched reporting tools to show enterprise customers what their token spending actually bought, with team-level views of work produced, cost and where workflows broke down, pooled group budgets, and a promise not to use that data for training by default. TechCrunch cited Jellyfish data from 276,000 customers showing spending on agentic coding tools up ten times since the start of the year, to about 2 percent of total engineering cost. The frontier fight is splitting in two: who can find and fix flaws, and who can prove the bill was worth it.
For builders, the practical move is patience plus instrumentation. Do not plan a migration around a model you cannot call, and do not swap a working coding stack on an index lead of one point. When Carbon does open, judge it the way CWE-bench does: a fix only counts if it blocks the exploit and keeps the software working, tested on your own codebase and your own failure cases. In the meantime, the lesson you can use today is Anthropic's side of the day's news. Instrument agent spend by team and task now, because the labs are about to sell you both the capability and the accounting for it, and only one of those is available this week.
