Google's Gemini 4 Carbon edges past Claude Opus 5.5 on one ranking, 59 to 58

A one-point lead, with details

Google DeepMind has released Gemini 4 Carbon, its newest flagship model. It arrives about two weeks after Gemini 4 Argon, so the pace of releases is itself part of the story. Our newsletter flagged the headline claim as unverified because the digest's title gave no scores. Ifanr's full briefing does supply them.

On the Intelligence Index v5.0 from the independent testing firm Artificial Analysis, Carbon (High) scored 59. Claude Opus 5.5 (Max) scored 58. The labels in brackets are the reasoning settings each model was run at, so the two entries are not necessarily tested under identical conditions.

One point is a thin margin. The index also changed this week, so the number is worth reading carefully, as the next section explains.

Why the ranking needs care

Artificial Analysis published version 5 of its index on 28 October. By its own description, 60% of the weight now comes from private test sets or held-back answers. That makes it harder for a model to look good by having seen the questions in advance.

The update added two new tests. FrontierCode 2.0 has 132 tasks drawn from real open-source projects, and a change passes only if a project maintainer would be willing to merge it. Terminal-Bench-Science has 70 research tasks done in a command-line window, and every test inside a task must pass.

When the new index launched, Opus 5.5 scored 58, GPT-6 Astra scored 57 and GPT-6.1 Sol scored 55, all at maximum reasoning. Carbon then took first place with 59. Artificial Analysis itself warns that scores from the old and new versions should not be compared, because the questions changed.

What Carbon is built to do

Google says this upgrade centres on long-running programming work and cybersecurity. Internal teams reportedly kept Carbon running for several days to hunt down and fix hidden bugs in large codebases and low-level systems. When it hit a dead end, it was able to try other solutions instead of stopping.

Google also published three security results. Carbon scored 88.2% on SEC-bench Pro, which tests finding real software vulnerabilities. It scored 54.8% on ExploitGym, which checks whether a flaw can actually be exploited. It scored 71.0% on CWE-bench, which tests writing a fix for the root cause. These came from Google, and they were measured with safety restrictions switched off.

That last detail helps explain the cautious rollout. A model that is good at finding and exploiting flaws is useful to defenders and risky in other hands.

Limited access and launch pricing

Carbon is opening this week to a select group of cybersecurity defence organisations. It will then expand to more trusted partners through something called the Fairwind Program. Google is also taking part in voluntary pre-release testing run by the US government. No date has been given for developers, companies or ordinary consumers.

Pricing is the same as for Argon. Developers pay $2 per million input tokens and $10 per million output tokens. Tokens are the small chunks of text a model reads and writes. After the introductory period, the rates become $4 and $20.

OpenAI and Google's chips

As Ifanr reported, citing The Information, OpenAI plans to buy Google's TPUs. These are custom chips Google designed for AI work, a rival to the Nvidia processors most of the industry uses. OpenAI has also posted a job for an engineer to lead the work of making its frontier models run well on TPUs. The report says this effort is at an early stage, and that experience in training or in running finished models would both qualify.

The two companies already do business. OpenAI rents Nvidia chips through Google Cloud, and last week said it would offer its models there too. Oracle has said it will rent TPUs to customers, and The Information says Oracle will buy the chips. OpenAI said last year it had no plans to rely on TPUs alone, which is a narrower statement than saying it would never use them.

Alphabet's latest results point the same way. Google Cloud revenue grew 103%, and the company credited demand for enterprise AI infrastructure and TPU system sales. It also said the Gemini app has reached 1 billion users.