Image to help understand the article
Google limits initial access to cybersecurity specialists
Google unveiled Gemini 4 Carbon on Oct. 28, 2026, initially making the artificial intelligence model available to a group of cybersecurity defenders selected by the company.
Google said it plans to expand access through its Fairwind Program before offering the model more broadly to developers, businesses and consumers. The company said it would refine the model’s safety measures before a wider release.
Google did not announce which subscription plans would include Gemini 4 Carbon, whether it would be offered through its cloud platform or what model ID developers would use to access it through an application programming interface. The company also said its published cybersecurity benchmark results were produced with the model’s guardrails disabled.
Introductory pricing and cached-input discount
Google set introductory pricing at $2 per 1 million input tokens and $10 per 1 million output tokens. Tokens are the units AI companies use to measure the text and other data processed or generated by their models.
Cached inputs receive a 95% discount from the standard input-token price. Google also listed subsequent pricing of $4 per 1 million input tokens and $20 per 1 million output tokens.
Carbon’s coding and agent benchmark results
According to results released by Google DeepMind, Gemini 4 Carbon scored 59 on AA Index v5.0, compared with 57 for GPT-6 Astra, 58 for Claude Opus 5.5 and 49 for Claude Fable 5.1. On Vals Index v2.2, Carbon scored 70.4%, Astra 65.0% and Opus 70.2%; no Fable result was provided. SWE-sweep results were 9.6% for Carbon, 9.5% for Astra, 3.6% for Opus and 2.9% for Fable.
Carbon trailed Astra and Opus on FrontierSWE 2, scoring 61.2% against Astra’s 65.5% and Opus’ 62.3%. Fable scored 56.3%. Carbon led the Terminal-Bench 4.0 AA results with 65.2%, followed by Opus at 59.6%, Astra at 59.1% and Fable at 52.0%.
On Agents’ Last Exam v1.1, Carbon scored 77.4%, Astra 73.1% and Opus 71.6%; Google did not provide a Fable score. In the offline subset of OSWorld-2.1, Opus led with 87.2%, followed by Carbon at 81.5%, Astra at 80.9% and Fable at 76.8%.
Carbon scored 53.9% on AutomationBench’s Zapier evaluation, compared with 41.4% for Astra, 42.5% for Opus and 31.4% for Fable. In Vending-Bench 2, Carbon recorded a bank balance of $16,484, ahead of Astra at $15,515, Opus at $9,235 and Fable at $5,422.
Science and cybersecurity testing
Google DeepMind reported that Carbon scored 72.5% on Chartography, compared with 71.0% for Astra, 66.3% for Opus and 46.2% for Fable. Blueprint-Bench 2 similarity results were 58.6% for Carbon, 49.7% for Astra, 51.2% for Opus and 41.9% for Fable.
On the GraphWalks evaluation under its 256k-to-1M breadth-first search condition, Carbon posted an F1 score of 96.2%, compared with 71.8% for Astra, 66.8% for Opus and 65.0% for Fable. RiemannBench scores were 80.0%, 72.0%, 69.6% and 65.6%, respectively. Terminal-Bench-Science 0.1 AA results were 64.8% for Carbon, 63.3% for Astra, 59.0% for Opus and 43.3% for Fable.
In cybersecurity testing, Carbon scored 71.0% on CWE-bench, compared with 68.0% for Astra, 67.0% for Opus and 58.0% for Fable. ExploitGym results were 54.8% for Carbon, 42.4% for Astra and 34.5% for Opus; no Fable result was released. Carbon scored 88.2% on SEC-bench Pro, while Astra scored 85.4%. Google did not provide scores for Opus or Fable on that test.
Google said Gemini 4 Carbon placed fourth on WebDev Arena with a score of 1,768, representing an 89-point improvement in its Elo rating. All benchmark figures were released by Google or Google DeepMind, and no results from an independent evaluator were presented.
Comments
Post a Comment