Google rolls out Gemini 4 Carbon, reclaims top benchmark spot but limits early access

Google’s latest AI claim to fame

Google announced Gemini 4 Carbon as its newest frontier AI model, positioning it as an upgrade over the earlier Argon flagship. The company says Carbon improves cybersecurity and complex software‑engineering tasks while keeping the same introductory token pricing as Argon.

Benchmark supremacy

Independent evaluator Artificial Analysis placed Carbon at the top of its Intelligence Index v5.0 with 59 points, ahead of Claude Opus 5.5 (58), GPT‑6 Astra (57) and Argon (50). In Google’s own benchmark suite, Carbon achieved the highest result in 15 of 17 rows. OpenAI retained an advantage on one software‑engineering test, and Anthropic led on a computer‑use evaluation.

Pricing details

Google disclosed an introductory rate of $2 per million input tokens and $10 per million output tokens, with a 95 % discount for cached inputs. A footnote indicates that after the introductory period the rates will rise to $4 and $20 respectively, though the company has not said when the introductory window ends.

Access restrictions

Unlike Argon’s broader rollout on October 13, Carbon is initially available only to selected cybersecurity defenders through Google’s curated “Fairwind Program.” Google said a wider release will follow after gathering feedback and iterating on guardrails.

Performance trade‑offs

Testing shows Carbon makes substantial gains over Argon in coding and knowledge‑work scenarios, yet it falls behind several rivals on certain professional tasks and exhibits a higher hallucination rate than Argon on a factual‑knowledge test.

What’s still missing

Google has not released a technical report for Carbon, nor disclosed its parameter count or whether it uses a new architecture compared with Argon. Argon remains the publicly available model for customers, offering a shared context capacity of over 1 million tokens and an extended generation limit beyond the default 65,536‑token output cap.

Internal perspective

The announcement, authored by Koray Kavukcuoglu—Google DeepMind SVP and chief AI architect—notes that internal teams have observed Carbon running for days, handling prolonged debugging, bug diagnosis, and patch development in large codebases.

While Carbon’s benchmark lead is notable, the limited early access and pending price changes mean many developers and businesses will have to wait before evaluating whether the model truly delivers on its promises.