Gemini 4 Carbon Takes the Top Spot on the Intelligence Index With 59 Points – Broad Launch Still Has No Date
The new frontier model tops Artificial Analysis' independent ranking with 59 points, 15 days after Argon was rolled out. But access is for now reserved for selected security experts in Google's Fairwind program – and the company has neither released a technical report, a public API, nor given a date for a broad launch.
What Happened
Google DeepMind launched Gemini 4 Carbon on October 28, 2026 – just 15 days after its predecessor Argon was rolled out more broadly on October 13. The blog post is signed by Koray Kavukcuoglu, SVP at Google DeepMind, who according to Mashable now leads the lab after Demis Hassabis stepped aside from the role. Kavukcuoglu calls Carbon Google's strongest model to date (Mashable).
The launch continues a remarkable pace on the frontier front: Anthropic launched Claude Opus 5.5 on September 22, and Argon was announced on September 30. OpenAI's GPT-6 Astra is also part of the picture of the top field – but none of the coverage gives a date for that launch.
Leads the Independent Ranking
The independent evaluator Artificial Analysis places Carbon, with high reasoning, in first place on the Intelligence Index v5.0 with 59 points – ahead of Claude Opus 5.5 (58), GPT-6 Astra (57) and Argon (50). That is according to VentureBeat, which at the same time points out that the index has been updated and has become harder (VentureBeat).
In Google's own comparison table, where all figures are as stated by the company itself, Carbon has the top score on 15 of 17 benchmarks against Argon, GPT-6 Astra, Claude Opus 5.5 and Claude Fable 5.1. GPT-6 Astra leads FrontierSWE 2, while Claude Opus 5.5 leads OSWorld-2.1 (Mashable). Note that The Decoder, in its coverage, refers to 15 of 16 tests in Google's selection – the discrepancy between the sources cannot be reconciled from the available reporting.
Cybersecurity in Focus – Google's Numbers
What sets Carbon apart from recent launches is its stated emphasis on cybersecurity: finding vulnerabilities, exploiting them and producing patches. Google claims the model has a "firm lead" on security benchmarks, with 88.2% on SEC-bench Pro (vulnerability discovery), 54.8% on ExploitGym (exploitation) and 71% on CWE-bench (patching). These are the company's own figures, reported by Anadolu Agency – not independently verified (Anadolu Agency).
The company describes the model as a "significant leap" in cyberdefense, making it better at finding vulnerabilities and generating secure code, and links the pace of the launch to its own researchers and AI models working "hand in hand to advance capabilities faster" (The Verge, Anadolu Agency). Google further claims that internal testing shows Carbon roughly halves the number of unsupported claims compared with Argon, and reduces both task overreach (26%) and under-service (29%) in multi-step workflows – again, company-reported figures (Anadolu Agency).
Locked Behind Fairwind
This is where the story becomes unusual: the model that leads the rankings is, to begin with, not available. Google is granting access only to selected cybersecurity researchers in the curated "Fairwind Program". The reason is precisely the advanced security capabilities: Carbon is going through the U.S. government's voluntary pre-testing process and will first become available to selected security experts (VentureBeat, The Decoder).
For a broader launch, Google promises "as soon as possible", after "gathering feedback and iterating on guardrails" – but no date is given, in either The Decoder's or VentureBeat's coverage.
Independent Numbers Show Trade-offs
The independent testing that The Decoder refers to paints a more nuanced picture than Google's own claims. Answer accuracy rises from 48 to 58 percent compared with Argon, but the hallucination rate climbs from 4 to 11 percent. That is higher, but still the second-lowest in the ranking (The Decoder).
And on agentic coding in Antigravity, Carbon still trails Claude Code with Opus 5.5 – by three points, even though the score, according to The Decoder, jumped from 53 to 59 compared with Argon. The Decoder also points out that Gray Swan's prompt-injection test "appears to be saturated", as an indication that the test's ability to distinguish between models is limited (The Decoder).
Pricing – With No End Date
When the model opens up more broadly, it will, according to VentureBeat, launch at an introductory price of $2 per million input tokens and $10 per million output tokens, with a 95% discount on cached inputs. Later, the standard price will rise to $4 and $20. VentureBeat points out that Google has not given any end date for the introductory pricing – and that Carbon, for now, lacks a comparable public API endpoint (VentureBeat).
Market Reaction
The stock market took the launch in stride: Alphabet shares rose around 1.5% in early afternoon trading on Wednesday, October 28, following the announcement (Seeking Alpha).
What We Still Don't Know
Several key pieces of information are missing. Google has not released a technical report, has not said whether Carbon uses a new model architecture, and has not given a parameter count (VentureBeat). Among other things, this means it remains unclear whether Argon's limitations in context window and modalities also apply to Carbon.
The company's security and benchmark claims have, for now, largely not been independently verified in the coverage. And the most important date – when Carbon becomes available to ordinary developers and enterprises – is one Google simply has not given. Until then, most people can read the benchmarks, but not use the model.