The Short Version
On October 28, 2026, Google announced Gemini 4 Carbon and took the top spot on the Artificial Analysis Intelligence Index v5.0 with 59. Claude Opus 5.5 scores 58 at its max setting, and GPT-6 Astra 57. The gap between them is small enough that the effort setting you choose matters as much as the model. The bigger differences are price and access: Opus 5.5 and Astra are on general sale, while Carbon is limited to selected cyber defenders.
Side by Side
| Gemini 4 Carbon | Claude Opus 5.5 | GPT-6 Astra | |
|---|---|---|---|
| Maker | Anthropic | OpenAI | |
| Released | October 28, 2026 (limited) | September 22, 2026 | Early September 2026 |
| AA Intelligence Index v5.0 | 59 (high), 58 (medium) | 58 (max), 57 (xhigh) | 57 (max) |
| Price per M tokens, input / output | $2 / $10 introductory, then $4 / $20 | $4 / $20 | $10 / $50; prompts over 272K tokens cost more |
| Context | 1M (Artificial Analysis listing) | 1M, 128K max output | 1.05M (922K max input), 128K max output |
| Knowledge cutoff | Not stated | June 2026 | April 30, 2026 |
| Who can use it | Selected cyber defenders, then partners through Google's Fairwind Program. No public date | Claude apps, API, Bedrock, Google Cloud, Microsoft Foundry | ChatGPT and the OpenAI API |
The scores are from the Artificial Analysis leaderboard, read on November 2, 2026. They change as models and settings are added, so name the variant when you quote one.
Gemini 4 Carbon
Carbon is aimed at security and long engineering work. Google reports 88.2% on SEC-bench Pro, 71.0% on CWE-bench, 54.8% on ExploitGym and a 1768 Elo on WebDev Arena, all vendor-reported. Google says Carbon makes about half as many unsupported claims as Gemini 4 Argon. One secondary report cites an AA-Omniscience hallucination rate of 11% for Carbon against 4% for Argon, which we could not confirm on Artificial Analysis' page. Most people can't use Carbon yet. For the full release history, see the Gemini 4 release brief and the Gemini 4 Carbon FAQ.
Claude Opus 5.5
Opus 5.5 is the strongest of the three that anyone can buy today at its price. Adaptive thinking is always on, with medium as the default effort. It is available on every major cloud. SemiAnalysis' new GDPx agent benchmark reportedly put Opus 5.5 first at 12.0%, which says as much about how hard the benchmark is as about the model. See the Claude Opus 5.5 release brief.
GPT-6 Astra
Astra is the most expensive by a wide margin, at $10 input and $50 output per million tokens. Batch and Flex processing cut that in half. It powers ChatGPT, which still has the largest consumer audience of any assistant, so it shapes more brand answers than either rival. The Wall Street Journal reported in late September that OpenAI cancelled a planned October release of GPT-6.1 Astra after safety testing. OpenAI has not commented publicly.
What It Means for Brand Visibility
The model that matters most to a brand is the one its buyers use, not the one at the top of a leaderboard. Astra reaches ChatGPT's users. Opus 5.5 reaches Claude's users and many coding and agent tools. Carbon reaches almost nobody yet, but it previews what Gemini and AI Mode will run on later. Run the same prompt set on each, and compare which competitors each one names and which sources it cites. See ChatGPT vs Claude vs Gemini vs Perplexity for the assistant-level view.