Gemini 4 Carbon replaces Claude Opus 5.5 as the best measured model, but it is private
The Aggregate Digest — Saturday, October 31, 2026
Gemini 4 Carbon replaces Claude Opus 5.5 as the best measured model, but it is private
Google's Gemini 4 Carbon is now the best measured model. Its Aggregate Rating of 2030 is the highest of any model, and it is first on tests where agents run a simulated business or automate business workflows. It is private. The best available model therefore stays Claude Opus 5.5 at 1995. Gemini 4 Carbon first led Claude Opus 5.5 on October 29, by 5 rating points, and the lead has grown to 35 rating points.
First at running a business. In Andon Labs' Vending-Bench 2, an agent runs a vending machine business for a simulated year from a $500 start. It has to find suppliers, negotiate, order stock and set prices. Gemini 4 Carbon ends the year with $16,484 on average, against $15,515 for GPT-6 Astra and $9,235 for Claude Opus 5.5. It is also first on AutomationBench, where agents complete business workflows in connected apps. That holds in both Zapier's run and Artificial Analysis' run.
Andon Labs, October 31, 2026. A simulated year running a vending machine business from a $500 start. The board's top seven of 73 and Claude Opus 5.5; no reasoning setting is listed for any of them.
Fewer wrong guesses. Artificial Analysis' AA-Omniscience asks 3,000 factual-recall questions in six domains. Of the questions Gemini 4 Carbon at high effort did not answer correctly, it answered 11.42% wrongly and declined the rest. For Claude Opus 5.5 at max effort with default fallback that share is 50.34%. Claude Opus 5.5 does answer more questions correctly: 65.5% against 58.27%.
Claude Opus 5.5 is still ahead on code and long documents. Artificial Analysis' coding-agent index scores each model inside a coding agent. There Claude Opus 5.5 at max effort with default fallback scores 61.65 in Claude Code, and Gemini 4 Carbon at high effort scores 59.4 in Antigravity CLI. Artificial Analysis' long-context test gives about 100,000 tokens of documents with each question. Claude Opus 5.5 at max effort with default fallback answers 84.67% correctly, and Gemini 4 Carbon at high effort 79.33%.
Evaluators split on professional tasks. On Mercor APEX-Agents, Gemini 4 Carbon at high effort passes 84.2% against 73.5% for Claude Opus 5.5 at max effort. Claude Opus 5.5 at max effort with default fallback is first on AA GDPval, Artificial Analysis' run of tasks from 44 occupations.
Artificial Analysis, Mercor and Vals AI, October 31, 2026. Gemini 4 Carbon at high effort; Claude Opus 5.5 at max effort, with default fallback at Artificial Analysis. AA workplace is Artificial Analysis' AutomationBench run, and the two Terminal-Bench 4.0 runs disagree. Points on each board's 0–100 scale; different tests, not a combined score.
Google's Gemini 4 Argon has API access. It comes next after Gemini 4 Carbon on both AutomationBench runs and on Mercor APEX-Agents, within three percentage points of it.
