Claude Opus 5.5 is the new best available model, 2082 to GPT-6's 2072, led by terminal and professional agent work
The Aggregate Digest — Friday, September 25, 2026
Claude Opus 5.5 is the new best available model, 2082 to GPT-6's 2072, led by terminal and professional agent work.
Claude Opus 5.5 entered the table on Wednesday and has rated above GPT-6 on all three run dates since, so the crown moves today. GPT-6 had held best available model since September 11. The new holder is #3 on the table, below Claude Mythos 5.1 and Claude Mythos 5, which are both private models. Claude Opus 5.5 measures best on long terminal sessions and professional agent work. It is well behind GPT-6 on closed-book facts and on answers grounded in a supplied document.
The Aggregate fitted ranking, September 25 snapshot. The three Claude Mythos rows are private. Claude Opus 5.5 carries a 9-point standard error and GPT-6 a 6-point one; the dot axis is zoomed.
Terminal work leads in two evaluators' runs. Terminal-Bench 4.0 is a set of long command-line tasks in software engineering, science, operations and hardware design. Vals AI's run puts Claude Opus 5.5 first of 34 at 61.62 percent. GPT-6 scores 57.07 there and Claude Opus 5 45.45. LLM Stats' board also has it first of 19. The official tbench.ai board has no Claude Opus 5.5 row yet. On Vals AI's Terminal-Bench Science, 70 long scientific tasks, GPT-6 is ahead at 65.71 to 48.57. Claude Opus 5.5 is second of 30 there.
Professional agent and legal tasks are the widest margins. Mercor APEX runs long professional workflows with tools in finance and law. Claude Opus 5.5 is first of 46 at 73.5 percent. Claude Fable 5.1 has 68.6 and GPT-6 64.7. Legal Benchmarks' Data Extraction board sets lawyer-written extraction tasks over real legal documents. A task counts only when every criterion is met in both attempts. Claude Opus 5.5 passes 76.7 percent of tasks and is first of 20. Claude Fable 5.1 has 63.3. GPT-6 and Claude Opus 5 both have 56.7.
September 25 results from Legal Benchmarks, Mercor, SimpleBench, Vals AI, OpenRouter and Kaggle's FACTS boards. Each row compares the same task and metric across the two models. Different tasks, not a combined score; two rows each come from Vals AI and Kaggle.
GPT-6 keeps facts and grounding. Kaggle's FACTS Parametric asks short closed-book factual questions with no tools. GPT-6 leads it at 83.08 percent. Claude Opus 5.5 scores 57.87 and is eleventh of 40. FACTS Grounding requires the answer to come only from a supplied document. GPT-6 is first of 48 at 81.98. Claude Opus 5.5 is twentieth at 68.46.
Against Claude Opus 5, agent work is up and research is down. Vals AI's Code Migration runs large framework and language upgrades agentically. It rises from 57.47 to 66.65 percent over Claude Opus 5. Research goes the other way. Aiera's institutional-finance research board has models answer through a tool loop over filings and transcripts. Claude Opus 5.5 scores 69.46 on its 0 to 100 research score and is 22nd of 37. Claude Opus 5 scores 87.0. Vals AI's Legal Research Bench falls from 55.29 to 50.48 percent.
September 25 results from Legal Benchmarks, Vals AI, Cursor, SimpleBench, Mercor and Coarena. Each row compares the same task and metric across the two models. Different tasks, not a combined score; three rows come from Vals AI.
The measured case for it is long terminal sessions and document-heavy legal and professional work. GPT-6 still scores higher on closed-book facts and document-grounded answers, and Claude Opus 5 does on research tool loops.
