GPT-6.1 Sol: second only to GPT-6 on expert questions and professional PDFs, and a big step up from GPT-6 Sol
The Aggregate Digest — Wednesday, September 30, 2026
GPT-6.1 Sol: second only to GPT-6 on expert questions and professional PDFs, and a big step up from GPT-6 Sol.
GPT-6.1 Sol enters the table at #3 at 2035. That is seven points above Claude Opus 5.5, inside its own 15-point standard error, and 92 above GPT-6 Sol. Vals AI's cost table puts a Vals Index task at $3.24 with GPT-6.1 Sol, $18.46 with GPT-6 and $32.14 with Claude Opus 5.5, though its own Vals Index score is not yet stored. Best available model stays Claude Opus 5.5 at 2028; a crown moves only after three run dates of lead.
The Aggregate fitted ranking, September 30 snapshot. The two Claude Mythos rows are private. GPT-6.1 Sol carries a 15-point standard error on 114 results and Claude Opus 5.5 a 7-point one; the dot axis is zoomed.
Second of 24 on HLE-Diamond. HLE-Diamond is CAIS and Scale AI's cleaned 1,000-question subset of Humanity's Last Exam. Half the questions test reasoning and half expert knowledge, all answered closed-book without tools. At max effort GPT-6.1 Sol scores 61.1 percent. GPT-6 has 65.8 and Claude Opus 5.5 60.9. GPT-6 Sol at the same setting scores 43.2. The gain is on the reasoning half, where GPT-6.1 Sol is second at 75.0 to GPT-6's 80.8. On the knowledge half it is third behind Claude Opus 5.5 and GPT-6.
lastexam.ai HLE-Diamond board, September 30 snapshot. Eight of 24 entries shown, all max-effort runs; percent correct, closed-book, no tools.
Professional PDFs: two evaluators put it second to GPT-6. GDP.pdf is Surge AI's document-reasoning benchmark over native professional PDF pages: real documents from finance, healthcare, law, engineering or insurance work with their tables and forms intact. A task scores only if it satisfies every criterion. On Surge's board GPT-6.1 Sol passes 32.0 percent of tasks and is second of 39 behind GPT-6 at 34.2. Claude Opus 5.5 has 30.6 there and GPT-6 Sol 26.4. Artificial Analysis' run agrees: 32.0 at high effort, second of 195 behind GPT-6 at 32.2, with Claude Opus 5.5's best run at 28.8.
The gains over GPT-6 Sol come on reasoning and agent tasks. Two coding tests slip. EnigmaForge hides the question inside a story, and the model has to reconstruct the world from narrative clues. At low effort GPT-6.1 Sol is first of 24 at 84.5 percent. GPT-6 has 80.3 and GPT-6 Sol 74.2. Claude Opus 5.5 at the same setting is 14th at 39.8. Bug Hunt Bench has coding models find and repair 105 planted defects in two production repositories. At max effort GPT-6.1 Sol fixes 44 of them, third of 98 behind Claude Sonnet 5.5 and GPT-6, while Claude Opus 5.5 fixes 41.7 and GPT-6 Sol 29.3. Artificial Analysis' SciCode sets scientist-written Python subproblems. There GPT-6.1 Sol falls from GPT-6 Sol's 57.64 to 54.17 percent at max effort, and LiveBench code generation falls from 83.1 to 80.28.
September 30 results from lastexam.ai, EnigmaForge, the Explicit Edit Benchmark, LLM Stats, Artificial Analysis and LiveBench. Each row compares the same task and metric across the two models, at the same effort where the board records one. Different tasks, not a combined score; two rows each come from LLM Stats and Artificial Analysis.
Claude Opus 5.5 keeps code generation and computer use. DeepSWE 1.1 is DataCurve's set of long-horizon software engineering tasks in realistic repositories. In LLM Stats' listing GPT-6.1 Sol is second of 40 at 75.22 percent to Claude Opus 5.5's 74.2. OSWorld 2.0 runs 108 full desktop workflows scored on completion. In LLM Stats' listing GPT-6.1 Sol is fourth of 15 at 71.42 and Claude Opus 5.5 first at 81.8. LiveBench code generation goes 91.55 to 80.28 at max effort and SciCode 66.9 to 54.17.
September 30 results from lastexam.ai, Artificial Analysis, Surge AI, LLM Stats and LiveBench. Each row compares the same task and metric across the two models, at the same effort where the board records one. Different tasks, not a combined score; two rows each come from lastexam.ai, LLM Stats and Artificial Analysis.
On this evidence the tasks worth testing are closed-book reasoning and PDF document work. Bug repair and story puzzles come next. Code generation and desktop computer use still favour Claude Opus 5.5, as do scientific coding and knowledge questions.
