GPT-5.6 Sol: benchmarks, price and rank

Its headline results

The benchmarks that count most in the rating, with its place among the models measured the same way, and where each number came from.

• Terminal-Bench 4.037.3%

  Hard jobs done in a real computer terminal: building and fixing software, setting up systems, and science and engineering problems.

  Independent result from LLM Stats.

• BrowseComp90.4%

  Questions whose answers are hidden deep on the web and take many searches and pages to track down.

  5th of 56. The maker's own number, from OpenAI, via LLM Stats.

• SWE-Bench Pro64.6%

  Long changes to real codebases in several languages, often across many files, including code from private and copyleft projects that models are unlikely to have seen.

  9th of 50. The maker's own number, from OpenAI, via LLM Stats.

• FrontierSWE32.2%

  Open-ended technical projects that take hours to tens of hours, from making systems faster to building large programs and applied machine learning research. Each run gets a 20-hour budget.

  11th of 21. Independent result from Epoch AI.

• ARC-AGI-37.8%

  Interactive puzzle games the model has never seen, with no instructions. It has to explore, work out the goal and the rules, and win.

  4th of 6. The maker's own number, from OpenAI, via LLM Stats.

• Vending-Bench 2$9,619

  The model runs a simulated vending machine business for a year on its own: finding suppliers, ordering stock, setting prices and paying the daily fees.

  10th of 60. Independent result from Epoch AI.

• Toolathlon58%

  Long tasks across many real apps and their tools, such as email, calendars, spreadsheets, code repositories and online stores.

  10th of 37. The maker's own number, from OpenAI, via LLM Stats.

• CursorBench41.7%

  Ambiguous coding tasks that span many files, taken from real sessions in the Cursor code editor.

  13th of 20. Independent result from Epoch AI.

• DeepSWE 1.173%

  Software engineering tasks run in the same simple agent setup for every model.

  Independent result from LLM Stats.

• SimpleBench64.8%

  Multiple-choice puzzles about everyday situations, space, time and people, written to trip up models that pattern-match instead of think. Ordinary adults do well on them.

  30th of 68. Independent result from Epoch AI.