GPT-6.1 Sol Benchmarks, Explained Simply

OpenAI's claimed results (Sept 2026 announcement), translated out of benchmark-speak. These are vendor numbers — useful for comparison, not gospel.

GPT-6.1 Sol benchmark claims vs GPT-6 Sol, Astra and others (OpenAI, Sept 2026)
Benchmark (what it tests)	Claimed result	Plain meaning
DeepSWE v1.1 — real coding tasks	Matches Astra at ~1/5 cost; +6.4pp over GPT-6 Sol	Codes about as well as the flagship, much cheaper
GDP.pdf — answers from complex PDFs	Beats Opus 5.5 at <½ cost; near Astra at ~1/5 cost	Strong at reading messy business documents
AutomationBench — multi-step business workflows	+2.2pp over Opus 5.5 at ~⅓ cost; +4.8pp over GPT-6 Sol	Good office-automation agent material
OSWorld 2.0 — using computer apps	+7pp over GPT-6 Sol; within 2.1pp of Astra at ~1/7 cost	Noticeably better at clicking around software
Terminal-Bench Science — data analysis, simulations	2× GPT-6 Sol's score at <½ cost ($5.47 vs $23+ per task)	Cheap science grunt work
Factuality — share of answers with an error (hard prompts)	Errors 11.4% → 7.7% at low effort (−32%)	Fewer hallucinations, still not zero

Data sources & limitations

All scores above are OpenAI's stated numbers from the Sept 2026 launch announcement — not independent tests. OpenAI notes evals ran in its research environment and competitor figures came from public reports.
Benchmarks measure narrow tasks (e.g. scripted coding or PDF QA), not everyday chat quality. Treat gaps of a few points as noise.
Costs are per-task averages at stated reasoning settings; your bill depends on effort level, caching, and retries.
If a third-party replication disagrees, trust the replication and check back — this page is updated when credible new numbers appear.

Sources: OpenAI announcement with links to each benchmark's own page. See also: Sol vs Luna vs Astra.