gpt-5.6-sol
Meant to excel at
-
Flagship for complex reasoning, coding and agentic work, described as particularly strong at command-line and multi-step coding.
Independent evaluator openrouter.ai -
Strong browsing and computer use: BrowseComp 90.4 percent and OSWorld 2.0 62.6 percent as reported by OpenAI.
Vendor claim, relayed chatbase.co -
Highest presentation Elo in Artificial Analysis's AA-Briefcase evaluation; programmatic tool calling cuts round trips.
Independent evaluator chatbase.co -
Much stronger execution than GPT-5.5 on cross-stack debugging and notably more polished Office documents.
Independent evaluator laprompt.com -
High taste scores on prototypes, with better visual hierarchy and interactivity than alternatives tested.
Independent evaluator lennysnewsletter.com
Watch out for
-
Sonar found higher bug and vulnerability density in its generated Java code, so testing and review remain necessary.
Independent evaluator chatbase.co -
Long-context recall fell to 73.8 percent at 512K-1M tokens, and requests over 272K tokens bill higher.
Independent evaluator chatbase.co -
Long threads degrade, subagents multiply token use, and high effort can exhaust five-hour windows.
Practitioner report laprompt.com -
Priciest in its family at $5 and $30 per million tokens; poor fit for simple classification or extraction.
Independent evaluator chatbase.co
Benchmarks
Agentic coding percentile 0.70
| Benchmark | Score | Percentile | Source | Date |
|---|---|---|---|---|
| artificial analysis coding agent index | 80 | 1.00 | Independent evaluator artificialanalysis.ai | 2026-07-09 |
| deepswe v1.1 | 72.7 | 0.50 | Vendor's own site openai.com | 2026-09-03 |
| terminal-bench 2.1 | 88.8 | 0.82 | Vendor claim, relayed rits.shanghai.nyu.edu | 2026-07-09 |
| terminal-bench 2.1 (aggregator, anotherwrapper.com) | 88.8 | 0.85 | Independent evaluator anotherwrapper.com | 2026-10-19 |
| terminal-bench 4.0 | 37.3 | 0.31 | Vendor's own site openai.com | 2026-09-03 |
Reasoning percentile 0.83
| Benchmark | Score | Percentile | Source | Date |
|---|---|---|---|---|
| gpqa diamond | 94.6 | 0.83 | Vendor's own site openai.com | 2026-09-03 |
Percentile is the share of tracked models this model beats on that benchmark, averaged across the group. Only benchmarks at least four tracked models share are scored.
Practitioner reports
- +
A code-review benchmark gave GPT-5.6 Sol a relative bug rate of 0.72, better than GPT-6 Sol's 1.38 (secondhand roundup).
Compared with gpt-6-sol | 2026-09-29
Practitioner report hackernoon.com - -
Sol's prose scored a readability grade of 8.74 versus 6.95 for Opus 5.5, meaning more complex, harder-to-read writing.
Compared with claude-opus-5.5 | 2026-09-22
Practitioner report every.to - +/-
Sol beat Fable 5.1 on structured tasks, compression and short posts, while Fable was better on judgment-heavy slide decks and persona answers.
Compared with claude-fable-5.1 | 2026-09-01
Practitioner report every.to - +/-
Execution and Office-file polish improved over GPT-5.5, but long threads degraded and a compaction loop burned an estimated 10 to 15 percent of an allowance.
Compared with gpt-5.5 | 2026-07-31 older than 90 days
Practitioner report laprompt.com - +/-
Finds Sol at max effort useful as a planner and reviewer but not as the hands-on coder; suggests pairing it with GPT-5.5 High for bounded implementation.
Task: code-review | Compared with gpt-5.5 | 2026-07-15 older than 90 days
Practitioner report community.openai.com - +/-
On extra high in Codex, follows specs that express intent (overly detailed specs hurt), handles UX feedback better than 5.5, and sometimes makes useful out-of-scope refactors; but misses visual problems such as overlaps unless shown screenshots.
Task: feature | Compared with gpt-5.5 | 2026-07-13 older than 90 days
Practitioner report community.openai.com - +/-
Same-thread user running Sol at extra high for 6-8 hour days reports no weekly limit hits and says false 'fixed' claims are rare and tied to missing tools to see UX bugs, while another user saw 5.6 claim an issue resolved when it was still present.
Task: bug-fix | 2026-07-13 older than 90 days
Practitioner report community.openai.com - +
Made Sol a daily driver after it won a taste-scored benchmark of PRDs, prototypes and debugging, and said it reconsidered approaches more readily than Fable.
Compared with claude-fable-5 | 2026-07-13 older than 90 days
Practitioner report lennysnewsletter.com - +/-
Used Codex with the Chrome tool to answer around 500 LinkedIn messages, fill forms and test web apps, though it has a green palette bias and weak voice output.
Compared with claude-sonnet-5 | 2026-07-13 older than 90 days
Practitioner report lennysnewsletter.com - -
Says Sol is not usable for serious terminal coding workflows: it loses constraints, repeats mistakes and drifts from instructions. Several replies agree, and one says it ignores architecture docs and invents services.
Task: agentic-long-running | Compared with gpt-5.5 | 2026-07-12 older than 90 days
Practitioner report community.openai.com - +/-
Roundup of seven creator tests: Sol turns a defined job into a finished artifact quickly and cheaply in Codex (one game build about $4.50 vs about $14.22 for Claude Fable) but Fable was more imaginative; Ultra mode multi-agent runs can cost hundreds of dollars. Not a controlled comparison.
Task: feature | Compared with claude-fable-5 | 2026-07-12 older than 90 days
Practitioner report ai.joaoqueiros.com - +/-
CodeRabbit recommends Sol for long-horizon coding and for review passes where missing a bug is costly, but notes noisy comments (precision 31.6%, 61 nitpicks), fuzzy judgment and an occasional stuck loop (one simple change took eight turns).
Task: code-review | Compared with gpt-5.6-terra | 2026-07-09 older than 90 days
Practitioner report coderabbit.ai