Model cards

gpt-6-sol

Vendor openai Tier standard Price per million tokens, in / out $2 / $10 Released 2026-09-22 Context 1,050,000 Max output 128,000 Vendor page

Last updated 2026-10-26 How this is made

Meant to excel at

• Hallucination rate on Artificial Analysis AA-Omniscience dropped from 92% (GPT-5.6 Sol) to 60%, and Terminal-Bench 4.0 rose to 43-44% from 37-40%.

  Independent evaluator artificialanalysis.ai

Watch out for

• Knowledge-work deliverables regressed: GDPval-AA fell about 100 Elo vs GPT-5.6 Sol, which Artificial Analysis attributes to weaker presentation and omitted rubric elements.

  Independent evaluator artificialanalysis.ai

• Security-focused tests lag Astra: Endor Labs SecPass 25.1% vs 34.6%, and Sonar found 285 vulnerabilities per million Java lines vs 178 for Astra (Endor is a security vendor testing its own benchmark).

  Independent evaluator endorlabs.com

• System card reports statistically significant regressions in some disallowed-content categories (e.g. self-harm) vs GPT-5.6 (secondary report of the OpenAI system card).

  Vendor claim, relayed andrew.ooo

Benchmarks

Coding percentile 0.70

Agentic coding percentile 0.20

Knowledge percentile 0.54

Overall percentile 0.25

Tool use percentile 0.33

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

• +/-

  Codex run on 179 security-sensitive tasks: Sol was the cheapest run Endor Labs had measured ($0.52 per task vs $2.34 for Astra) and improved functional and security pass rates over GPT-5.6 Sol, but trailed Astra on both (72.1% vs 82.1% functional).

  Compared with gpt-6-astra | 2026-10-01

  Practitioner report endorlabs.com
• -

  Several commenters found GPT-6 Sol worse than 5.6 Sol: one spent a day fixing its code before switching back to 5.6, another saw degraded results when refactoring, and a third calls it a regression. Comments under a launch write-up, so anecdotal.

  Task: refactor | Compared with gpt-5.6-sol | 2026-09-29

  Practitioner report handyai.substack.com
• +/-

  Sonar's Java analysis found Sol passed 83.09% of tests (1.1 points above GPT-5.6 Sol) but its vulnerability density was 60% higher than Astra's and it wrote about 27% less code with sparse comments.

  Compared with gpt-6-astra | 2026-09-24

  Practitioner report sonarsource.com
• +/-

  Calls Sol at high decent for agent orchestration but says it needs a larger context window.

  Task: agentic-long-running | 2026-09-23

  Practitioner report community.openai.com
• +

  On a poster's three-task Codex test, Sol at high passed all three at roughly $0.02-0.06 per task, while GPT-5.6 Luna had failed the dependency-scheduler task.

  Task: feature | Compared with gpt-5.6-luna | 2026-09-22

  Practitioner report community.openai.com
• -

  Reports that on AutomationBench-style workflow problems GPT-6 Sol at max reasoning made no headway and produced worse output at higher cost than lower effort.

  Task: operations-support | 2026-09-22

  Practitioner report community.openai.com
• -

  Analysis built on Artificial Analysis data: coding is the one clear gain over 5.6 Sol (Coding Agent Index 57 vs 55, similar cost-halving) while knowledge-work deliverables regressed (GDPval-AA about 100 Elo lower, weaker presentation); calls the launch a price cut with roughly level capability.

  Task: writing-content | Compared with gpt-5.6-sol | 2026-09-22

  Practitioner report codingfleet.com