Model cards

gpt-6.1-sol

Vendor openai Tier standard Price per million tokens, in / out $2 / $10 Released 2026-09-29 Context 1,050,000 Max output 128,000 Vendor page

Last updated 2026-10-26 How this is made

Meant to excel at

• Near-Astra performance on agentic coding, computer use and professional work at one-fifth of Astra's price ($2 and $10 per million).

  Vendor's own site openai.com

• Strong at multi-step business workflows and scientific work, with cheap cached input at $0.10 per million tokens.

  Vendor's own site openai.com

• Highest Terminal-Bench 4.0 score in its comparison table (58.18 percent) and top code-migration reviewer rating (4.5 of 5).

  Independent evaluator gradually.ai

• Follows specs literally and writes the smallest code, suiting high-volume agentic coding with precise requirements.

  Independent evaluator thomas-wiegold.com

• Fewer alignment failures than GPT-6 Sol, such as respecting restrictions and disclosing broken tools.

  Vendor's own site openai.com

• Perfect 43/43 on Plotly's analytics test at about 10 percent of Astra's cost, finding all five hidden-data issues.

  Practitioner report plotly.com

• On Artificial Analysis, accuracy and knowledge reliability improved sharply over GPT-6 Sol (AA-Omniscience 41.5 vs 27.1; hallucination rate 54% vs 60%) at about $0.72 per index task vs $3.26 for Astra (relayed via a reseller blog).

  Independent evaluator orcarouter.ai

• Sonar's Java analysis shows vulnerability density down about 32% vs GPT-6 Sol (193 vs 285 per million lines) with pass rate 85.29%, just under Astra's 85.85%.

  Independent evaluator sonarsource.com

Watch out for

• Slow, about 51 tokens per second at max effort; literal interpretation means prompts need more detail.

  Independent evaluator thomas-wiegold.com

• Weak on cybersecurity: 39.29 percent on CyberBench with 0 percent on proof-of-concept tasks.

  Independent evaluator gradually.ai

• Astra remains better for the hardest science and computer-use work; Sol trails by about 2 points on OSWorld 2.0.

  Vendor's own site openai.com

• Heavy tasks can be costly and slow, for example about $6.51 and 96 minutes per code migration at max effort.

  Independent evaluator gradually.ai

• Cost per run on small bug fixes was lower than Argon and Sonnet 5.5 in one user's eval, but see Argon note on pi cost reporting.

  Practitioner report news.ycombinator.com

• Writes much more code than GPT-6 Sol (about 39% more lines in Sonar's Java run); total bugs and code smells rose in raw count even though per-line density fell, so review load grows.

  Independent evaluator sonarsource.com

• Trails Astra on Artificial Analysis Terminal-Bench 4.0 (56.1 vs 59.1), SciCode and Humanity's Last Exam, and Astra has a lower hallucination rate (51% vs 54%) (relayed via a reseller blog).

  Independent evaluator orcarouter.ai

• Ultrafast service tier (API, from 2026-10-08) costs six times Standard ($12/$60 per million tokens) and burns subscription usage at 8x; the speedup multiple has not been independently measured (relayed via a reseller blog).

  Vendor claim, relayed orcarouter.ai

Benchmarks

Coding percentile 0.87

Agentic coding percentile 0.74

Knowledge percentile 0.61

Overall percentile 0.42

Tool use percentile 0.67

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

• +

  Sol scored 43/43 on Plotly's 43-question analytics test, matching Astra at about a tenth of the cost, with a median of 18s per question. Single run, private questions, one domain and harness, and cost estimates are rough.

  Task: data-analysis | Compared with gpt-6-astra | 2026-10-14

  Practitioner report plotly.com
• -

  GitHub issue reports a two-week drop in code comprehension quality in the Codex VS Code extension with 6.1 Sol at extra high (also Astra and 5.6 Sol): needs 3-4 correction turns, misses architectural risks, slow; same question in ChatGPT web was clearly better. Single-user observation, no traces attached.

  Task: code-review | 2026-10-13

  Practitioner report github.com
• +/-

  Sonar's large Java benchmark found 6.1 Sol close to Astra on pass rate with fewer vulnerabilities per line than GPT-6 Sol, but reasoning tokens fell 51% while output volume grew, producing more raw bugs and smells to triage.

  Compared with gpt-6-sol | 2026-10-09

  Practitioner report sonarsource.com
• +

  Uses a cheaper reverse-orchestration setup, with Luna and 6.1 Sol doing reviews, and says it works very well in practice.

  Task: code-review | Compared with gpt-6-astra | 2026-10-08

  Practitioner report community.openai.com
• +

  Ran three agent-style tasks five times each at max effort (CI-log triage, 3,664-line incident-log postmortem, dependency resolver written from a spec against a hidden 120-test suite): both models went 15/15, 6.1 Sol cost about 18% of Astra ($2.66 vs $14.77) and was faster on two of three tasks.

  Task: operations-support | Compared with gpt-6-astra | 2026-10-07

  Practitioner report thenewstack.io
• +

  On a poster's own DigBench runs (high effort) 6.1 Sol won both tasks, finishing in about 10 and 25 minutes at roughly $1.11 and $2.46 API-equivalent, ahead of GPT-6 Sol at high. Task type not described; poster's unverified numbers.

  Task: agentic-long-running | Compared with gpt-6-sol | 2026-10-06

  Practitioner report community.openai.com
• -

  Sol produced minimal design with no animations, invented grammar in an interpreter, and wrote the thinnest tests; Opus 5.5 won the design and interpreter tests.

  Compared with claude-opus-5.5 | 2026-10-05

  Practitioner report thomas-wiegold.com
• +/-

  Sol followed the spec most literally and wrote the least code, but its tests missed 6 of 36 seeded bugs and its search ran far slower than Opus.

  Compared with claude-opus-5.5 | 2026-10-05

  Practitioner report thomas-wiegold.com
• +/-

  Across six reasoning efforts, max produced the best-ranked code in both blind reviews (maintainable structure, real tests) but took 1h16 and 37k reasoning tokens; low and medium looked fine in the browser but had hard-to-maintain code.

  2026-10-04

  Practitioner report spectrumailab.com
• +

  Versus GPT-6 Sol, 6.1 Sol was equally correct but used about 30% fewer tokens on a messy sales CSV aggregation and prioritised planted bugs better in a code review; estimated API cost per run about 20% lower. Versus Astra the review feedback was nearly the same at roughly one fifth of the cost (about $0.06 vs $0.30).

  Compared with gpt-6-sol | 2026-10-02

  Practitioner report note.com
• +

  Hand-built test of 15 support tickets with a policy doc and one lookup tool: 6.1 Sol at low, medium and high all got the policy outcome right 15/15, same as Astra at medium, at about $1.80 per 1,000 tickets versus $9.09 for Astra. Some setups over-escalated, and none caught a letter O typed in an order ID. Author says it is not a benchmark and the publisher sells support agents.

  Task: operations-support | Compared with gpt-6-astra | 2026-10-01

  Practitioner report eesel.ai
• -

  Pro user says 6.1 Sol increasingly refuses or sanitizes ordinary site-admin, SEO, debugging and moderation-config work when the site contains lawful adult content; GPT-5.6 Sol completed these but burns far more Work/Codex quota, so fallback is not workable.

  Task: operations-support | Compared with gpt-5.6-sol | 2026-10-01

  Practitioner report community.openai.com
• +/-

  In three Codex builds (browser OS, 3D apartment, simulated robot arm) the author found 6.1 Sol faster and more cost-efficient than Sonnet 5.5, but Sonnet's visuals, environments and spatial accuracy were stronger.

  Compared with claude-sonnet-5.5 | 2026-10-01

  Practitioner report aifire.co
• +/-

  Sol scored 97.5 percent on an eight-task Codex set, above GPT-6 Sol but below Opus 5.5's reported score; weak spots were dispatch logic and collision geometry.

  Compared with claude-opus-5.5 | 2026-09-30

  Practitioner report daily.dev
• +/-

  Code-migration benchmark gave Sol the best reviewer rating but it trailed Astra on overall score, taking about 96 minutes per task at max effort.

  Compared with gpt-6-astra undated

  Practitioner report gradually.ai