‹ PublicAI Index
The LLM benchmark aggregator.
GPT-5.6 Sol
OpenAI
Strongest in Science (#2 of 122), weakest in Toxicity avoidance (#225 of 268). Above par in 31 of 36 scopes. Among the models it meets almost everywhere, it finishes behind Claude Fable 5.1 and Claude Opus 5.5 and ahead of Claude Opus 4.8 and Claude Sonnet 5.5.
§ 1 · Profile
What it is good at
Bars run from 50 — the average of the models each source lists — so right of the line is above par. The middle column is the gap to whoever leads that scope.
Knowledge59.7−2.8#9/1972/3
Document parsing60−3.3#9/831/1
Academic knowledge59.8−4.1#16/1231/1
Multimodal understanding53.2−18.9—/19✱0/1
Reasoning60.6−4.8#11/1954/5
Science63.4−0.1#2/1221/1
Mathematics60−3.8#13/1522/2
Reasoning59.7−9.3#21/1472/3
Expert reasoning44.4−22—/0✱✱
Coding59.5−6.7#12/1793/6
Agentic coding59.9−7#11/1733/5
Repository Q&A44.5−22.2—/0✱✱
Professional57.1−5.3#15/1941/2
Cybersecurity60.3−3.6#4/471/1
IT operations62.1−11.7#6/331/1
Biology research52.2−18.5#13/311/1
Legal59.5−6.4#14/1731/1
Finance57−5.4#31/1621/1
Medical54−12.5#55/1491/1
Human preference64.8−4#19/3541/1
Human preference64.8−4#19/3541/1
Core abilities58.1−7.2#21/1992/4
Language62.7−3.9#7/621/1
Instruction following56.8−17#21/621/1
General intelligence57−12.4#42/1781/3
Agents58.2−8.2#27/2983/6
Web research52.9−11#15/351/2
Knowledge work64.4−6.3#26/2082/2
Workflow automation48.1−17.9—/0✱✱
Safety53−8.4#100/3292/3
Secure code62.8−3.4#33/2701/1
Harm refusal55−6.5#73/2921/2
Fairness53.7−16.3#79/2921/2
Factual grounding43.8−26.9#85/1041/1
Jailbreak resistance54.3−10.7#125/2681/1
Toxicity avoidance46−13#225/2681/1
§ 2 · Head to head
What it beats, and what beats it
The same models turn up scope after scope. Counted once: where both were placed, who finished higher. Bars run right for GPT-5.6 Sol, left for the other.
§ 3 · Sources
Where the numbers come from
13 publications, 62 figures. Every one links to the page it was read from.
LMArena Text 1485
GDPval-AA 1618
AA-Briefcase 1479
Terminal-Bench 38.5% (Codex)
ARC-AGI-2 92.5%
LiveBench 1.0 81LiveBench 1.0 · Reasoning 91.7LiveBench 1.0 · Coding 83.9LiveBench 1.0 · Agentic Coding 56.2LiveBench 1.0 · Mathematics 96.2LiveBench 1.0 · Data Analysis 79.8LiveBench 1.0 · Language 87.7LiveBench 1.0 · Instruction Following 71.8
Kagi LLM Benchmark 67%
SimpleBench 64.8%
Vals · Legal Research Bench 48.08%Vals · LegalBench 86.97%Vals · Finance Agent 53.76%Vals · CorpFin 64.38%Vals · TaxEval 74.78%Vals · MortgageTax 67.29%Vals · MedCode 43.97%Vals · MedScribe 85.23%Vals · BioMysteryBench 71.11%Vals · CyberBench 76.31%Vals · SRE Bench 30.53%Vals · Web Search Index 49.08% (Keenable)Vals · SWE-bench Verified 96.2% (Mini-SWE-agent)Vals · Vibe Code Bench 80.5% (OpenHands)Vals · Code Migration 52.92%Vals · GPQA Diamond 95.2%Vals · MMLU Pro 89.1%Vals · ProofBench 83%
ParseBench · Overall 74.4%ParseBench · Tables 91.41%ParseBench · Charts 68.68%ParseBench · Content faithfulness 87.84%ParseBench · Semantic formatting 77.16%ParseBench · Visual grounding 46.9%
Enkrypt · Jailbreak risk 9.2%Enkrypt · Harmful content risk 0.6%Enkrypt · CBRN risk 19.2%Enkrypt · Toxicity risk 8.9%Enkrypt · Bias risk 72.1%Enkrypt · Insecure code risk 7.1%
Vectara · Factual consistency 87.6%
GPQA Diamond (Pass@1) 94.1%HLE (Pass@1) 44.5%Terminal-Bench 2.1 (Pass@1) 88.8%Terminal-Bench 3.0 (Pass@1) 34.4%Terminal-Bench 4.0 (Pass@1) 39.9%DeepSWE v1.1 (Resolved) 73%ProgramBench (Almost@1) 23%NL2Repo-Bench (Score) 56.8%CyberGym (Pass@1) 84.5%SEC-Bench Pro (Pass@1) 74.3%ExploitGym (Pass@1) 33.7%AutomationBench (Pass@1) 45.8%Agent's Last Exam (Pass@1) 26.7%Chartography w/ tools (Pass@1) 79.9%BabyVision w/ tools (Pass@1) 88.9%ZeroBench-main w/ tools (Pass@5) 53%
Badge
[](https://publicai.io/model-index/m/gpt-5-6-sol)