‹ PublicAI Index
The LLM benchmark aggregator.
GPT-5.6 Terra
OpenAI
Strongest in Data analysis (#11 of 31), weakest in Toxicity avoidance (#212 of 268). Above par in 35 of 40 scopes. Among the models it meets almost everywhere, it finishes behind Claude Fable 5.1 and Claude Opus 5.5 and ahead of Claude Sonnet 4.5 and MiMo V2.6 Pro.
§ 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.
Coding57.7−8.5#20/1794/6
Agentic coding58−8.9#19/1734/5
Code generation61.7−3.6—/20✱0/1
Reasoning57.3−8.1#28/1955/5
Science60.3−3.2#24/1221/1
Mathematics57.4−6.4#26/1522/2
Reasoning56.1−12.9#42/1473/3
Expert reasoning60.9−5.5—/0✱✱
Knowledge55.5−7#36/1972/3
Document parsing55.5−7.8#33/831/1
Academic knowledge56.8−7.1#39/1231/1
Factuality52.5−8.3—/0✱✱
Safety55.8−5.6#38/3291/3
Secure code64.1−2.1#18/2701/1
Fairness61.1−8.9#37/2921/2
Harm refusal54.4−7.1#80/2921/2
Jailbreak resistance55.1−9.9#114/2681/1
Toxicity avoidance47.9−11.1#212/2681/1
Professional54.4−8#39/1942/2
IT operations46.3−27.5#15/331/1
Finance57.9−4.5#16/1621/1
Cybersecurity56.4−7.5#18/471/1
Document design46.7−22#22/311/1
Legal56.4−9.5#30/1731/1
Medical52.6−13.9#67/1491/1
Human preference63.2−5.6#43/3541/1
Human preference63.2−5.6#43/3541/1
Agents54−12.4#71/2983/6
Long-horizon work43.7−30.3#18/311/1
Web research51−12.9#20/351/2
Computer use54.4−13.7#25/681/2
Knowledge work61.1−9.6#41/2082/2
Tool use57.8−16.2—/81✱0/1
Workflow automation56.3−9.7—/0✱✱
Core abilities51.7−13.6#80/1993/4
Data analysis57.6−7.6#11/311/1
Language56.4−10.2#19/621/1
Instruction following44−29.8#43/621/1
General intelligence50.3−19.1#84/1782/3
Long context56−8.6—/0✱✱
§ 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 Terra, left for the other.
§ 3 · Sources
Where the numbers come from
13 publications, 66 figures. Every one links to the page it was read from.
LMArena Text 1468
GDPval-AA 1461
AA-Briefcase 1333
Terminal-Bench 21.5% (Codex)
ARC-AGI-2 83.9%
LiveBench 67.2LiveBench · Agentic 25.9LiveBench · Reasoning 90.6LiveBench · Coding 84.2LiveBench · Browser Use 70.3LiveBench · Data Analysis 79.3LiveBench · Research 79.7LiveBench · Design 70.9
LiveBench 1.0 77.9LiveBench 1.0 · Reasoning 90.6LiveBench 1.0 · Coding 78.2LiveBench 1.0 · Agentic Coding 54.9LiveBench 1.0 · Mathematics 94.9LiveBench 1.0 · Data Analysis 79.3LiveBench 1.0 · Language 82.9LiveBench 1.0 · Instruction Following 64.6
Kagi LLM Benchmark 51.3%
SimpleBench 48.9%
Vals · Legal Research Bench 41.35%Vals · LegalBench 85.11%Vals · Finance Agent 54.44%Vals · CorpFin 65.31%Vals · TaxEval 76.17%Vals · MortgageTax 67.33%Vals · MedCode 43.41%Vals · MedScribe 82.87%Vals · CyberBench 72.08%Vals · SRE Bench 9.54%Vals · SWE-bench Verified 95.4% (Mini-SWE-agent)Vals · Vibe Code Bench 74.59% (OpenHands)Vals · Code Migration 47.8%Vals · GPQA Diamond 90.91%Vals · MMLU Pro 86.66%Vals · ProofBench 74%
ParseBench · Overall 67.06%ParseBench · Tables 86.74%ParseBench · Charts 67.86%ParseBench · Content faithfulness 86.77%ParseBench · Semantic formatting 68.47%ParseBench · Visual grounding 25.45%
Enkrypt · Jailbreak risk 8.5%Enkrypt · Harmful content risk 2.2%Enkrypt · CBRN risk 20%Enkrypt · Toxicity risk 7.6%Enkrypt · Bias risk 60.7%Enkrypt · Insecure code risk 4.4%
GDPVal-AA 1503tau3-Banking 28.7%Toolathlon Verified 64.8%Automation Bench Public 28%Apex-Agents (pass@1) 25.4%MCPMark 74%WildClawBench 60%Terminal-Bench 2.1 75.7%SciCode 50.1%Humanity's Last Exam (without tools) 38.5%GPQA Diamond 89.6%CritPt 22.9%AA-LCR 73.3%AA-Omniscience Accuracy 45%AA-Omniscience Non-Hallucination 10%
Badge
[](https://publicai.io/model-index/m/gpt-5-6-terra)