Loading page…
Preferred languages
More languages supported by Google Translate
Open the selector to load the current language catalog.
Translations are provided by Google Translate.
Loading page…
Preferred languages
More languages supported by Google Translate
Open the selector to load the current language catalog.
Translations are provided by Google Translate.
OpenAI · Proprietary · rank 19 · bench-align-v5
GPT-5.6 Sol is built for agent work. It ranks in the top quarter for agent work (top 3% of 127 scored models) and multimodal tasks (top 4% of 136 scored models), by BenchLM's published scores. It accepts text, images and PDFs and produces text.
GPT-5.6 Sol by OpenAI scores 73 overall, costs $4.00 per 1M input tokens and $20.00 per 1M output tokens, and carries a context window of 1.05M tokens — observed 2026-11-21.
Every published price component, including cache reads and writes.
Lifecycle events the source associates with this model, plus what this profile does not claim.
That is not evidence the model has no lifecycle plan — only that this source published none.
Seven axes from the ranking source. A missing axis is drawn as a gap.
Each axis shows the ranking source’s score and, beside it, BenchLM’s separate verified score for that axis. The two are published side by side and are not interchangeable.
Measured values with the date they were observed. Nothing is inferred from a sibling model or a provider claim.
Time to first token: latency to the first answer chunk (Artificial Analysis, via BenchLM). Reasoning models include thinking time, so values can run to tens or hundreds of seconds.
Evidence key: Observed
| Measurement | Value | Observed | Last good | Evidence |
|---|---|---|---|---|
| Time to first token | 100.66 s | 2026-11-21 | 2026-11-21 | |
| Throughput | 94 tok/s | 2026-11-21 | 2026-11-21 |
Regional or per-endpoint measurements appear only when the API supplies them; none are modelled here.
| Batch output / 1M tokens (OpenRouter batch) (base rate) | $10.00 |
| Batch cache read / 1M tokens (OpenRouter batch) (base rate) | $0.20 |
| Batch cache write / 1M tokens (OpenRouter batch) (base rate) | $2.50 |
Batch API rate (asynchronous jobs), from OpenRouter's batch listing.
The standard input, output and cache rates above are the base rate. models.dev publishes higher rates for requests whose prompt exceeds the thresholds below. The batch rates have their own long-context tiers, from OpenRouter's batch listing, shown separately. The cost simulator prices a standard-rate request whose whole prompt is above a threshold entirely at the highest of these tiers it exceeds, and a batch request only at a batch tier; it never applies these tiers to batch rates. The cost example on this page uses the base rate: its 1,200-token prompt is below every threshold.
The batch rates above are the base batch rate. OpenRouter's batch listing, the same listing as those rates, publishes higher batch rates for requests whose total prompt is above the thresholds below. The cost simulator prices a batch request whose whole prompt is above a threshold entirely at the highest batch tier it exceeds.
10 conversations per day × 8 messages × 22 active days, 1200 input and 400 output tokens per message, no cache. This uses the same calculator as the cost simulator, so an unavailable applicable rate makes the example unavailable too.
Plans from the model's maker, and the API providers that list this exact model today. Links leave ModelScale; each provider sets its own prices and limits.
| Benchmark | Value | Rank | Tasks | Difficulty | Provenance |
|---|---|---|---|---|---|
| GPQA Graduate-Level Google-Proof Q&A | 94.6 | #6 | 448 questions | Graduate level | GPQA: A Graduate-Level Google-Proof Q&A Benchmark |
| GPQA-D GPQA Diamond | 94.6 | #6 | Graduate-level science questions | Graduate level | GPQA: A Graduate-Level Google-Proof Q&A Benchmark |
| HLE-Verified HLE-Verified | 54.5 | #2 | 1,811 verified or revised expert questions | Frontier multidisciplinary expert reasoning | HLE-Verified: A Systematic Verification and Structured Revision of Humanity's Last Exam |
| LABBench2 LABBench2: An Improved Benchmark for AI Systems Performing Biology Research | 82.1 | #4 | Nearly 1,900 biology-research tasks | Real-world biology research | LABBench2: An Improved Benchmark for AI Systems Performing Biology Research |
| Artificial Analysis Intelligence Index Artificial Analysis Intelligence Index | 58.9 | #6 | Cross-benchmark intelligence index | Display-only external reference | Artificial Analysis |
| AA-GPQA Diamond Artificial Analysis GPQA Diamond | 94.1 | #5 | Graduate-level science questions | Graduate-level science reasoning | Artificial Analysis GPQA Diamond Benchmark Leaderboard |
| AA-HLE Artificial Analysis Humanity's Last Exam | 49.5 | #13 | Expert-level questions | Frontier expert reasoning | Artificial Analysis Humanity's Last Exam Benchmark Leaderboard |
| AA-Omniscience Index Artificial Analysis Omniscience Index | 22.0 | #35 | Knowledge questions | Broad factual knowledge | AA-Omniscience: Knowledge and Hallucination Benchmark |
| AA-Omniscience Accuracy Artificial Analysis Omniscience Accuracy | 59.4 | #10 | Knowledge questions | Broad knowledge | Artificial Analysis model benchmarks |
| AA-Omniscience Hallucination Rate Artificial Analysis Omniscience Hallucination Rate | 92.2 | #23 | Knowledge questions | Factuality | Artificial Analysis model benchmarks |
| HealthBench Hard HealthBench Hard | 33.1 | #5 | 1,000 health prompts | Advanced health reasoning | Muse Spark Eval Methodology |
| HealthBench Professional HealthBench Professional | 60.5 | #12 | Clinician chat tasks | Professional clinical workflows | HealthBench Professional: Evaluating Large Language Models on Real Clinician Chats |
| GPQA Diamond (Vals) GPQA Diamond, Vals AI run | 95.2 | #2 | Graduate-level science questions | Expert reasoning | Vals AI GPQA Diamond, Vals AI run leaderboard |
| MMLU-Pro (Vals) MMLU-Pro, Vals AI run | 89.1 | #15 | Academic multiple-choice questions | Broad academic knowledge | Vals AI MMLU-Pro, Vals AI run leaderboard |
| Benchmark | Value | Rank | Tasks | Difficulty | Provenance |
|---|---|---|---|---|---|
| Terminal-Bench 2.1 Terminal-Bench 2.1 | 91.9 | #2 | Terminal-based software-agent tasks | Professional software engineering | Terminal-Bench |
| Terminal-Bench 2.0 Terminal-Bench 2.0 | 91.9 | #1 | Terminal-based software tasks | Professional software engineering | Terminal-Bench 2.0 |
| SWE-bench Pro SWE-bench Pro | 64.6 |
| Benchmark | Value | Rank | Tasks | Difficulty | Provenance |
|---|---|---|---|---|---|
| FrontierMath (legacy) FrontierMath legacy aggregate | 89 | #1 | Historical aggregate | Research-level mathematics | FrontierMath: A Benchmark for Evaluating Advanced Mathematical Reasoning in AI |
| FrontierMath v2 (Tiers 1-3) FrontierMath v2 Tiers 1-3 | 89.000 | #10 | 285 private advanced mathematics problems | From olympiad-plus to early research mathematics | FrontierMath v2 benchmark hub |
| FrontierMath v2 (Tier 4) FrontierMath v2 Tier 4 | 83.000 |
| Benchmark | Value | Rank | Tasks | Difficulty | Provenance |
|---|---|---|---|---|---|
| ARC-AGI-1 ARC-AGI-1 Semi-Private Evaluation | 96.50 | #11 | Semi-private ARC-AGI-1 evaluation set | Abstract visual reasoning | ARC Prize leaderboard |
| ARC-AGI-2 Abstraction and Reasoning Corpus for AGI v2 | 92.5 | #4 | Visual pattern completion and abstract reasoning | Expert-level — hardest public reasoning benchmark | ARC-AGI-2: A Harder General Intelligence Benchmark |
| ARC-AGI-3 Abstraction and Reasoning Corpus for AGI v3 | 7.78 |
| Benchmark | Value | Rank | Tasks | Difficulty | Provenance |
|---|---|---|---|---|---|
| AA-IFBench Artificial Analysis IFBench | 72.7 | #36 | Verifiable instruction constraints | Instruction precision | Artificial Analysis IFBench Benchmark Leaderboard |
| Gray Swan IPI (15 attempts) Gray Swan Indirect Prompt Injection, attack success at 15 attempts | 27.0 | #4 | Indirect prompt-injection attacks | Adversarial prompt injection | Gemini 4 Argon evaluation methodology |
| Benchmark | Value | Rank | Tasks | Difficulty | Provenance |
|---|---|---|---|---|---|
| Terminal-Bench 3.0 Terminal-Bench 3.0 | 34.6 | #2 | 74 professional computer-work tasks across 7 domains | Frontier autonomous knowledge work | Terminal-Bench 3.0 |
| AA Briefcase Artificial Analysis Briefcase | 1487 | #28 | Professional knowledge-work tasks | Professional work | Artificial Analysis Briefcase Benchmark Leaderboard |
| AA AutomationBench Artificial Analysis AutomationBench | 60.1 |
| Benchmark | Value | Rank | Tasks | Difficulty | Provenance |
|---|---|---|---|---|---|
| MMMU-Pro Massive Multi-discipline Multimodal Understanding Pro | 83 | #5 | Multimodal academic reasoning | Frontier multimodal | MMMU-Pro: A More Robust Multi-discipline Multimodal Understanding Benchmark |
| AA-MMMU-Pro Artificial Analysis MMMU-Pro | 83.4 | #11 | Multimodal academic reasoning | Frontier multimodal | Artificial Analysis MMMU-Pro Benchmark Leaderboard |
| MMMU-Pro w/ Python MMMU-Pro with Python | 84.6 |
| Benchmark | Value | Rank | Tasks | Difficulty | Provenance |
|---|---|---|---|---|---|
| ExploitBench ExploitBench v8-bench | 74 | #10 | V8 exploit synthesis runs | Browser exploitation and cybersecurity | ExploitBench |
| The Last Ones completion The Last Ones Cyber Range Completion Rate | 70 | #1 | 10 long-horizon runs | Long-horizon cyber operations | UK AISI and CAISI Preliminary Assessment of Kimi K3's Cyber Capabilities |
| ACCR standard Advanced Cyber Completion Rate — Standard Access |
A second, independent score source. Epoch’s own figures in Epoch’s units, one line per model version Epoch evaluated (effort setting as Epoch labels it). They are not part of this site’s capability scores or rank.
| Benchmark | Variant | Score | Epoch date | Evidence |
|---|---|---|---|---|
| GPQA Diamond Accuracy on graduate-level science multiple-choice questions (best score across scorers). Unit: fraction (0–1); higher is better. | low | 0.899 ± 0.0215 | 2026-08-07 | |
| max | 0.935 ± 0.0157 | 2026-07-09 | ||
| none | 0.8283 ± 0.0269 | 2026-08-07 | ||
| FrontierMath Tiers 1-3 (v2) Share of private research-level math problems solved, tiers 1-3. Unit: fraction (0–1); higher is better. | max | 0.8912 ± 0.0185 | 2026-07-09 | |
| FrontierMath Tier 4 (v2) Share of private research-level math problems solved, tier 4. Unit: fraction (0–1); higher is better. | max | 0.8293 ± 0.0595 | 2026-07-09 | |
| promax | 0.8049 ± 0.0627 | 2026-07-09 | ||
| SimpleQA Verified Short factual questions answered correctly. Unit: fraction (0–1); higher is better. | max | 0.697 ± 0.0145 | 2026-08-10 | |
| OSWorld 2.0 Binary task success on computer-use tasks, as compiled by Epoch from the OSWorld 2.0 leaderboard. Unit: fraction (0–1); higher is better. | max batch tool | 0.2734 | — |
Source: Epoch AI, CC BY 4.0 (licence), benchmark data read once a day.
People's votes between two anonymous answers on arena.ai, one board per task. LMArena's own ratings and ranks, one line per model version it rates; not part of this site's capability scores or rank.
| Board | Variant | Rating | Votes | Rank | Published | Evidence |
|---|---|---|---|---|---|---|
| Vision Arena People's votes between two answers about the same image, style-controlled. | xHigh | 1,288 ± 7 | 12,244 | 17 | 2026-11-19 |
Source: LMArena, CC BY 4.0 (licence), leaderboard data read once a day.
| #18 |
| 1,865 repository problems |
| Long-horizon professional engineering |
| SWE-Bench Pro: Can AI Agents Solve Long-Horizon Software Engineering Tasks? |
| VulcanBench v3 VulcanBench v3 | 87.0 | #4 | 23 post-cutoff repository tasks in the v3 report | Professional multi-file software engineering | VulcanBench |
| VulcanBench CII v1 VulcanBench Coding Intelligence Index v1 | 86.5 | #3 | 38 validated post-cutoff repository tasks | Mid-band frontier software engineering | CII v1 frontier results |
| FrontierCode 1.1 Extended FrontierCode 1.1 Extended | 60.6 | #5 | 150 private software-engineering tasks | Frontier coding-agent quality | FrontierCode leaderboard |
| FrontierSWE v2 FrontierSWE v2 | 32.2 | #12 | 34 ultra-long-horizon engineering and research tasks | Ultra-long-horizon frontier software engineering | FrontierSWE v2 |
| Bug Hunt Bench Bug Hunt Bench | 42.0 | #9 | 105 planted bugs across two production TypeScript repositories | Blind production-repository bug finding and repair | Bug Hunt Bench method, caveats, and definitions |
| AA Coding Index Artificial Analysis Coding Index | 77.4 | #3 | Cross-benchmark coding index | Display-only external reference | Artificial Analysis model leaderboards |
| AA-SciCode Artificial Analysis SciCode | 57.1 | #22 | Scientific coding subproblems | Scientific programming | Artificial Analysis SciCode Benchmark Leaderboard |
| LiveCodeBench (Vals) LiveCodeBench, Vals AI run | 82.6 | #49 | Competitive programming problems (easy, medium, hard) | Frontier coding | Vals AI LiveCodeBench, Vals AI run leaderboard |
| SWE-bench (Vals) SWE-bench, Vals AI run | 96.2 | #4 | Real repository issues by human time bucket | Frontier coding agents | Vals AI SWE-bench, Vals AI run leaderboard |
| DeepSWE DeepSWE | 72.7 | #18 | 113 software engineering tasks across 91 repositories and 5 languages | Long-horizon software engineering | DeepSWE benchmark blog |
| PostTrainBench v1.1 PostTrainBench v1.1 | 36.2 | #6 | Post-training Qwen3 1.7B and 4B, SmolLM3 3B, and Gemma3 4B | Frontier agent evaluation | PostTrainBench v1.1 leaderboard and methodology |
| VulcanBench Frontier v4 VulcanBench Frontier v4 | 87.2 | #13 | 23 behavioural-reconstruction coding tasks | Professional multi-file software engineering | VulcanBench Frontier v4 leaderboard |
| FrontierCode 2.0 FrontierCode 2.0 | 28.9 | #14 | 132 private software-engineering tasks | Frontier coding-agent quality | FrontierCode 2.0 |
| PostTrainBench PostTrainBench v1.2 | 31.0 | #13 | Post-training four base models on six benchmarks | Autonomous ML engineering | PostTrainBench |
| #12 |
| 41 private extreme-difficulty mathematics problems |
| Research-level mathematics requiring hours or days of expert work |
| FrontierMath Tier 4 v2 leaderboard |
| #10 |
| Interactive game-like tasks with hidden rules |
| Frontier agentic reasoning |
| ARC-AGI-3: A New Challenge for Frontier Agentic Intelligence |
| GeneBench-Pro GeneBench-Pro | 28.7 | #2 | 129 genomics statistical-analysis workflows | Long-horizon scientific reasoning | GeneBench-Pro: Evaluating Multistage Statistical Reasoning |
| AA-LCR Artificial Analysis Long Context Reasoning | 84.0 | #8 | Long-context reasoning tasks | Long-context reasoning | Artificial Analysis model benchmarks |
| MLCR-AA Medical Long Context Reasoning (MLCR-AA) | 26.1 | #18 | Long, fragmented medical-record reasoning | Long-context medical reasoning | Medical Long Context Reasoning (MLCR-AA) |
| CritPt Critical Physics Tasks | 32.3 | #3 | Research-level physics questions | Research-level physics reasoning | CritPt Benchmark Leaderboard |
| BioBasicsBench BioBasicsBench: everyday wet-lab science | 45.6 | #3 | 100 routine wet-lab analysis tasks (88 multimodal) across five domains | Routine wet-lab data interpretation | Measuring how AI agents handle everyday wet-lab science |
| #23 |
| Business-process automation tasks |
| Agentic automation |
| Artificial Analysis AutomationBench Benchmark Leaderboard |
| AA EnterpriseOps-Gym Artificial Analysis EnterpriseOps-Gym | 42.9 | #17 | Enterprise operations workflows | Enterprise agent operations | Artificial Analysis EnterpriseOps-Gym Benchmark Leaderboard |
| AA Harvey LAB v1.0 Artificial Analysis Harvey LAB-AA v1.0 | 87.2 | #13 | Legal agent tasks | Professional legal work | Artificial Analysis Harvey LAB-AA Benchmark Leaderboard |
| AA ITBench Artificial Analysis ITBench-AA | 56.2 | #1 | IT incident-response tasks | Enterprise IT operations | Artificial Analysis ITBench-AA Benchmark Leaderboard |
| AA Tau3 Banking Artificial Analysis Tau3-Banking | 44.3 | #10 | Banking tool-use workflows | Agentic banking workflows | Artificial Analysis Tau3-Banking Benchmark Leaderboard |
| Terminal-Bench 2.0 Terminal-Bench 2.0 | 91.9 | #1 | Terminal-based software tasks | Professional software engineering | Terminal-Bench 2.0 |
| Terminal-Bench 2.1 Terminal-Bench 2.1 | 91.9 | #2 | Terminal-based software-agent tasks | Professional software engineering | Terminal-Bench |
| BrowseComp BrowseComp | 92.2 | #2 | Research questions requiring browsing | Hard web research | BrowseComp |
| GDPval-AA GDPval-AA | 1735 | #14 | Agentic real-world work tasks | Professional agentic workflows | DeepSeek-V4 Technical Report |
| GDPval-AA GDPval-AA normalized | 55.4 | #19 | Economically valuable tasks | Professional agentic workflows | Artificial Analysis model benchmarks |
| AA Agentic Index Artificial Analysis Agentic Index | 50.5 | #9 | Cross-benchmark agentic index | Display-only external reference | Artificial Analysis model leaderboards |
| AA Terminal-Bench 4.0 Artificial Analysis Terminal-Bench v4.0 | 39.9 | #23 | Terminal-based agent tasks | Agentic software engineering | Artificial Analysis Terminal-Bench v4.0 Benchmark Leaderboard |
| GDP.pdf Artificial Analysis GDP.pdf | 27.2 | #12 | Professional document-production tasks | Professional knowledge work | GDP.pdf Benchmark Leaderboard |
| AA-AnalystAgent Artificial Analysis AnalystAgent | 47.5 | #16 | Spreadsheet and document analysis questions | Business and data analysis | AA-AnalystAgent Benchmark Leaderboard |
| OSWorld 2.0 OSWorld 2.0 | 62.6 | #6 | 108 long-horizon computer-use workflows | Long-horizon professional workflows | OSWorld2.0: Benchmarking Computer Use Agents on Long-Horizon Real-World Tasks |
| CyberGym CyberGym | 84.5 | #9 | 1,507 vulnerability analysis instances | Real-world cybersecurity | CyberGym: Evaluating AI Agents' Real-World Cybersecurity Capabilities at Scale |
| ExploitGym ExploitGym | 33.7 | #7 | 898 exploitation tasks | Advanced cybersecurity exploitation | ExploitGym: Can AI Agents Turn Security Vulnerabilities into Real Attacks? |
| Toolathlon Toolathlon | 58 | #3 | Multi-tool workflows | Advanced tool use | Introducing GPT-5.4 mini and nano |
| τ²-bench results τ²-Bench Tool-Agent-User Evaluation | 85.1 | #58 | Airline, retail, and telecom customer-service task sets | Dual-control customer-service workflows | τ²-Bench: Evaluating Conversational Agents in a Dual-Control Environment |
| Terminal-Bench 2.1 (Vals) Terminal-Bench 2.1, Vals AI run | 85.8 | #3 | Difficult terminal tasks | Frontier agentic | Vals AI Terminal-Bench 2.1, Vals AI run leaderboard |
| ApprenticeBench ApprenticeBench: end-to-end computer use, continual learning, and long-horizon agency on a real accounts-payable job | 26 | #5 | 100 vendor bills processed in sequence inside a simulated construction company | Long-horizon computer use with offline and online continual learning | ApprenticeBench: a step change in AI's job readiness |
| MazeBench MazeBench: visual spatial reasoning in a 3D open world | 13 | #10 | 100 hidden gems across 256 rooms in one open world | Long-horizon spatial planning, with single puzzles that can take over one hundred moves | MazeBench Results |
| Terminal-Bench 4.1 Terminal-Bench 4.1 | 38.50 | #21 | 66 professional computer-work tasks | Frontier autonomous knowledge work | Terminal-Bench 4.1 |
| LifeVal LifeVal: everyday life tasks | 80.0 | #4 | 169 everyday life tasks (caregiving, health, education, household administration, personal finance, shopping, government services) | Agentic personal-assistant work | LifeVal |
| #2 |
| Multimodal academic reasoning |
| Frontier multimodal |
| Introducing GPT-5.4 mini and nano |
| ParseBench-1 ParseBench-1: document parsing for AI agents (LlamaIndex leaderboard, first release) | 74.40 | #3 | 2,078 human-verified enterprise pages from insurance, finance and government documents | No method is strong on all five dimensions; charts and visual grounding separate the field | ParseBench: A Document Parsing Benchmark for AI Agents |
| ExtractBench (LlamaIndex) ExtractBench: schema-guided enterprise document extraction (LlamaIndex leaderboard) | 93.77 | #5 | 370 documents, 4,869 pages, 8 business domains, 67 document types, each with a JSON Schema | Short documents are near solved; long record lists and grounding separate systems | ExtractBench: A Benchmark for Schema-Guided Enterprise Document Extraction |
| #2 |
| Internal advanced-cyber request set |
| Cyber access and safeguards |
| Expanding Daybreak as the cyber defense window narrows |
| ACCR Daybreak Blue Advanced Cyber Completion Rate — Daybreak Blue | 2.0 | #2 | Internal advanced-cyber request set | Cyber access and safeguards | Expanding Daybreak as the cyber defense window narrows |
| SEC-Bench Pro SEC-Bench Pro | 71.2 | #9 | Security engineering tasks | Advanced cybersecurity | Introducing GPT-5.6 |
| FrontierCyber FrontierCyber | 9.6 | #3 | 197 dynamic cyber tasks | Easy through elite cyber operations | FrontierCyber |
| CyScenarioBench success CyScenarioBench Average Success Rate | 28 | #6 | 11 cyber scenarios | Long-horizon cybersecurity | Assessing GPT-5.6 Sol |
| CyScenarioBench solved CyScenarioBench Scenarios Ever Solved | 7 | #3 | 11 cyber scenarios | Long-horizon cybersecurity | Assessing GPT-5.6 Sol |
| Atomic network attacks Atomic Network Attack Simulation | 98 | #3 | Atomic cyber tasks | Network attack simulation | Assessing GPT-5.6 Sol |
| Atomic vulnerability research Atomic Vulnerability Research and Exploitation | 91 | #3 | Atomic cyber tasks | Vulnerability research and exploitation | Assessing GPT-5.6 Sol |
| Atomic evasion Atomic Evasion | 56 | #1 | Atomic cyber tasks | Cybersecurity evasion | Assessing GPT-5.6 Sol |
| MaxxBench MaxxBench (mutated DeepSWE tasks) | 69.1 | #3 | 89 mutated DeepSWE feature-implementation tasks, two attempts per model | Long-horizon software engineering under changed requirements | Why MaxxBench (unLAP research post) |
| Chess Puzzles Chess Puzzles (Epoch AI) | 59.0 | #11 | 100 programmatically generated chess positions with a single best move | Novel chess positions, single best move | Chess Puzzles (Epoch AI) |
| Mystery Game Puzzles Mystery Game Puzzles (Epoch AI) | 58.0 | #12 | 100 single-best-move positions from an undisclosed game | Unfamiliar game positions, single best move | Mystery Game Puzzles (Epoch AI) |
| Furniture Assembly Furniture Assembly Benchmark (Epoch AI) | 56.7 | #14 | 60 photos of three IKEA builds, some with deliberate assembly mistakes | Spatial reasoning from technical diagrams to real photos | Furniture Assembly Benchmark (Epoch AI) |
| EBR-bench Earthborne Rangers (EBR-bench) | 44.8 | #13 | Ten playthroughs of an Earthborne Rangers campaign segment per sample (eight learning, two scored) | Learning an unfamiliar game over multi-hour playthroughs | Earthborne Rangers (EBR-bench), Epoch AI |
| MirrorCode MirrorCode (Epoch AI and METR) | 20.0 | #38 | 30 reimplementation tasks (15 programs in two languages), three attempts each | Days-long software engineering without source code | MirrorCode (Epoch AI) |
| Clocktower Bench Clocktower Bench (Blood on the Clocktower) | 1591 | #4 | 386 rated head-to-head games of Blood on the Clocktower (10 to 12 players) | Social deduction, deception and multi-agent reasoning | Clocktower Bench leaderboard |
| ExploitBench Internal Port ExploitBench — Internal Port (June–August 2026) | 5.5 | #7 | Recently disclosed V8 and Chrome vulnerabilities (June–August 2026) | Browser exploitation | GPT-6 Cyber System Card |
| ExploitGym Honeypot ExploitGym Honeypot out-of-scope exploitation rate | 49.50 | #1 | Hard ExploitGym tasks with an out-of-scope honeypot service | Cyber-agent scope adherence | GPT-6 Cyber: Advancing the frontier of aligned cybersecurity agents |
| FACTS Multimodal FACTS Multimodal | 46.4 | #13 | About 1,500 image questions: 711 public and 811 held out | Image-grounded factual answering | BenchLM |
| DAYJOB DAYJOB | 4.2 | #9 | 130 long-horizon tasks: 80 finance and 50 healthcare | Long-horizon professional work | BenchLM |
| DAYJOB Healthcare DAYJOB Healthcare | 1.2 | #10 | 50 long-horizon healthcare tasks across five subfields | Long-horizon clinical and healthcare operations work | BenchLM |
| Code2Video Code2Video | 1548 | #4 | Creative briefs rendered as motion-graphics video | Code-driven motion design | BenchLM |
| Joke Arena Joke Arena | 1502 | #8 | 100 joke prompts in 10 categories, rated by human votes | Humour writing | BenchLM |
| Pelican Arena Pelican Arena | 1635 | #9 | One prompt: an animated SVG of a pelican riding a bicycle, rated by human votes | Vector drawing and animation in code | BenchLM |
| Game Arena Go Game Arena Go | 1005 | #18 | Head-to-head Go games on a 13 x 13 board | Strategic planning and rule-following | BenchLM |
| BabaIsBench BabaIsBench (Baba Is You intro levels, pass@1) | 100 | #1 | 8 intro levels of Baba Is You, three attempts per level (Stage 0); 15 further levels run once (Stage 1) | Rule-rewriting puzzles that need planning over many moves | Baba Is Solved by Fable 5 and GPT-5.6 Sol, but at what cost? (Quesma) |
| NYT Connections NYT Connections LLM Benchmark (extended version) | 93.8 | #12 | 940 NYT Connections puzzles, each with up to four added trick words | Wordplay, word knowledge and decoys that fit almost-but-not-quite | NYT Connections LLM Benchmark (Lech Mazur, GitHub) |
| StealthMark StealthMark Intelligence Benchmarks (overall score) | 53.7 | #12 | 8 original tasks: Newton's Mirror, Ground Truth, Latch, Foresight, Mona Lisa, Selfie, Beach and Raccoon | Chained physical reasoning, planning without tools and precise visual construction | StealthMark: The AI Benchmark Explained (Stealth Models) |
| FrontierPhysics End-to-End Physics Research Benchmark for AI Agents | 7.1 | #9 | 56 research projects from 42 physicists across nine fields, three trials each (168 rollouts per setup) | Open-ended research projects that took their authors about 270 days on average | FrontierPhysics: Evaluating Agents for End-to-End Physics Research (arXiv, November 4, 2026) |