BenchmarksEvery score sourcedPrices verified
AI benchmarks35
Leaderboards for the tests AI labs use to rank their models, each score linked to the page that published it. Where a benchmark runs its own leaderboard we take its number; otherwise the lab's own report, marked as such. Next to every score: the model's current API price, read on the vendor's page. As of November 8, 2026.
Benchmarks
35
Systems ranked
116
Published scores
655
Independent
262
Sources
25
Newest score
Nov 5
§ 01 / LLM LEADERBOARD
Top models across benchmarks
The models with scores on the most benchmarks, against the 8 benchmarks with the most models. Each cell is the model's score on that benchmark's main leaderboard; the leader of each column is highlighted. Blank means no source has published a score for that model on that test.
| Model | ARC-AGI-2 | ARC-AGI-1 | TB 4.0 | DeepSWE | ARC-AGI-3 | AutomationBench | HLE | TB 2 | Price / 1M |
|---|---|---|---|---|---|---|---|---|---|
| DeepSeek V4.1 Flash DeepSeek | 78.1% | 94.5% | 42.3% | 77% | 0.2% | 60.5% | 63.9% | 90.6% | $0.3 / $1.2 |
| GPT-6 Astra OpenAI | 95% | 98.5% | 58.2% | 74.1% | 62.7% | 41.4% | 57.2% | $10 / $50 | |
| DeepSeek V4 Pro DeepSeek | 61.3% | 90.5% | 12.4% | 62.7% | 43.2% | 60% | 87.9% | $1.32 / $3.96 | |
| DeepSeek V4 Flash DeepSeek | 61.4% | 89% | 7% | 54.4% | 37.7% | 51.5% | 82.7% | $0.44 / $1.32 | |
| Gemini 4 Argon Google DeepMind | 91.3% | 98.5% | 57.4% | 77.9% | 49.8% | 51.3% | $2 / $10 | ||
| GPT-6.1 Sol OpenAI | 94.2% | 98.5% | 58.2% | 75.2% | 52.7% | 36.1% | $2 / $10 | ||
| Claude Opus 5 Anthropic | 90.4% | 97.5% | 53.9% | 30.2% | 26.9% | 63.6% | $5 / $25 | ||
| Gemini 3.8 Flash Google DeepMind | 89.2% | 98.5% | 19.1% | 73.7% | 10.4% | 89.4% | $0.75 / $3.75 | ||
| GPT-6 Sol OpenAI | 89.6% | 95.5% | 49.4% | 68.8% | 4.6% | 33.2% | $2 / $10 | ||
| GPT-5.6 Sol OpenAI | 92.5% | 97.5% | 37.3% | 72.7% | 7.8% | 28.8% | $4 / $20 | ||
| GLM-5.2 Zhipu (Z.ai / GLM) | 22.8% | 77% | 46.2% | 26.2% | 54.7% | 81% | $1.4 / $4.4 | ||
| Gemini 3.1 Pro Preview Google | 77.1% | 98% | 12% | 0.4% | 51.4% | 73.8% | $2 / $12 | ||
| GPT-6 Luna OpenAI | 59.3% | 86.7% | 16.4% | 66.6% | 0.2% | 20.7% | $0.1 / $0.5 | ||
| GPT-5.6 Luna OpenAI | 59.6% | 90.7% | 17.3% | 62.2% | 0.2% | 17% | $0.2 / $1.2 | ||
| Claude Opus 5.5 Anthropic | 93.3% | 98.5% | 64.8% | 40% | 67.7% | $4 / $20 | |||
| Claude Fable 5.1 Anthropic | 90% | 97.5% | 57.9% | 31.4% | 65.6% | $10 / $50 | |||
| Kimi K3 Moonshot (Kimi) | 60.4% | 94.5% | 67.5% | 56% | 88.3% | $3 / $15 | |||
| GLM-5.3 Zhipu (Z.ai / GLM) | 41.8% | 66.9% | 48.2% | 62.5% | 88.2% | $1.4 / $4.4 | |||
| Claude Sonnet 5.5 Anthropic | 83.9% | 96.5% | 61.8% | 24.3% | 64.5% | $2 / $10 | |||
| Gemini 3.7 Flash Google DeepMind | 84.6% | 95.5% | 11.2% | 65.3% | 85.8% | $0.75 / $3.75 | |||
| Qwen3.8 Max Alibaba (Qwen) | 27% | 56.6% | 27.3% | 56.2% | 86.6% | $2 / $6 | |||
| Grok 4.7 SpaceXSI | 61.4% | 90.2% | 37.6% | 71% | 1.8% | $2 / $6 | |||
| Claude Haiku 5.5 Anthropic | 61.3% | 89% | 39.2% | 0.5% | 57.4% | $0.1 / $0.5 | |||
| Grok 4.6 SpaceXSI | 67.1% | 87.5% | 20.3% | 65.9% | 2.1% | $2 / $6 | |||
| Grok 4.5 SpaceXSI | 52.6% | 87.2% | 12.4% | 54% | 0.3% | $2 / $6 | |||
| MiMo-V2.6-Pro Xiaomi MiMo | 34.9% | 71.9% | 53.1% | 89.9% | $0.435 / $0.87 | ||||
| MiMo-V2.6-Flash Xiaomi MiMo | 28.8% | 67.9% | 52.3% | 87.6% | $0.14 / $0.28 | ||||
| Gemini 3.6 Flash Google | 60.4% | 91.2% | 48.6% | 78% | $0.75 / $3.75 | ||||
| Gemini 3.5 Flash Google | 72.1% | 92.5% | 37% | 76.2% | $1.5 / $9 | ||||
| GPT-5.6 Terra OpenAI | 83.9% | 96.5% | 21.5% | 0.8% | $2 / $12 |
§ 02 / CODING
Coding benchmarks
| Benchmark | Leader | Top score | Ranked | Independent | Newest score |
|---|---|---|---|---|---|
| DeepSWE DeepSWE tests coding agents on new features and behaviour changes in real open-source repositories, written as short behaviour-focused requests that usually touch several files. | Gemini 4 ArgonGoogle DeepMind | 77.9% | 27 | vendor-only | 2026-11-02 |
| FrontierCode FrontierCode judges coding agents the way an open-source maintainer would: on whether the change they submit would be merged, not only whether tests pass. | Claude Opus 5.5Anthropic | 54.4% | 15 | vendor-only | 2026-10-07 |
| NL2Repo-Bench NL2Repo-Bench tests whether a coding agent can build a complete, installable Python library from scratch, given nothing but a natural-language requirements document. | DeepSeek V4.1 FlashDeepSeek | 68.2% | 8 | vendor-only | 2026-10-19 |
| SWE-Bench Pro SWE-Bench Pro tests coding agents on long, real software tasks in professional repositories, including private and copyleft code chosen to be hard to find in training data. | Gemini 3.6 FlashGoogle | 58.7% | 8 | vendor-only | 2026-08-08 |
| CursorBench 4.0 CursorBench is Cursor's internal coding benchmark, built from short and often ambiguous requests taken from real Cursor sessions; version 4.0 is the current one. | Claude Opus 5.5Anthropic | 57.8% | 7 | vendor-only | 2026-09-28 |
| ProgramBench ProgramBench asks whether a model can rebuild a program from scratch, from small command-line tools up to FFmpeg, SQLite and the PHP interpreter, given only the program and its documentation. | Kimi K3.1 ProMoonshot (Kimi) | 31.5% | 7 | vendor-only | 2026-11-02 |
| FrontierSWE FrontierSWE gives AI agents ultra-long engineering and research projects, such as reimplementing git in Zig or speeding up a compiler, with up to 20 hours to work. | Gemini 4 ArgonGoogle DeepMind | 55% | 6 | vendor-only | 2026-10-13 |
§ 03 / AGENTS AND COMPUTER USE
Agents and computer use benchmarks
| Benchmark | Leader | Top score | Ranked | Independent | Newest score |
|---|---|---|---|---|---|
| Terminal-Bench 4.0 Terminal-Bench 4.0 is the current Terminal-Bench release: the 3.0 task format with fixed tasks, saturated tasks removed and resources recalibrated, run on an official leaderboard. | Kimi K3.1 ProMoonshot (Kimi) | 65.7% | 33 | 23 of 45 | 2026-11-02 |
| AutomationBench AutomationBench tests whether an AI agent can carry out a business workflow across several apps through their APIs, finding the endpoints itself and following company rules. | MiniMax M3.1 FlashMiniMax | 62.8% | 23 | vendor-only | 2026-10-19 |
| Agents' Last Exam Agents' Last Exam measures whether AI agents can finish long, economically valuable professional workflows with a checkable result, across 55 sub-fields in 13 industry clusters. | GPT-6 AstraOpenAI | 59.3% | 20 | vendor-only | 2026-11-02 |
| OSWorld 2 OSWorld 2 tests computer-use agents on long, realistic desktop and web workflows that take a person about an hour and a half and an agent hundreds of actions. | Claude Opus 5.5Anthropic | 81.8% | 19 | vendor-only | 2026-11-02 |
| Terminal-Bench 2 Terminal-Bench 2 tests AI agents on hard, realistic jobs done entirely in a command-line terminal: building software, setting up machine learning, security work, data processing. | DeepSeek V4.1 FlashDeepSeek | 90.6% | 18 | vendor-only | 2026-09-22 |
| Terminal-Bench 3.0 Terminal-Bench 3.0 is a harder, longer-running set of terminal tasks than version 2, covering databases, machine learning, hardware design, formal proofs, CAD, science and business processes. | DeepSeek V4.1 FlashDeepSeek | 30% | 9 | vendor-only | 2026-09-10 |
| Toolathlon Toolathlon (the Tool Decathlon) tests language agents on long workflows across real software through MCP tools: calendars and Notion, but also WooCommerce, Kubernetes and BigQuery. | MiMo-V2.6-ProXiaomi MiMo | 76.9% | 9 | vendor-only | 2026-09-22 |
| OSWorld-Verified OSWorld-Verified tests computer-use agents on real tasks in a live Ubuntu desktop: web and desktop apps, files and multi-app workflows, controlled through screenshots, mouse and keyboard. | Kimi K3Moonshot (Kimi) | 84.8% | 8 | vendor-only | 2026-09-22 |
| BrowseComp BrowseComp tests how well a browsing agent can find a hard-to-locate fact on the web: 1,266 questions whose short answers are buried in tangled information but easy to check. | Kimi K3.1 ProMoonshot (Kimi) | 92.1% | 6 | vendor-only | 2026-11-02 |
§ 04 / REASONING
Reasoning benchmarks
| Benchmark | Leader | Top score | Ranked | Independent | Newest score |
|---|---|---|---|---|---|
| ARC-AGI-2 ARC-AGI-2 tests abstract reasoning on grid puzzles that people solve easily and AI systems find hard: infer a hidden rule from a few examples and apply it to a new grid. | GPT-6 AstraOpenAI | 95% | 105 | 105 of 109 | 2026-11-05 |
| ARC-AGI-1 ARC-AGI-1 is the original Abstraction and Reasoning Corpus: grid puzzles where a model must infer a hidden rule from a few examples and apply it to a new input. | 6 systems tiedincl. Claude Fable 5 | 98.5% | 99 | 99 of 101 | 2026-11-05 |
| ARC-AGI-3 ARC-AGI-3 is an interactive reasoning benchmark: the agent plays turn-based grid games with no instructions and has to work out the rules and the goal by exploring. | GPT-6 AstraOpenAI | 62.7% | 24 | 35 of 37 | 2026-10-17 |
§ 05 / EXPERT KNOWLEDGE
Expert knowledge benchmarks
| Benchmark | Leader | Top score | Ranked | Independent | Newest score |
|---|---|---|---|---|---|
| Humanity's Last Exam Humanity's Last Exam is a set of 2,500 expert-written questions across dozens of academic subjects, built to be hard for frontier models; each has one checkable answer. | Kimi K3.1 ProMoonshot (Kimi) | 70% | 20 | vendor-only | 2026-11-02 |
| GPQA Diamond GPQA Diamond is a set of 198 graduate-level multiple-choice questions in biology, physics and chemistry, written so that skilled non-experts with web access still get them wrong. | GPT-6 AstraOpenAI | 96% | 11 | vendor-only | 2026-11-02 |
§ 06 / VISION AND DOCUMENTS
Vision and documents benchmarks
| Benchmark | Leader | Top score | Ranked | Independent | Newest score |
|---|---|---|---|---|---|
| Chartography Chartography tests whether a model can read the charts professionals use, in medicine, engineering, finance, manufacturing and science, and answer questions that need several readings and arithmetic. | Kimi K3.1 ProMoonshot (Kimi) | 90.2% | 10 | vendor-only | 2026-11-02 |
| CharXiv Reasoning CharXiv tests chart understanding on real figures from arXiv papers; the reasoning questions need a model to compare, estimate and combine information across a chart. | Kimi K3.1 ProMoonshot (Kimi) | 89% | 10 | vendor-only | 2026-11-02 |
| GDP.pdf GDP.pdf tests whether a model can answer the questions working professionals ask about real PDF documents: leases, datasheets, clinical guidelines, construction plans. | Gemini 3.8 FlashGoogle DeepMind | 35% | 7 | vendor-only | 2026-11-02 |
§ 07 / LONG CONTEXT
Long context benchmarks
| Benchmark | Leader | Top score | Ranked | Independent | Newest score |
|---|---|---|---|---|---|
| MRCR v2 MRCR v2 (multi-round coreference resolution) tests whether a model can find one specific earlier reply in a very long conversation full of near-identical ones and reproduce it exactly. | Gemini 3.7 FlashGoogle DeepMind | 97% | 7 | vendor-only | 2026-08-13 |
§ 08 / CYBERSECURITY
Cybersecurity benchmarks
| Benchmark | Leader | Top score | Ranked | Independent | Newest score |
|---|---|---|---|---|---|
| CyberGym CyberGym tests whether an AI agent can reproduce real software vulnerabilities: given a description of a historical bug and the unpatched code, write an input that triggers it. | MiMo-V2.6-FlashXiaomi MiMo | 95.1% | 8 | vendor-only | 2026-10-19 |
| ExploitGym ExploitGym measures whether an AI agent can turn a known software vulnerability into a working exploit, across userspace programs, Google's V8 engine and the Linux kernel. | GPT-6 AstraOpenAI | 42.4% | 8 | vendor-only | 2026-11-02 |
| ExploitBench ExploitBench grades exploitation as a ladder of 16 capabilities, from reaching buggy code and crashing it up to arbitrary code execution, on real bugs in Chrome's V8 engine. | GPT-6 AstraOpenAI | 100% | 7 | vendor-only | 2026-11-02 |
| SEC-bench Pro SEC-bench Pro tests whether a model can reproduce a real, disclosed vulnerability by writing a working proof-of-concept input, on the V8 and SpiderMonkey browser engines and the Linux kernel. | GPT-6 AstraOpenAI | 85.4% | 7 | vendor-only | 2026-09-22 |
§ 09 / PROFESSIONAL WORK
Professional work benchmarks
| Benchmark | Leader | Top score | Ranked | Independent | Newest score |
|---|---|---|---|---|---|
| Harvey Legal Agent Benchmark Harvey's Legal Agent Benchmark (LAB) tests whether an AI agent can produce legal work a lawyer could review, from a short instruction and the documents of a closed client matter. | Kimi K3.1 ProMoonshot (Kimi) | 23.7% | 7 | vendor-only | 2026-11-02 |
| APEX-Agents APEX-Agents (AI Productivity Index for Agents) tests whether an agent can do long, cross-application work set by investment banking analysts, management consultants and corporate lawyers. | Kimi K3.1 ProMoonshot (Kimi) | 69.4% | 6 | vendor-only | 2026-11-02 |
| JobBench JobBench tests AI agents on the work that experts in 35 occupations say they most want to hand off, packaged as workspaces of real reference files. | MiMo-V2.6-ProXiaomi MiMo | 62% | 5 | vendor-only | 2026-09-22 |
§ 10 / AI RESEARCH
AI research benchmarks
| Benchmark | Leader | Top score | Ranked | Independent | Newest score |
|---|---|---|---|---|---|
| Terminal-Bench-Science Terminal-Bench-Science tests AI agents on real scientific research workflows, contributed by domain experts, carried out in a terminal: data analysis, simulation, theorem proving, image reconstruction, model fitting. | GPT-6 AstraOpenAI | 68.1% | 9 | vendor-only | 2026-11-02 |
| PostTrainBench PostTrainBench tests whether an AI agent can post-train a small base language model on its own, with 10 hours on one H100 GPU, to score as high as possible on a target benchmark. | Gemini 4 ArgonGoogle DeepMind | 45.3% | 6 | vendor-only | 2026-11-02 |
| MLE-bench MLE-bench tests whether an AI agent can do machine-learning engineering end to end on 75 real Kaggle competitions: prepare data, train models and submit. | Gemini 3.6 FlashGoogle | 63.9% | 5 | vendor-only | 2026-07-21 |
Benchmarks FAQ
Answers derived from the 655 published scores in this section.
Q · 01 Which AI model is best on benchmarks right now?
+
It depends on the test. On ARC-AGI-2 GPT-6 Astra leads at 95%; On ARC-AGI-1 6 systems tie for first at 98.5%; On Terminal-Bench 4.0 Kimi K3.1 Pro leads at 65.7%; On DeepSWE Gemini 4 Argon leads at 77.9%; On ARC-AGI-3 GPT-6 Astra leads at 62.7%. Each leaderboard ranks its own test; a model that leads one often trails on another.
Q · 02 Are AI benchmark scores independently verified?
+
Some are. 262 of the 655 scores here come from a benchmark's own leaderboard (ARC-AGI-2, ARC-AGI-1, ARC-AGI-3, Terminal-Bench 4.0), where one team runs every model on the same harness. The rest are reported by the labs that make the models. Every row carries a badge saying which.
Q · 03 Why do two sources give the same model different scores?
+
Different harnesses, effort levels, subsets and dates. A lab may report its model at maximum reasoning effort on its own scaffold, while the benchmark's leaderboard runs a fixed setup for everyone. Each leaderboard here keeps the other numbers and lists them under “Where the sources disagree”.
Q · 04 Where do the prices next to each score come from?
+
From our model price index: per-token rates read on each vendor's own pricing page, each with the date it was read. A score never changes the price, and a repricing moves every leaderboard on the next update.