Reported scores0–1
- GPT-5.6 Sol0.73
- GPT-5.6 Terra0.7
- Claude Fable 50.7
Source order · setup not recorded
View all 25 observationsTHE BENCHMARK DIRECTORY
AI benchmarks
Task definitions, reported model results
and the sources behind them.
3,772 source records / 5 catalog sources
Featured benchmarks
Explore the tasks. See the results. Go straight to the evidence.
Reported scores0–100%
- Claude 4 Opus79.6%
- Claude 4 Sonnet75.4%
- Gemini 2.5 Pro86.4%
Single attempt · pass@1
View all 33 observationsReported scores0–100%
- Apodex 1.110.3%
- Claude Opus 517.5%
- DeepSeek V4 Flash-073112.4%
97 tasks · Claude Code
View all 13 observationsAbout these previews
Three observations per preview, in source order, not a universal model ranking. Bars use each record’s original units and declared maximum. GPQA Diamond shows one reported pass@1 setup; FrontierChallenge shows the Claude Code setup. Matching labels do not establish identical conditions, and a report may cite other organizations’ results. DeepSWE 1.1 has no recorded protocol. Open a record for all setups and individual citations. Dates describe collection or report publication, not a live update.
Make sense of the results
All articles →Benchmarks and leaderboards measure different things
Separate the task definition from the ranking before comparing models.
How to compare AI scores
Check versions, tools, budgets, and uncertainty alongside the score.
Choosing benchmarks for agents
Match the environment and success criteria to the work you need done.
The directory3,772 records
Download catalog ↓Claire Radar
τ^τ-Bench: An Environment for End-To-End, Realistic Agent Construction
A benchmark where a developer agent builds a complete customer-service agent from business records, client requirements, a production API, inherited codebase, and cost/model limits, scored by deployment against held-out simulated users across 53 tasks in four domains.
PaperDetails
Claire Radar
2d-fet-bench
Field-effect transistor (FET) layouts on exfoliated two-dimensional flakes are typically drawn by hand for each flake, placing contacts and gates to match its position and outline in optical micrographs.
PaperDetails
Claire Radar
360CityArena
360CityArena evaluates embodied agents in a photorealistic virtual urban environment built from 360-degree video of Tokyo's Akihabara district. It includes 175 tasks across environment understanding, path reasoning, and spatial reasoning, testing localization, landmark search,…
Paper · CodeDetails
Claire Radar
3DCodeBench
3DCodeBench evaluates vision-language model agents on procedural 3D modeling by converting text and image references into Blender Python code. It includes 212 object categories with ground-truth scripts, and scores outputs on executability, image similarity (SigLIP-2/DINOv3),…
Paper · CodeDetails
Claire Radar
4DSynth-Nav
Evaluates embodied agents on interactive navigation tasks in procedurally generated 4D environments with independently tunable difficulty axes.
PaperDetails
Claire Radar
5-Dialects-BN
Evaluation object not clearly defined; abstract mentions alignment of Romanized transliteration with dialectal text, Standard Bangla, English, and subjectivity labels across five Bangla regional varieties, but no explicit task or scoring setup is provided.
PaperDetails
Claire Radar
A Large-Scale Dataset and Benchmark
InteractBind provides a large-scale dataset of ~100k protein-ligand pairs with fine-grained binding-site localization tasks and interaction maps for six non-covalent interaction types, plus affinity and similarity-controlled splits.
PaperDetails
Claire Radar
A Real-World CAD Segmentation Benchmark
CAD segmentation datasets are predominantly synthetic or automatically labelled, leaving it unclear how well existing methods transfer to real-world engineering tasks.
PaperDetails
OpenCompass Hub
A-Bench
A-Bench is a benchmark designed to diagnose whether LMMs are masters at evaluating AIGIs. 2,864 AIGIs from 16 text-to-image models are sampled, each paired with question-answers annotated by human experts, and tested across 18 leading LMMs.
Paper · Code · DatasetDetails
OpenCompass Hub
A-OKVQA
A-OKVQA assesses commonsense reasoning abilities. It is a crowdsourced dataset composed of a diverse set of about 25K questions requiring a broad base of commonsense and world knowledge to answer.
Paper · CodeDetails
Claire Radar
A3S-Bench
A3S-Bench is an executable security test suite for evaluating autonomous agents across multi-turn scenarios and agent-level risks using an action-grounded oracle.
Paper · CodeDetails
Artificial Analysis
AA-AnalystAgent
Quantitative analysis on spreadsheets & documents
Source record30 results
One record per source. Related records retain their own evidence and evaluation context.Missing scores remain in the catalog.