Leaderboard · Source coverage · Contribute

BenchBoard indexes published benchmark results for current AI models. Every score keeps its model release, benchmark variant, evaluation setting, and original source attached. Results from different harnesses or versions remain separate, and the site does not invent a cross-benchmark composite score.

_{126 curated releases + 27 comparison-only models · 1,203 separately ranked metric and version views}

	2026-10-12 VOICE BENCHMARK AUDIT	Added the exact values from OpenAI's GPT-Live-1 and GPT-Realtime-2 release charts: 25 results across nine voice benchmark views. Tau3 voice backends, Full Duplex Bench versions, reasoning settings, and turn-taking latency remain distinct.
	2026-10-11 LEADERBOARD AUDIT	Completed the 35-row Terminal-Bench 4.0 official leaderboard, including separate effort settings for GPT-6 Astra, Claude Fable 5.1, and Claude Opus 5. Maintainer-run scores retain their harness and run details. Flagged the PinchBench v2 task-count conflict: 147 on the live board versus 148 in its release notes.
	2026-10-10 MODEL RELEASES	Added Claude Sonnet 5.5, Claude Haiku 5.5, Gemini 4 Argon, and Mistral Large 4 Preview. Imported the exact capability results reported by their official release pages and system cards.
	2026-09-28 MODEL RELEASES	Added Claude Opus 5.5, Grok 4.7, GPT-6 Sol, and GPT-6 Luna, plus the full MentalHealthBench views and complete CursorBench 4.0 table.
	2026-09-21 SOURCE AUDIT	Completed a broad first-party sweep across release pages, model cards, repositories, and technical reports. Added a generated source coverage matrix.
	2026-09-20 COLLECTION	Added the Multimodal Input × Harness collection for tasks that give an agent images, audio, video, screens, or other non-text files.
	2026-09-18 PROJECT LAUNCH	Published the first BenchBoard dataset with representative GPT, Claude, Gemini, DeepSeek, Qwen, GLM, Seed, Kimi, and Hy releases.

Search or scan the complete benchmark index, compare up to three releases, and export the current view as CSV. Opening a model shows every recorded result and setting for that release. The coverage matrix records how completely each source has been checked, and conflicting published values remain visible.

The index focuses on recent general, coding, multimodal, and agent models. Older models remain when an official result table uses them as comparisons. Current coverage includes GPT-6, Claude 5.5, Gemini 3.8, DeepSeek V4.1, Qwen3.8, GLM-5.3, Seed2.1, Kimi K3, Hy4, Grok 4.7, and many related releases.

Benchmark families appear once in navigation. Versions, subsets, and metrics are selectable inside that entry. Settings such as reasoning effort, harness, tool access, task count, and snapshot date remain attached to each result.

Source and comparison policy

Model providers, benchmark maintainers, model cards, technical reports, and official repositories are preferred. A provider can report results for competing models; those rows retain the reporting source. Conflicting first-party values stay separate until the settings or data can be reconciled.

A source audit records whether a page was checked in full, checked for selected model columns, used only for metadata, partially imported, inaccessible, or still pending. The generated model source coverage matrix makes that status visible.

BenchBoard is an evidence index. It does not claim independent reproduction or recommend one model from a single score.

Multimodal Input × Harness criteria

This collection requires both a named agent harness, such as OpenClaw, Claude Code, Codex, OpenCode, or OpenHands, and a task that supplies non-text input. Advertising a model as multimodal is not enough. Text-only generation judged by a vision model is also excluded.

Mixed suites are labeled separately from dedicated multimodal subsets. The input modalities and harness remain visible in the benchmark entry.

Missing result or incorrect setting? Read CONTRIBUTING.md, then open an issue or pull request with the original source.

BenchBoard is a static site. To run it locally:

python3 -m http.server 8000

After editing the registry or a data pack:

npm run check
npm run sync-readme
npm run sync-source-matrix
npm run audit

Pushes to main are checked and deployed to GitHub Pages. The code and data organization are available under the MIT License.