Model scores by category.
Third-party leaderboard results, copied as published with the board, metric, harness, as-of date and source. We don't make our own score.
Boards checked · 409 results from 13 boards · 114 model pages
Code
Repository bug-fix and code-editing benchmarks (SWE-bench Verified, Aider polyglot).
- GPT-6 Astra (xhigh)57.88%
- GPT-6 Astra (high)57.88%
- Fable 5.1 (xhigh)57.88%
- GLM-5.2 [high]62.9%
- Junie Agent61.8%
- Codex Agent58%
- qwen3-coder-480b-a35b38.7%
- minimax-2.136.81%
- kimi-k2-instruct27.67%
Agentic / terminal
End-to-end tasks an agent completes in a terminal. Scores are per model and harness.
- Opus 5.5 (max)64.85%
- Sonnet 5.5 (max)61.82%
- GPT-6 Astra (max)58.18%
Websites
Human-preference arenas for front-end and full-stack web builds. Elo measures preference, not correctness.
- claude-opus-5.5-max1807
- gpt-6-astra-max1774
- claude-sonnet-5.5-xhigh1769
- Claude Sonnet 5.592.39%
- Gemini 4 Argon91.91%
- Claude Haiku 5.590.44%
- gpt-6-astra1357
- claude-fable-5-11326
- kimi-k31315
Design
Human-preference votes on generated UI. Measures taste, not whether the code works.
- claude-opus-5-51389
- gpt-6-astra1383
- muse-wonder-1.01377
Android
Android app tasks run by Google. Vendor-run, small task sets, wide confidence intervals.
- Claude Opus 5.532.7%
- GPT 6 Astra28%
- Claude Fable 5.122.7%
- Kimi K390.2%
- Qwen3.8 Max87%
- GLM 5.272.2%
iOS
No mature public leaderboard exists.
No mature public leaderboard
Which labs hold the top spots
Count of top-10 rows per lab on each category's active boards. This counts rows; it is not a score or a ranking of labs.
Code
| Lab | Rows |
|---|---|
| OpenAI | 5 |
| Anthropic | 4 |
| 1 | |
| Meta | 2 |
| Chinese labs | 12 |
| Others | 3 |
Sources: tbench.ai · swe-rebench.com · scale.com
Agentic / terminal
| Lab | Rows |
|---|---|
| OpenAI | 2 |
| Anthropic | 3 |
| 1 | |
| xAI | 1 |
| Chinese labs | 3 |
Sources: tbench.ai
Websites
| Lab | Rows |
|---|---|
| OpenAI | 5 |
| Anthropic | 9 |
| 3 | |
| xAI | 1 |
| Meta | 4 |
| Chinese labs | 6 |
| Others | 2 |
Sources: lmarena.ai · vals.ai · designarena.ai
Design
| Lab | Rows |
|---|---|
| OpenAI | 3 |
| Anthropic | 3 |
| Meta | 3 |
| Chinese labs | 1 |
Sources: designarena.ai
Android
| Lab | Rows |
|---|---|
| OpenAI | 2 |
| Anthropic | 2 |
| 1 | |
| Chinese labs | 12 |
Sources: developer.android.com · developer.android.com
What we did not import13
- SWE-bench Verified: Vals AI no longer scores new releases (page 'Updated 9/1/2026'; parent cites 2026-09-05). Treat as final standings.
- WebDev Arena: Gemini 4 Argon (High) 1678, qwen3.8-max(-0902) and hy4-preview are marked 'Preliminary' on the board (as of 2026-10-13).
- Aggregator-sourced numbers (theindexofai, genztech.blog, kunya.ai, llmboard.ai, paperswithcode.co, benchlm.ai, llm.ing, anotherwrapper.com) were NOT imported. Each needs the original board; Terminal-Bench 2.0, Design Arena Website and LiveCodeBench 93.5 (DeepSeek V4 Pro Max) could not be confirmed on the original boards.
- GPT-6 Astra DeepSWE 74.1% is OpenAI-reported via an aggregator (kunya.ai); not imported.
- Design Arena per-category boards (Website, UI Component, Mobile Apps, Native Apps) require a Bearer API key per Design Arena API docs (https://docs.designarena.ai/api-reference/leaderboard); only the server-rendered All-categories and Fullstack boards were captured.
- Aider polyglot: not updated since 2025-10-03 (last commit to leaderboard YAML 2025-10-04) - frozen; prefer Vals Vibe Code Bench / Terminal-Bench 4.0
- swebench.com: all boards end 2026-02-26 or earlier - frozen; Vals SWE-bench Verified frozen 9/1/2026 (saturated)
- LiveCodeBench official site data still ends 2025-04-07; Vals LCB frozen 9/1/2026 - no active LCB board found
- SWE-Bench Pro (Scale) public board is active (Muse Spark 1.1 marked NEW) but lists no 2026 Anthropic/OpenAI frontier models; no as-of date on page
- SWE-rebench last changelog 2026-07-01 - stale
- superseded flag is an automatic same-family/newer-version heuristic within each board; verify before hiding rows
- Design Arena per-category boards (android, mobile, ui-components) still need an API key
- hosting_region: no model card states a hosting region; Muse Spark Open lists US-and-other hosted partners (Baseten, Fireworks, Groq, Modal, Nebius, OpenRouter, Together) without regions
License and hosting region
license_as_listed is copied from the board's listing (LMArena WebDev shows a license column); hosting_region is not_confirmed until checked against each model card. Proprietary-API models from Chinese vendors are tagged with license 'Proprietary' as listed.