Results
| 1 | claude-fable-5-2 | anthropic | 2.8 | |
|---|---|---|---|---|
| 2 | gpt-6-astra | openai | 3 | |
| 3 | claude-opus-5-5 | anthropic | 3.1 | |
| 4 | gemini-4-argon | 3.07 | ||
| 5 | claude-fable-5-1 | anthropic | 3.1 | |
| 6 | claude-opus-5 | anthropic | 3 |
Public leaderboard · Aggregated leaderboards
Scale AI tests AI models on private, expert-written questions that models cannot have memorised, across areas like coding, reasoning and safety. One representative board (Humanity's Last Exam, Diamond) is shown.
Collection of expert-built, mostly private evaluations (such as SWE-Bench Pro, MCP Atlas, MultiChallenge, Remote Labor Index, HLE) run by Scale.
Share of expert-level questions answered correctly in Humanity's Last Exam (Diamond) (%) — higher is better.
| 1 | claude-fable-5-2 | anthropic | 2.8 | |
|---|---|---|---|---|
| 2 | gpt-6-astra | openai | 3 | |
| 3 | claude-opus-5-5 | anthropic | 3.1 | |
| 4 | gemini-4-argon | 3.07 | ||
| 5 | claude-fable-5-1 | anthropic | 3.1 | |
| 6 | claude-opus-5 | anthropic | 3 |
Note: SEAL hosts ~30 leaderboards; only HLE (Diamond) shown, non-deprecated entries. as_of is the board's updatedAt field; newest entries dated Nov 2026. Some entries carry a contamination warning on the site. Data read from page-embedded JSON.
Source: Scale AI (SEAL) · labs.scale.com · results as of 2026-09-23 · retrieved 2026-11-17