BrowseComp leaderboard
BrowseComp tests persistent web research. Its 1,266 questions have short, verifiable answers that are genuinely difficult to locate, requiring an agent to chase entangled information across many pages rather than retrieve a single fact. OpenAI built it and released it in 2025, and it has become the standard reference for deep-research and agentic-search products.
Last refreshed 2026-08-26. 11 models scored on this benchmark.
Full leaderboard
| # | Model | Provider | Score | Released |
|---|---|---|---|---|
| 1 | Kimi K3 | Moonshot AI | 91.2%vendor | 2026-07 |
| 2 | Claude Opus 5 | Anthropic | 90.8%vendor | 2026-07 |
| 3 | GPT-5.6 Sol | OpenAI | 90.4%vendor | 2026-07 |
| 4 | GPT-5.6 Terra | OpenAI | 87.5%vendor | 2026-07 |
| 5 | Claude Fable 5 | Anthropic | 87.4% | 2026-06 |
| 6 | Claude Sonnet 5 | Anthropic | 84.7%vendor | 2026-06 |
| 7 | Claude Opus 4.8 | Anthropic | 84.3% | 2026-05 |
| 8 | MiniMax M3 | MiniMax | 83.5%vendor | 2026-06 |
| 9 | GPT-5.6 Luna | OpenAI | 83.3%vendor | 2026-07 |
| 10 | LongCat-2.0 | Meituan | 79.9%vendor | 2026-06 |
| 11 | Inkling | Thinking Machines | 77.1%vendor | 2026-07 |
Score interpretation
Two caveats that matter more here than on any other benchmark we track. First, there is no official leaderboard and no independent evaluator: every published BrowseComp figure is self-reported by the lab that produced it. Second, scores move sharply with context-management strategy, so a single-agent run, a multi-agent run, and a run using context compaction are not the same measurement even for the same model. What a BrowseComp number really ranks is a full system, the model plus its tools plus its context policy, not a bare model. Frontier systems now cluster within a couple of points of each other, which suggests the benchmark is approaching saturation at the top.
Why this matters for AI agents
If you are building research agents, this is the closest published proxy for whether the thing will actually find an obscure answer instead of confidently inventing one. Just do not treat small gaps between top models as real. The measurement noise from differing harnesses and context strategies is larger than the differences between the leaders.
Other benchmarks
Premium API: time-series for BrowseComp
The leaderboard above is a snapshot. Want to see how a model's BrowseComp score has moved over the last 30-90 days, or set a webhook that fires when a score crosses a threshold? The premium API has both:
/api/premium/history/benchmarks/series?model=&benchmark=browsecomp: daily score evolution for one model on this benchmark, 1 credit per call