Benchmark

BrowseComp Benchmark

What is BrowseComp?

BrowseComp is a benchmark comprising 1,266 questions that challenge AI agents to persistently navigate the internet in search of hard-to-find, entangled…

Released
2025-04-10
Modality
text
Evaluates
reasoning, search, agents
Openness
restricted
Importer
LLM Stats
Reported scores
62

Source provenance

BrowseComp paper and code

BrowseComp results and reported scores

LLM Stats

ModelOrganizationReported valueReportedEvidence
Claude Mythos PreviewAnthropic0.8692026-04-07evidence
Claude Opus 4.6Anthropic0.842026-02-05evidence
Claude Opus 4.7Anthropic0.7932026-04-16evidence
Claude Opus 4.8Anthropic0.8432026-05-28evidence
Claude Opus 5Anthropic0.9082026-07-24evidence
Claude Sonnet 4.6Anthropic0.7472026-02-17evidence
Claude Sonnet 5Anthropic0.8472026-06-30evidence
DeepSeek-R1-0528DeepSeek0.0892025-05-28evidence
DeepSeek-V3.1DeepSeek0.32025-01-10evidence
DeepSeek-V3.2DeepSeek0.5142025-12-01evidence
DeepSeek-V3.2 (Thinking)DeepSeek0.5142025-12-01evidence
DeepSeek-V3.2-ExpDeepSeek0.4012025-09-29evidence
DeepSeek-V4-Flash-0423DeepSeek0.5352026-04-23evidence
DeepSeek-V4-Flash-MaxDeepSeek0.7322026-04-23evidence
DeepSeek-V4-Pro-MaxDeepSeek0.8342026-04-23evidence
Gemini 3.1 ProGoogle0.8592026-02-19evidence
GLM-4.5Z.ai0.2642025-07-28evidence
GLM-4.5-AirZ.ai0.2132025-07-28evidence
GLM-4.6Z.ai0.4512025-09-30evidence
GLM-4.7Z.ai0.522025-12-22evidence
GLM-4.7-FlashZ.ai0.4282026-01-19evidence
GLM-5Z.ai0.7592026-02-11evidence
GLM-5.1Z.ai0.7932026-04-07evidence
GPT-5OpenAI0.5492025-08-07evidence
GPT-5.2OpenAI0.6582025-12-11evidence
GPT-5.2 ProOpenAI0.7792025-12-11evidence
GPT-5.4OpenAI0.8272026-03-05evidence
GPT-5.5OpenAI0.8442026-04-23evidence
GPT-5.5 ProOpenAI0.9012026-04-23evidence
GPT-5.6 LunaOpenAI0.8332026-07-09evidence
GPT-5.6 SolOpenAI0.9042026-07-09evidence
GPT-5.6 TerraOpenAI0.8752026-07-09evidence
Grok 4 FastxAI0.4492025-08-28evidence
Hy3Tencent0.8422026-07-06evidence
Inkling-SmallThinking Machines Lab0.7742026-07-30evidence
Kimi K2-Thinking-0905Moonshot AI0.6022025-09-05evidence
Kimi K2.5Moonshot AI0.7492026-01-27evidence
Kimi K2.6Moonshot AI0.8632026-04-20evidence
Kimi K3Moonshot AI0.9122026-07-16evidence
LongCat-Flash-Thinking-2601Meituan0.5662026-01-14evidence
MiMo-V2-FlashXiaomi0.5832025-12-16evidence
MiniMax M2MiniMax0.442025-10-27evidence
MiniMax M2.1MiniMax0.622025-12-23evidence
MiniMax M2.5MiniMax0.7632026-02-12evidence
MiniMax M3MiniMax0.83522026-06-01evidence
Mistral Medium 3.5Mistral0.4862026-04-29evidence
Nemotron 3 Super (120B A12B)NVIDIA0.31282026-03-11evidence
Nemotron 3 Ultra (550B A55B)NVIDIA0.4442026-06-04evidence
Nemotron 3.5 Lightning (30B A3B)NVIDIA0.36972026-08-11evidence
o3OpenAI0.4972025-04-16evidence
o4-miniOpenAI0.5152025-04-16evidence
Qwen3.5-122B-A10BQwen0.6382026-02-24evidence
Qwen3.5-27BQwen0.612026-02-24evidence
Qwen3.5-35B-A3BQwen0.612026-02-24evidence
Qwen3.5-397B-A17BQwen0.692026-02-16evidence
Sarvam-105BSarvam AI0.4952026-03-06evidence
Sarvam-30BSarvam AI0.3552026-03-06evidence
Seed 2.0 ProByteDance0.7732026-02-14evidence
Seed 2.1 ProByteDance0.8622026-06-24evidence
Seed 2.1 TurboByteDance0.8492026-06-24evidence
Solar Pro 4Upstage0.4922026-08-06evidence
Step-3.5-FlashStepFun0.692026-02-02evidence

Scores are partitioned by the source that reported them and are never merged into a single cross-source ranking, because the sources measure different things and say so.

Which sources cite BrowseComp?

Related reasoning, search, agents benchmarks