BrowseComp

Canonical page on the main site: sinoaihub.com/benchmarks/browsecomp

Type

benchmark

Task type: Browsing agent benchmark (locate hard-to-find, entangled information on the internet)
Dataset size: 1,266 problems
Evaluation method: Short-answer questions with a single correct answer; graders verify the exact answer (encrypted set)
Scoring: Accuracy (% correct)
Recorded evaluations: 4

Benchmark of browsing and retrieval ability: locating obscure information using web search and browsing.

Evaluations

benchmark	model	score	metric	date	source_type	source_url
BrowseComp	kimi-k3	91.2 (90.4 with full 1M context, no compaction)	accuracy	2026-07	vendor_reported	https://github.com/MoonshotAI/Kimi-K3
BrowseComp	kimi-k2.5	60.6 (74.9 with context management; 78.4 Agent Swarm)	accuracy	—	vendor_reported	https://github.com/MoonshotAI/Kimi-K2.5
BrowseComp	minimax-m3	83.5	accuracy	2026-06-01	vendor_reported	https://www.minimax.cn/models/text/m3

Task type: Browsing agent benchmark (locate hard-to-find, entangled information on the internet)

Dataset size: 1,266 problems

Evaluation method: Short-answer questions with a single correct answer; graders verify the exact answer (encrypted set)

Scoring: Accuracy (% correct)

All scores are vendor-reported and not independently verified. Evaluation setups (context management, agent scaffolding) differ between vendors.

verified

No documented benchmark changes on record as of 2026-09-29.

2026-09-20

No documented source-change events located as of 2026-09-20.

evidence_id	source_name	source_url	source_type	published	verified	confidence	conflict
src-benchmarks-browsecomp-1	Kimi K3 GitHub README	https://github.com/MoonshotAI/Kimi-K3	Official documentation	—	2026-09-20	high	—
src-benchmarks-browsecomp-2	MiniMax official M3 model page	https://www.minimax.cn/models/text/m3	Official documentation	—	2026-09-20	high	—