Skip to main content
BenchLM
Data
Specialist vertical · BenchAlign v6.0 · 74 ranked (36 Supported, 38 Estimated)
Scored from the dataset of

Search and deep research leaderboard

As of November 21, 2026, Claude Fable 5.2 leads our search and deep research leaderboard at 86.6 (Supported, $8/$40 per 1M tokens), ahead of GPT-6.2 Astra at 88.6. Gemini 4 Carbon (#3, 83.7) is the lowest-cost model near the top at $2/$10 per 1M tokens; MiMo-V2.6-Pro leads open-weight models at #15 (69.0). We rank 74 models (36 Supported, 38 Estimated) across 9 domain protocols, and this vertical carries zero weight in the overall score.

Shortlist by workload and budget

Each pick comes directly from the live search & deep research rows below, with listed API rates and the case where another option wins.

Highest point score#1 · Supported

Anthropic · $8/$40 per 1M tokens

Costs $8/$40 per 1M tokens; Gemini 4 Carbon (#3, 83.7) cuts token spend to $2/$10 per 1M tokens.

Closest contender#2 · Estimated

OpenAI · $8/$40 per 1M tokens

Sits -2.0 points behind Claude Fable 5.2; test both on your own prompts when provider lock-in or latency matters.

Best API value#3 · Supported

Google · $2/$10 per 1M tokens

Keeps 97% of the top score (83.7 vs 86.6), but gives up 2.9 points on the hardest tail cases.

Top open-weight model#15 · Estimated

Xiaomi · Open weights · $0.43/$0.87 per 1M tokens

Trails Claude Fable 5.2 by 17.6 points (#15 overall here), so pick it when weight control or on-prem residency outweighs peak score.

Ranked models

Ordered by BenchAlign v6.0 point estimate as of November 21, 2026. An Estimated row backed by fewer than two source families never ranks above a row scoring within 2 points below it. Conditional ranges do not establish rank confidence.

Search & Deep Research LLM Leaderboard
RankModelScoreEvidence
#1Anthropic86.6Supported
#2OpenAI88.6Estimated
#3Google83.7Supported
#4OpenAI78.2Supported
#5Anthropic76.7Supported
#6OpenAI76.6Supported
#7OpenAI76.4Supported
#8Google75.3Supported
#9Google75.0Supported
#10Anthropic73.5Supported
#11Anthropic73.3Supported
#12Alibaba72.0Estimated
#13OpenAI71.4Supported
#14OpenAI69.2Supported
#15
MiMo-V2.6-ProOpen weights
Xiaomi
69.0Estimated
#16DeepSeek68.3Supported
#17Anthropic68.3Supported
#18MiniMax68.2Estimated
#19
Kimi K3.1 ProOpen weights
Moonshot AI
68.1Supported
#20
Kimi K3Open weights
Moonshot AI
67.8Supported
#21Anthropic67.4Supported
#22Anthropic66.0Estimated
#23OpenAI65.4Supported
#24xAI64.3Supported
#25OpenAI64.0Supported
#26OpenAI63.2Supported
#27
Step 5 PreviewOpen weights
StepFun
62.7Estimated
#28
Qwen3.8 MaxOpen weights
Alibaba
62.6Supported
#29
GLM-5.3Open weights
Z.AI
64.5Estimated
#30
GLM-5.3-FlashOpen weights
Z.AI
62.3Estimated
#31Anthropic62.3Supported
#32Meta63.9Estimated
#33Google61.7Supported
#34OpenAI61.2Estimated
#35OpenAI61.0Estimated
#36Anthropic58.9Supported
#37Anthropic58.7Supported
#38Anthropic58.7Supported
#39Google58.5Supported
#40Google60.1Estimated
#41
Muse Spark OpenOpen weights
Meta
55.9Estimated
#42Anthropic55.8Supported
#43Anthropic55.8Supported
#44OpenAI55.7Supported
#45
Kimi K2.6Open weights
Moonshot AI
55.2Estimated
#46xAI56.8Estimated
#47
InklingOpen weights
Thinking Machines Lab
53.6Estimated
#48Alibaba55.5Estimated
#49
GLM-5.2Open weights
Z.AI
55.5Estimated
#50
Qwen3.8-27BOpen weights
Alibaba
52.9Estimated
#51Google52.4Supported
#52Google54.3Estimated
#53OpenAI51.4Supported
#54OpenAI52.8Estimated
#55Anthropic52.3Estimated
#56Anthropic49.9Supported
#57Anthropic49.9Supported
#58xAI50.5Estimated
#59xAI48.7Estimated
#60OpenAI46.4Estimated
#61xAI45.1Estimated
#62
Gemma 4 31BOpen weights
Google
43.0Supported
#63
Mistral Large 4Open weights
Mistral
42.7Estimated
#64Google43.5Estimated
#65OpenAI43.4Estimated
#66Google43.2Estimated
#67Google42.7Estimated
#68Google41.9Estimated
#69OpenAI39.7Estimated
#70
Gemma 4 26B A4BOpen weights
Google
37.2Estimated
#71OpenAI33.4Estimated
#72
Gemma 3 27BOpen weights
Google
29.4Estimated
#73
GPT-OSS 120BOpen weights
OpenAI
27.3Estimated
#74
GPT-OSS 20BOpen weights
OpenAI
27.0Estimated

Scored domain benchmarks (9)

Each percentage is the protocol's relative reference weight inside this vertical's calibrated BenchAlign v6.0 model, not a raw score average. Independent evaluator shares are capped at 40% whenever three or more independent evaluators contribute.

FACTS Search V2 measures factual answering with a shared Brave Search tool. Overall, public, private, hedging, search-count, attempted-accuracy, and F1 measurements remain distinct.

17% ref. weight

Epoch AI's EBR-bench tests whether models get better at an unfamiliar game with practice: they play the campaign card game Earthborne Rangers repeatedly, take notes, and are scored on later playthroughs.

16% ref. weight

BrowseComp tests whether a browsing agent can find difficult, fact-seeking answers across the web. Its 1,266 problems use short answers that are easier to grade than an open-ended research report.

13% ref. weight

A broad research-agent benchmark for open-ended information gathering, synthesis, and answer construction across wide search spaces.

11% ref. weight

Multi-step information synthesis and source verification on fresh LiveBench tasks.

10% ref. weight

Interactive browser navigation and page-state extraction on fresh LiveBench web tasks.

10% ref. weight

An agentic browsing benchmark where models search the web, gather evidence, and answer list-style questions using browser tools.

8% ref. weight

Agentic data-analysis tasks scored against per-task rubrics at a 980K-token budget.

7% ref. weight

The archived FACTS Search V1 protocol measures factual answering with Brave Search and a Gemini 2.0 Flash grader. Its scores remain separate from Search V2.

7% ref. weight

Multi-hop web browsing, enterprise retrieval, and citation-backed synthesis

Nine search and research benchmarks power this board, led by FACTS Search V2 (17%), EBR-bench (16%), BrowseComp (13%) alongside WideResearch, LiveBench 2.0 Research and Browser, DeepSearchQA, and DRACO. These evaluations test whether an agent can decompose ambiguous multi-constraint queries, reformulate failed searches, reconcile conflicting web pages, and synthesize long-form reports whose citations actually support each sentence.

Thirty-four of the 74 ranked models carry Supported status, giving buyers multi-evaluator evidence across Google DeepMind's FACTS Search suite, BrowseComp, EBR-bench, and LiveBench.

Your search index, page parser, and retry cap move accuracy as much as the LLM

A search score never belongs to model weights in isolation. Switching from SerpAPI to an internal Elasticsearch cluster, stripping JavaScript tables during HTML-to-Markdown conversion, or capping an agent at five tool calls instead of twenty will reorder close models immediately.

Benchmark snapshots also cannot guarantee live breaking-news freshness or paywall access in your own environment; evaluate candidate models inside your exact retrieval and reranking stack before shipping customer-facing answers.

Questions

Which LLM is best for web search and deep research right now?

Claude Fable 5.2 holds #1 on our web search and deep research leaderboard at 86.6 (Supported) as of November 21, 2026, ahead of GPT-6.2 Astra at 88.6 and Gemini 4 Carbon at 83.7. Both top models hold Supported status across independent public evaluations.

What is the best open-weight LLM for web search and deep research?

MiMo-V2.6-Pro from Xiaomi leads open-weight models on this web search and deep research board at #15 with a calibrated score of 69.0 (Estimated). That trails Claude Fable 5.2 by 17.6 points, so choose MiMo-V2.6-Pro when self-hosting, VPC data residency, or weight customization matters more than peak frontier accuracy.

Which API model offers the best value for web search and deep research?

Gemini 4 Carbon (#3, 83.7) is our top API value pick for web search and deep research, priced at $2/$10 per 1M tokens compared with Claude Fable 5.2 at $8/$40 per 1M tokens. You give up 2.9 calibrated points, making Gemini 4 Carbon the better buy whenever those extra points do not cut downstream retries.

Which benchmarks make up the web search and deep research leaderboard score?

BenchAlign v6.0 calibrates 9 open and primary domain protocols for this board, led by FACTS Search V2 (17%), EBR-bench (16%), BrowseComp (13%), WideResearch (11%). Each percentage reflects share of the vertical's reference weight inside our item-response model rather than a raw score average, and no single independent evaluator can exceed 40% when three or more evaluators contribute.

How much does the browsing harness affect search leaderboard scores?

Substantially. FACTS Search, BrowseComp, EBR-bench, and WideResearch each pair models with a specific search index, page-fetch parser, and step limit. Changing search APIs, stripping HTML tables, or lowering tool-call budgets can flip two-point leads, so benchmark your top two finalists inside your own retrieval pipeline.

How does the search leaderboard differ from the factuality leaderboard?

Our factuality board tests closed-book parametric accuracy and faithfulness to static passages already in the prompt. This search board tests active query decomposition, multi-hop web navigation, enterprise index retrieval, and cross-document synthesis across nine agentic research protocols, carrying zero direct weight in the overall score.

Search & deep research leaderboard updates

Weekly updates when new models or domain evaluations move the search & deep research rankings.

Read a sample issue

Join 7,000+ subscribers.

One email each week. Unsubscribe anytime.