Benchmark Almanac

Deep Research Bench (FutureSearch)

Claude Opus 4.6 0.55 leads since February 2026
Not in model cards

Web research agents work on multi-step research tasks against RetroSearch, a frozen copy of 10,000 to 100,000 web pages per task. Answers are compared with answers worked out by people.

The paper had 89 tasks and the June 2026 leaderboard has 169. Some task types use LLM grading. Not the same benchmark as the USTC DeepResearch Bench. Epoch AI tracks this benchmark as DeepResearch Bench in its capabilities index.

Made by
FutureSearch
Released
May 2025
Kind
Deep research
Level
Expert
Scored
Checked automatically
Answers
Agent task
Tasks
Kept private
Strongest model tested
GPT-5.5 #20 on the capability index
Last model added
May 2026
Scores refreshed
November 12th, 2026

In the authors’ words

Amongst the most common use cases of modern AI is LLM chat with web search enabled. However, no direct evaluations of the quality of web research agents exist that control for the continually-changing web. We introduce Deep Research Bench, consisting of 89 multi-step web research task instances of varying difficulty across 8 diverse task categories, with the answers carefully worked out by skilled humans. We provide a "RetroSearch" environment with a large frozen set of scraped web pages, and demonstrate that offline "RetroSearch" agents perform comparably to "live web" agents, enabling reliable evaluations of models over time. We provide robust agent tooling and scaffolding to benchmark major LLMs as they are released, including "thinking" models like o3 and Gemini 2.5 Pro. We include automated evaluations of the lengthy agent traces to report progress over time in hallucinations, tool use, and forgetting. Finally, we evaluate the major web research products branded as "Deep Research", "Deep Search", "Search", or "Research." Results are available on a public leaderboard at https://drb.futuresearch.ai/.

Abstract of Deep Research Bench: Evaluating AI Web Research Agents (arXiv:2506.06287, CC0)
22
Entries on its own leaderboard, 21 models we track

Claude Opus 4.6 leads with 0.55 among models released by May 2026

Best score per entry from the benchmark's own leaderboard, read on November 12th, 2026. Where a model ran with several settings or harnesses, we keep its best. Every run is in the next chart. Names in gray aren't in our model list: fine-tunes, agent systems, products and older models. They don't count toward leaders.

The newest model with a score here came out in May 2026, and nobody has tested a newer one. Among the 20 highest-ranked models on the capability index, these have no score yet:

22 entries. Scroll sideways for the rest.

  1. 0.55 Claude Opus 4.6
  2. 0.55 Claude Sonnet 4.6
  3. 0.55 Claude Opus 4.5
  4. 0.54 GPT-5.5
  5. 0.50 Claude Opus 4.8
  6. 0.50 Gemini 3 Flash
  7. 0.50 GPT-5
  8. 0.48 Claude Opus 4.1
  9. 0.48 Gemini 3.1 Pro
  10. 0.47 Claude Sonnet 4.5
  11. 0.47 Grok 4
  12. 0.47 Claude Opus 4
  13. 0.47 Claude Sonnet 4
  14. 0.46 Gemini 3 Pro
  15. 0.46 Claude Haiku 4.5
  16. 0.45 o3
  17. 0.43 GPT-5.1
  18. 0.41 Gemini 2.5 Pro (snapshot unspecified)
  19. 0.41 GPT-5.2
  20. 0.37 Gemini 3.1 Flash-Lite
  21. 0.36 GPT-5.4 Mini
  22. 0.35 GPT-5.4
  1. 1 Claude Opus 4.6 high 0.55
  2. 2 Claude Sonnet 4.6 high 0.55
  3. 3 Claude Opus 4.5 high 0.55
  4. 4 GPT-5.5 high 0.54
  5. 5 Claude Opus 4.8 high 0.50
  6. 6 Gemini 3 Flash low 0.50
  7. 7 GPT-5 low 0.50
  8. 8 Claude Opus 4.1 medium 0.48
  9. 9 Gemini 3.1 Pro high 0.48
  10. 10 Claude Sonnet 4.5 low 0.47
  11. 11 Grok 4 0.47
  12. 12 Claude Opus 4 medium 0.47
  13. 13 Claude Sonnet 4 medium 0.47
  14. 14 Gemini 3 Pro low 0.46
  15. 15 Claude Haiku 4.5 low 0.46
  16. 16 o3 0.45
  17. 17 GPT-5.1 low 0.43
  18. 18 Gemini 2.5 Pro (snapshot unspecified) dynamic 0.41
  19. 19 GPT-5.2 low 0.41
  20. 20 Gemini 3.1 Flash-Lite low 0.37
  21. 21 GPT-5.4 Mini low 0.36
  22. 22 GPT-5.4 low 0.35

Source: Official Deep Research Bench (FutureSearch) leaderboard

#EntryScoreSettingDateSource
1 Claude Opus 4.6 0.55 high Feb 2026 Official leaderboard
2 Claude Sonnet 4.6 0.55 high Feb 2026 Official leaderboard
3 Claude Opus 4.5 0.55 high Nov 2025 Official leaderboard
4 GPT-5.5 0.54 high Apr 2026 Official leaderboard
5 Claude Opus 4.8 0.50 high May 2026 Official leaderboard
6 Gemini 3 Flash 0.50 low Dec 2025 Official leaderboard
7 GPT-5 0.50 low Aug 2025 Official leaderboard
8 Claude Opus 4.1 0.48 medium Aug 2025 Official leaderboard
9 Gemini 3.1 Pro 0.48 high Feb 2026 Official leaderboard
10 Claude Sonnet 4.5 0.47 low Sep 2025 Official leaderboard
11 Grok 4 0.47 Jul 2025 Official leaderboard
12 Claude Opus 4 0.47 medium May 2025 Official leaderboard
13 Claude Sonnet 4 0.47 medium May 2025 Official leaderboard
14 Gemini 3 Pro 0.46 low Nov 2025 Official leaderboard
15 Claude Haiku 4.5 0.46 low Oct 2025 Official leaderboard
16 o3 0.45 Apr 2025 Official leaderboard
17 GPT-5.1 0.43 low Nov 2025 Official leaderboard
18 Gemini 2.5 Pro (snapshot unspecified) 0.41 dynamic Official leaderboard
19 GPT-5.2 0.41 low Dec 2025 Official leaderboard
20 Gemini 3.1 Flash-Lite 0.37 low Mar 2026 Official leaderboard
21 GPT-5.4 Mini 0.36 low Mar 2026 Official leaderboard
22 GPT-5.4 0.35 low Mar 2026 Official leaderboard

Research score vs. cost

The line joins the runs that no other run beats on both measures. We redrew this chart from the numbers on the official leaderboard.

0.4 0.5 0.6 $0$0.5$1$1.5$2 Cost per task ($), lower is better Score o3: Score 0.45, cost per task $0.12 Claude Opus 4 (medium): Score 0.47, cost per task $1.64 Claude Sonnet 4 (medium): Score 0.47, cost per task $0.45 Grok 4: Score 0.47, cost per task $0.49 Claude Opus 4.1 (medium): Score 0.48, cost per task $1.2 GPT-5 (low): Score 0.5, cost per task $0.25 GPT-5 (minimal): Score 0.49, cost per task $0.24 GPT-5 (medium): Score 0.49, cost per task $0.35 GPT-5 (high): Score 0.48, cost per task $0.39 Claude Sonnet 4.5 (low): Score 0.47, cost per task $0.36 Claude Haiku 4.5 (low): Score 0.46, cost per task $0.1 Claude Opus 4.5 (high): Score 0.55, cost per task $0.46 Claude Opus 4.5 (low): Score 0.54, cost per task $0.43 GPT-5.2 (low): Score 0.41, cost per task $0.26 Gemini 3 Flash (minimal): Score 0.49, cost per task $0.1 Gemini 3 Flash (high): Score 0.48, cost per task $0.14 Claude Opus 4.6 (medium): Score 0.53, cost per task $0.36 Claude Sonnet 4.6 (low): Score 0.5, cost per task $0.27 Gemini 3.1 Pro (high): Score 0.48, cost per task $0.23 Gemini 3.1 Flash-Lite (dynamic): Score 0.36, cost per task $0.03 GPT-5.4 (low): Score 0.35, cost per task $0.14 GPT-5.4 Mini (low): Score 0.36, cost per task $0.04 GPT-5.5 (medium): Score 0.5, cost per task $0.31 GPT-5.5 (low): Score 0.49, cost per task $0.23 Claude Opus 4.8 (high): Score 0.5, cost per task $0.35 Claude Opus 4.8 (low): Score 0.49, cost per task $0.24 Claude Opus 4.8 (medium): Score 0.47, cost per task $0.29 Gemini 2.5 Pro (snapshot unspecified) (dynamic): Score 0.41, cost per task $0.11. Not in our model list GPT-5.1 (low): Score 0.43, cost per task $0.04 Gemini 3 Flash (low): Score 0.5, cost per task $0.1 Claude Opus 4.6 (high): Score 0.55, cost per task $0.53 Claude Opus 4.6 (low): Score 0.51, cost per task $0.24 Claude Sonnet 4.6 (high): Score 0.55, cost per task $0.46 Gemini 3.1 Pro (low): Score 0.45, cost per task $0.09 Gemini 3.1 Flash-Lite (low): Score 0.37, cost per task $0.02 GPT-5.5 (high): Score 0.54, cost per task $0.36 Claude Opus 4.6 (high)Claude Sonnet 4.6 (high)GPT-5.5 (high)Claude Opus 4.6 (low)Gemini 3 Flash (low)Gemini 3.1 Pro (low)GPT-5.1 (low)Gemini 3.1 Flash-Lite (low) 0.4 0.5 0.6 $0$0.5$1$1.5$2 Cost per task ($), lower is better Score o3: Score 0.45, cost per task $0.12 Claude Opus 4 (medium): Score 0.47, cost per task $1.64 Claude Sonnet 4 (medium): Score 0.47, cost per task $0.45 Grok 4: Score 0.47, cost per task $0.49 Claude Opus 4.1 (medium): Score 0.48, cost per task $1.2 GPT-5 (low): Score 0.5, cost per task $0.25 GPT-5 (minimal): Score 0.49, cost per task $0.24 GPT-5 (medium): Score 0.49, cost per task $0.35 GPT-5 (high): Score 0.48, cost per task $0.39 Claude Sonnet 4.5 (low): Score 0.47, cost per task $0.36 Claude Haiku 4.5 (low): Score 0.46, cost per task $0.1 Claude Opus 4.5 (high): Score 0.55, cost per task $0.46 Claude Opus 4.5 (low): Score 0.54, cost per task $0.43 GPT-5.2 (low): Score 0.41, cost per task $0.26 Gemini 3 Flash (minimal): Score 0.49, cost per task $0.1 Gemini 3 Flash (high): Score 0.48, cost per task $0.14 Claude Opus 4.6 (medium): Score 0.53, cost per task $0.36 Claude Sonnet 4.6 (low): Score 0.5, cost per task $0.27 Gemini 3.1 Pro (high): Score 0.48, cost per task $0.23 Gemini 3.1 Flash-Lite (dynamic): Score 0.36, cost per task $0.03 GPT-5.4 (low): Score 0.35, cost per task $0.14 GPT-5.4 Mini (low): Score 0.36, cost per task $0.04 GPT-5.5 (medium): Score 0.5, cost per task $0.31 GPT-5.5 (low): Score 0.49, cost per task $0.23 Claude Opus 4.8 (high): Score 0.5, cost per task $0.35 Claude Opus 4.8 (low): Score 0.49, cost per task $0.24 Claude Opus 4.8 (medium): Score 0.47, cost per task $0.29 Gemini 2.5 Pro (snapshot unspecified) (dynamic): Score 0.41, cost per task $0.11. Not in our model list GPT-5.1 (low): Score 0.43, cost per task $0.04 Gemini 3 Flash (low): Score 0.5, cost per task $0.1 Claude Opus 4.6 (high): Score 0.55, cost per task $0.53 Claude Opus 4.6 (low): Score 0.51, cost per task $0.24 Claude Sonnet 4.6 (high): Score 0.55, cost per task $0.46 Gemini 3.1 Pro (low): Score 0.45, cost per task $0.09 Gemini 3.1 Flash-Lite (low): Score 0.37, cost per task $0.02 GPT-5.5 (high): Score 0.54, cost per task $0.36 Claude Opus 4.6 (high)Claude Sonnet 4.6 (high)GPT-5.5 (high)Claude Opus 4.6 (low)Gemini 3 Flash (low)Gemini 3.1 Pro (low)GPT-5.1 (low)Gemini 3.1 Flash-Lite (low)
ModelSettingScoreCost per task
Claude Opus 4.6 high 0.55 $0.53
Claude Sonnet 4.6 high 0.55 $0.46
Claude Opus 4.5 high 0.55 $0.46
GPT-5.5 high 0.54 $0.36
Claude Opus 4.5 low 0.54 $0.43
Claude Opus 4.6 medium 0.53 $0.36
Claude Opus 4.6 low 0.51 $0.24
Claude Sonnet 4.6 low 0.5 $0.27
Claude Opus 4.8 high 0.5 $0.35
Gemini 3 Flash low 0.5 $0.1
GPT-5 low 0.5 $0.25
GPT-5.5 medium 0.5 $0.31
Claude Opus 4.8 low 0.49 $0.24
Gemini 3 Flash minimal 0.49 $0.1
GPT-5 minimal 0.49 $0.24
GPT-5.5 low 0.49 $0.23
GPT-5 medium 0.49 $0.35
Claude Opus 4.1 medium 0.48 $1.2
GPT-5 high 0.48 $0.39
Gemini 3 Flash high 0.48 $0.14
Gemini 3.1 Pro high 0.48 $0.23
Claude Sonnet 4.5 low 0.48 $0.36
Claude Opus 4.8 medium 0.47 $0.29
Grok 4 0.47 $0.49
Claude Opus 4 medium 0.47 $1.64
Claude Sonnet 4 medium 0.47 $0.45
Claude Haiku 4.5 low 0.46 $0.1
o3 0.45 $0.12
Gemini 3.1 Pro low 0.45 $0.09
GPT-5.1 low 0.43 $0.04
Gemini 2.5 Pro (snapshot unspecified) dynamic 0.42 $0.11
GPT-5.2 low 0.41 $0.26
Gemini 3.1 Flash-Lite low 0.37 $0.02
Gemini 3.1 Flash-Lite dynamic 0.36 $0.03
GPT-5.4 Mini low 0.36 $0.04
GPT-5.4 low 0.35 $0.14

The best score, 0.55, costs $0.53 a task

Runs on the cost frontier score higher than any run that's cheaper. The costs are the leaderboard's own estimates, so compare them only within this benchmark.

OpenAI Google DeepMind Anthropic Cost frontier One model at each effort level
0.00 0.20 0.40 0.60 $0.01 $0.10 $1.00 $10 Cost per task, log scale Claude Opus 4.6, high Claude Opus 4.6, low Gemini 3.1 Pro, low Gemini 3.1 Flash-Lite, low 0.00 0.20 0.40 0.60 $0.01 $0.10 $1.00 $10 Cost per task, log scale Claude Opus 4.6, high Claude Opus 4.6, low

Source: Official Deep Research Bench (FutureSearch) leaderboard

ModelEffortHarnessScoreCost per taskOutput tokens
Claude Opus 4.6 high 0.55 $0.53
Claude Sonnet 4.6 high 0.55 $0.46
Claude Opus 4.5 high 0.55 $0.46
GPT-5.5 high 0.54 $0.36
Claude Opus 4.5 low 0.54 $0.43
Claude Opus 4.6 medium 0.53 $0.36
Claude Opus 4.6 low 0.51 $0.24
Claude Sonnet 4.6 low 0.50 $0.27
Claude Opus 4.8 high 0.50 $0.35
Gemini 3 Flash low 0.50 $0.10
GPT-5 low 0.50 $0.25
GPT-5.5 medium 0.50 $0.31
Claude Opus 4.8 low 0.49 $0.24
Gemini 3 Flash minimal 0.49 $0.10
GPT-5 minimal 0.49 $0.24
GPT-5.5 low 0.49 $0.23
GPT-5 medium 0.49 $0.35
Claude Opus 4.1 medium 0.48 $1.20
GPT-5 high 0.48 $0.39
Gemini 3 Flash high 0.48 $0.14
Gemini 3.1 Pro high 0.48 $0.23
Claude Sonnet 4.5 low 0.47 $0.36
Claude Opus 4.8 medium 0.47 $0.29
Grok 4 0.47 $0.49
Claude Opus 4 medium 0.47 $1.64
Claude Sonnet 4 medium 0.47 $0.45
Gemini 3 Pro low 0.46
Claude Haiku 4.5 low 0.46 $0.10
o3 0.45 $0.12
Gemini 3.1 Pro low 0.45 $0.09
GPT-5.1 low 0.43 $0.04
GPT-5.2 low 0.41 $0.26
Gemini 3.1 Flash-Lite low 0.37 $0.02
Gemini 3.1 Flash-Lite dynamic 0.36 $0.03
GPT-5.4 Mini low 0.36 $0.04
GPT-5.4 low 0.35 $0.14

7 models have held number one on the official leaderboard since its release in May 2025

Lab claims count from the day the card came out, independent runs from the model’s release.

OpenAI Anthropic xAI
Own leaderboard Not tested since May 2026
Claude Opus 4.6 0.55

Source: Official Deep Research Bench (FutureSearch) leaderboard

Scores over time

All the scores we have, one dot each, so a model tested at several settings appears more than once.

On the benchmark’s own leaderboard
0.40 0.50 0.60 2025 2026 0.40 0.50 0.60 2025 2026

Source: Official Deep Research Bench (FutureSearch) leaderboard

No lab model card reports it yet

We read the headline results of 522 launch posts from 52 labs, and none of them list this benchmark.

More in Tool use

All 144 in Tool use
  • AutomationBench In use Agent completes end-to-end business workflows inside a simulated company exposing dozens of REST APIs. Graded on resulting state. Gemini 4 Carbon 53.9%
  • MCP Atlas In use Multi-step workflows over real MCP servers. Scores tool-use competency on realistic tasks. Gemini 4 Argon 90.7%
  • BFCL V4 In use Models call functions and tools, including multi-turn agentic tasks with web search and memory. Checked against expected calls and outcomes. Qwen3.7-Max 75%
  • τ³-bench Banking In use Agents act as banking customer-service reps over many turns with a simulated user, following policy. Pass^1 on the final database state. Qwen 3.8 Max 55.1%
  • Toolathlon-Verified In use 108 long-horizon tasks across 32 apps through about 600 MCP tools. Execution-based checkers inspect final environment state. GLM-5.3-Flash 78.4%
  • DeepSearchQA In use Questions that need multi-step search and synthesis. Final answers graded against references. Kimi K3 95%