Skip to content
← AI by field
Analysts, consultants and researchers

AI for research and analysis, right now

Finding and checking information across many sources, and writing up the analysis.

Moving: very fastUpdated Nov 12, 2026 from 7 sources
Open research in Spezia Free account. No card.
Where AI is in research now

AI research agents now score 73-92% on hard research benchmarks, led by Claude Fable 5.2 and Kimi K3.1 Pro.

What models can do now

  • Claude Fable 5.2 leads the WANDR wide-search research benchmark at 82.1%.[4]
  • LongCat-2.5 leads the ResearchRubrics deep-research benchmark at 81.0%.[3]
  • Claude Opus 4.6 holds the top DeepResearch Bench score at 55.3%.[5]
  • Opus 5 leads the Mercor-run DeepResearch Bench II snapshot at 56.10%.[6]
  • Kimi K3.1 Pro set the BrowseComp record at 92.1%.[1]
  • Claude Fable 5.2 set the Humanity's Last Exam record at 73.2%.[2]

Where they still fall short

  • Grok 4.7 made a visual error in a scientific image reasoning task despite extensive inspection and annotation.[7]

How fast it is moving

The record on this field's benchmarks over the last year, from Spezia's model leaderboard.

BrowseComp

Questions whose answers are hidden deep on the web and take many searches and pages to track down.

Record: 92.1%, up from 60.2% a year ago (1.5×) (Kimi K3.1 Pro)

0%25%50%75%100%Kimi K3.1 ProGPT-5.5 Pro
Jul 25Oct 25Jan 26Apr 26Jul 26Oct 26

the record so far set a record other models

DeepSearchQA

Questions that need many searches and several steps of reasoning across sources to answer.

Record: 95%, up from 77.1% 6 months ago (Kimi K3)

0%25%50%75%100%Claude Opus 4.6Kimi K3
Apr 26Jul 26Oct 26

the record so far set a record other models

The best models for research now

On this field's benchmarks alone.

  1. 1Claude Opus 5
  2. 2Muse Spark 1.1
  3. 3Kimi K3
  4. 4GPT-5.5 Pro
  5. 5Seed 2.1 Pro
  6. 6Kimi K3.1 Pro

Milestones

What became possible, newest first, each marked by how firm it is. Open one for the words it rests on.

Showndemonstrated with published results, a paper or an independent checkClaimedsaid by the company or people who built it, not yet checked by othersReportedreported by others, not confirmed by the company or a published resultNot publicdone by a model that is not released to the public

  1. Nov 11, 2026ShownLongCat-2.5
    LongCat-2.5 leads the ResearchRubrics benchmark at 81.0%.
    “LongCat-2.5 leads the table at 81.0%” ResearchRubrics Leaderboard & Scores — November 2026
  2. Nov 9, 2026ShownClaude Fable 5.2
    Claude Fable 5.2 leads the WANDR benchmark at 82.1%.
    “Claude Fable 5.2 leads the table at 82.1%” WANDR Leaderboard & Scores — November 2026
  3. Oct 28, 2026ShownOpus 5
    Opus 5 leads the Mercor-run DeepResearch Bench II snapshot at 56.10%.
    “Opus 5 leads the public snapshot at 56.10%” DeepResearch Bench II (Mercor) Leaderboard — November 2026
  4. Feb 5, 2026ShownClaude Opus 4.6
    Claude Opus 4.6 achieved 55.3% on DeepResearch Bench.
    “Claude Opus 4.6 holds the top DeepResearch Bench score at 55.3%” DeepResearch Bench leaderboard (Nov 2026) | BenchLeader
Every milestone, story and model Pick a model to see what it can do in research.

Sources

  1. [1]BrowseComp: best result 92.1%, up from 60.2% a year ago (1.5×)Leaderboard · 2026-10-23
  2. [2]Humanity's Last Exam: best result 73.2%, up from 40% a year ago (1.8×)Leaderboard · 2026-11-09
  3. [3]ResearchRubrics Leaderboard & Scores — November 2026Web · 2026-11-09
  4. [4]WANDR Leaderboard & Scores — November 2026Web · 2026-11-07
  5. [5]DeepResearch Bench leaderboard (Nov 2026) | BenchLeaderWeb · 2026-10-31
  6. [6]DeepResearch Bench II (Mercor) Leaderboard — November 2026Web
  7. [7]VIALS-AA: different paths to scientific image reasoningWeb · 2026-11-05

Other fields