Skip to content
← AI by field
Lawyers, legal teams and firms

AI for law, right now

Research, drafting, contract review and long legal tasks with citations.

Moving: fastUpdated Nov 12, 2026 from 9 sources
Open law in Spezia Free account. No card.
Where AI is in law now

AI for law is advancing rapidly, with Claude Sonnet 5.5 leading law benchmarks and Gemini 4 Argon topping agentic legal work, but real-world contract review remains below 60%.

What models can do now

  • Claude Sonnet 5.5, Claude Opus 5.5, and Claude Opus 5 are the best models on law benchmarks.[2]
  • Gemini 4 Argon leads Harvey's Legal Agent Benchmark with a 27.1% score, ahead of Muse Spark Open at 25.0%.[3]
  • On LegalBench, Claude Fable 5 achieves 88.6%, with Claude Fable 5.1 close behind at 88.5%.[9]
  • In an independent study of 350 attorney-scored contract tasks, Grok led at 58.3%, just ahead of Claude Opus 5.5 and Kimi K3.[6]
  • Qwen3.8 Max leads the Access to Justice benchmark at 76.37%, followed by Claude Sonnet 5.5 at 74.94%.[8]
  • Parallel for Legal's web search evaluation shows accuracy up to 97.5% on legal hallucination questions, with OpenAI leading.[7]

Where they still fall short

  • No AI model has cracked 60% on real-world contract review, according to an independent study.[6]
  • On Harvey's Legal Agent Benchmark, the best score is only 27.1%, indicating that agentic legal work remains challenging.[3]
  • On the Legal Agent Benchmark, the best result is 13.3%, up from 0% six months ago, showing limited progress on complex legal tasks.[1]

What is expected next

  • Crosby's founder Ryan Daniels said the goal, ideally by the end of next year, is legal work done by agents without lawyers.[5]
  • Harvey said context graphs improved Legal Agent Bench rubric pass rates by up to 30% and cut task cost by about 50% in company tests.[4]

How fast it is moving

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

Legal Agent Benchmark

Long, complex legal tasks graded against rubrics written by legal experts, covering facts, conclusions, citations and structure.

Record: 13.3%, up from 0% 6 months ago (Claude Fable 5)

0%4%8%12%16%Claude Fable 5Claude Opus 4.7
Jan 26Apr 26Jul 26Oct 26

the record so far set a record other models

APEX-Agents

Long tasks written by investment banking analysts, management consultants and corporate lawyers from their own work, done across files and apps in a realistic work setup.

Record: 84.2%, up from 21.3% a year ago (4×) (Gemini 4 Carbon)

0%25%50%75%100%Gemini 4 CarbonClaude Sonnet 5.5
Oct 25Jan 26Apr 26Jul 26Oct 26

the record so far set a record other models

The best models for law now

On this field's benchmarks alone.

  1. 1Claude Sonnet 5.5
  2. 2Claude Opus 5.5
  3. 3Claude Opus 5
  4. 4Claude Fable 5
  5. 5Claude Sonnet 5
  6. 6GPT-5.5

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 10, 2026ShownQwen3.8 Max
    Qwen3.8 Max leads the Access to Justice benchmark with 76.37% rubric criteria met.
    “Qwen3.8 Max leads the public snapshot at 76.37%, followed by Claude Sonnet 5.5 (74.94%) and Muse Wonder Full (74.68%)” Access to Justice Leaderboard & Scores — November 2026
  2. Nov 10, 2026Reported
    EU advocate general opinion says AI training may be lawful but infringing AI output can violate copyright.
    “EU advocate general opinion says AI training may be lawful but infringing AI output can violate copyright” EU advocate general opinion says AI training may be lawful but infringing AI out
  3. Nov 8, 2026Reported
    California's SB 574 requires attorneys to verify AI-generated citations and disclose generative AI use in court filings.
    “California's SB 574 requires attorneys to verify AI-generated citations and disclose generative AI use in court filings.” California's SB 574 requires attorneys to verify AI-generated citations and disc
  4. Nov 7, 2026Reported
    Crosby, an AI-powered law firm, bought $5 million of AI liability insurance and removed lawyer supervision from gray-area tasks, about 20% of its work.
    “Crosby, an AI-powered law firm whose agents handle most of its contract review under about 60 supervising lawyers, bought $5 million of AI liability insurance from Testudo” AI law firm Crosby buys $5 million liability policy to cover its AI agents
  5. Nov 5, 2026Reported
    DOJ launches Super Intelligence Task Force to prosecute AI misuse.
    “The U.S. Department of Justice announced a new Super Intelligence Task Force on November 5” DOJ launches Super Intelligence Task Force to prosecute AI misuse and absorb pri
  6. Oct 30, 2026ShownGemini 4 Argon
    Gemini 4 Argon achieved 27.1% on Harvey's Legal Agent Benchmark v1.1, the top score among 33 model configurations.
    “Gemini 4 Argon holds the top Harvey's Legal Agent Benchmark v1.1 score at 27.1% (Vals AI, 30 Oct 2026)” Harvey's Legal Agent Benchmark leaderboard (Nov 2026)
Every milestone, story and model Pick a model to see what it can do in law.

Sources

  1. [1]Legal Agent Benchmark: best result 13.3%, up from 0% 6 months agoLeaderboard · 2026-06-09
  2. [2]Best models on law benchmarksLeaderboard · 2026-11-12
  3. [3]Harvey's Legal Agent Benchmark leaderboard (Nov 2026)Web · 2026-11-01
  4. [4]Harvey details context graphs for legal agents; cites early lab gains - NewsWeb · 2026-10-18
  5. [5]AI law firm Crosby buys $5 million liability policy to cover its AI agentsWeb · 2026-11-07
  6. [6]What’s New in Legal AI? Anthropic’s Legal Push, a Contract Review Reality Check, and More - Legal Talk NetworkWeb · 2026-11-09
  7. [7]Parallel for Legal: web search for legal AIWeb · 2026-11-05
  8. [8]Access to Justice Leaderboard & Scores — November 2026Web · 2026-11-08
  9. [9]LegalBench leaderboard (Nov 2026) | BenchLeaderWeb · 2026-11-03

Other fields