Menu
This weekThreadsBy practiceToolsArenaSignalsMethodMoversCompareBest by categoryBy yearFeature requestsBy modelEvolutionRatingsThe tableCorrectionsFor vendors
Updated

Home › Arena › Frontier models for legal work

Frontier models for legal work, cutoff Oct 26, 2026.

As of Oct 26, 2026, aiforlawyers.app counts 3473 lawyer voices that name Claude, ChatGPT or Gemini: Claude net +37 (band +33 to +41, 872 voices, High confidence); Gemini net +28 (band +15 to +39, 134 voices, Medium confidence); ChatGPT net -13 (band -16 to -10, 2467 voices, High confidence). Public benchmark scores are shown apart on this page and are not our score.

Download the PDF (cutoff Oct 26, 2026)Data (JSON)ArenaBy model

Cite this

aiforlawyers.app (AI Vortex), Frontier models for legal work, cutoff Oct 26, 2026, https://aiforlawyers.app/report/frontier-models, licensed CC BY 4.0.

For one figure: “As of Oct 26, 2026, aiforlawyers.app counts 3473 lawyer voices that name Claude, ChatGPT or Gemini: Claude net +37 (band +33 to +41, 872 voices, High confidence); Gemini net +28 (band +15 to +39, 134 voices, Medium confidence); ChatGPT net -13 (band -16 to -10, 2467 voices, High confidence). Public benchmark scores are shown apart on this page and are not our score.”

What lawyers say, our record

3473 voices name Claude, ChatGPT or Gemini as the main tool. Ordered by the floor of the 90% band, so a few happy voices cannot outrank hundreds. Net = like minus dislike per 100 voices. Data through Oct 26, 2026.

#Model familyVoicesLike / mixed / dislikeNet90% bandConfidence
1ClaudeAnthropic872492 / 103 / 169+37+33 to +41High
2GeminiGoogle13473 / 18 / 36+28+15 to +39Medium
3ChatGPTOpenAI2,467832 / 321 / 1,162−13−16 to −10High

By task, the same three

Net and 90% band where a family has 10+ voices on the task, otherwise the count only. Read the band and the voices, not the dot. Data through Oct 26, 2026.

TaskClaudeChatGPTGemini
Legal research+41 band +27 to +54 · 82 voices−34 band −39 to −28 · 623 voices+42 band +19 to +59 · 31 voices
Drafting+57 band +47 to +65 · 141 voices+22 band +16 to +28 · 573 voices+48 band +20 to +68 · 27 voices
Summarization+59 band +33 to +75 · 27 voices+5 band −8 to +18 · 114 voices+36 band −10 to +68 · 11 voices
Contract review+46 band +29 to +59 · 65 voices−11 band −29 to +9 · 57 voicesFew voices 3 of 10 needed, no band
Citation checking+16 band −12 to +41 · 31 voices−69 band −76 to −59 · 157 voicesFew voices 9 of 10 needed, no band
General use+28 band +22 to +34 · 455 voices−18 band −23 to −13 · 790 voices+11 band −11 to +33 · 45 voices

By version, when the post names it

Posts that name a version are filed under it too (the same data as By model). Posts naming it counts every voice that mentions the version; the net uses only voices whose main tool is that family. Data through Oct 26, 2026.

Version named in the postPosts naming itVoices about the familyNet90% bandLatest
Claude Opus 4.xClaude2715+47+12 to +70Oct 6, 2026
Claude 3.5 SonnetClaude1210+100+61 to +99Jun 7, 2026
Claude Opus (unversioned)Claude117–Few voices 7 of 10, no bandSep 3, 2026
Claude Sonnet 4.xClaude105–Few voices 5 of 10, no bandAug 5, 2026
Claude 3.7 SonnetClaude32–Few voices 2 of 10, no bandMay 4, 2025
GPT-4ChatGPT9565+12−4 to +28Sep 30, 2026
GPT-4oChatGPT4023+43+13 to +65Sep 30, 2026
GPT-5ChatGPT5723+13−16 to +39Sep 29, 2026
o3ChatGPT2011+82+37 to +95Sep 13, 2025
GPT-3.5ChatGPT77–Few voices 7 of 10, no bandJun 6, 2026
o1ChatGPT107–Few voices 7 of 10, no bandFeb 18, 2026
Gemini 2.xGemini2914+57+22 to +76Aug 5, 2026
BardGemini2711−36−69 to +11May 14, 2025
Gemini 3Gemini2210+80+41 to +91Sep 17, 2026
Gemini Advanced / ProGemini127–Few voices 7 of 10, no bandAug 24, 2026

The newest models, and how many voices name them

Release dates are from each lab’s own page (read 2026-10-23). A count is not a score.

Current modelReleased (lab’s own page)Voices naming it
Claude Haiku 5.5Anthropic2026-10-07none yet
Claude Sonnet 5.5Anthropic2026-09-28none yet
Claude Opus 5.5Anthropic2026-09-224 voices
Claude Fable 5.1 flagshipAnthropic2026-09-012 voices
Claude Opus 5Anthropic2026-07-2410 voices
Claude Sonnet 5Anthropic2026-06-302 voices
Claude Fable 5Anthropic2026-06-0910 voices
GPT-6.1 SolOpenAI2026-09-29none yet
GPT-6 SolOpenAI2026-09-22none yet
GPT-6 LunaOpenAI2026-09-22none yet
GPT-6 Astra flagshipOpenAI2026-09-032 voices
GPT-5.6 SolOpenAI2026-07-095 voices
Gemini 4 Argon flagshipGoogle2026-10-13none yet
Gemini 3.8 FlashGoogle2026-09-02none yet
Gemini 3.1 ProGoogle2026-02-194 voices

Public test scores, next to the record not our score

What public tests measure for the same three labs, read on 2026-10-23. These are question sets graded by their publishers, not lawyers’ experience. We copy their figures with their dates and never mix them with the net above. A dash means the board did not list that model when we read it; we never estimate a missing score.

ModelReleasedLegal Research BenchLegalBenchAA IntelligenceArena rankHarvey LAB
Claude Haiku 5.5Anthropic2026-10-0743.27%–43–2.8%lab page
Claude Sonnet 5.5Anthropic2026-09-2848.08%–56#322.8%lab page
Claude Opus 5.5Anthropic2026-09-2250.48%–58#24.2%lab page
Claude Fable 5.1 flagshipAnthropic2026-09-0155.29%88.51%53#66.4%lab page
Claude Opus 5Anthropic2026-07-2455.29%86.97%–#13–lab page
Claude Sonnet 5Anthropic2026-06-3041.83%83.92%–#57–lab page
Claude Fable 5Anthropic2026-06-0949.52%88.56%–#4–lab page
GPT-6.1 SolOpenAI2026-09-2938.46%–52#216.9%lab page
GPT-6 SolOpenAI2026-09-2228.85%––#673.6%lab page
GPT-6 LunaOpenAI2026-09-2230.29%–38#913.3%lab page
GPT-6 Astra flagshipOpenAI2026-09-0339.42%–53#358.6%lab page
GPT-5.6 SolOpenAI2026-07-0948.08%86.97%–#20–lab page
Gemini 4 Argon flagshipGoogle2026-10-1354.81%88.3%53#13.9%lab page
Gemini 3.8 FlashGoogle2026-09-0238.94%86.99%41#80%lab page
Gemini 3.1 ProGoogle2026-02-1920.67%87.4%30#170%lab page

The boards, with their own dates

Caveats from the publishers’ own pages

legal reasoning benchmark · Vals AI

Vals Legal Research Bench (agentic US legal research leaderboard)

published 2026-10-22 (page shows Updated 10/22/2026) · read 2026-10-22

Caveat. Tests raw models in an agent harness, not legal products like Harvey or CoCounsel; private test set; 'All-Pass' means every rubric item must pass. Top 3 are a three-way tie, so rank order is arbitrary.

https://www.vals.ai/benchmarks/legal_research

legal reasoning benchmark · Vals AI

LegalBench leaderboard on Vals (public LegalBench tasks)

published 2026-10-13 (page shows Updated 10/13/2026) · read 2026-10-22

Caveat. Mostly classification and rule-recall tasks on public data, so contamination is possible and the top scores are nearly saturated; measures models, not legal products.

https://www.vals.ai/benchmarks/legal_bench

general model leaderboard · LMArena

LMArena (arena.ai) Text Arena overall leaderboard

published 2026-10-15 (leaderboard date shown; archive snapshot 2026-10-17) · read 2026-10-22

Caveat. Preference votes on general chat, not legal accuracy; the top 3 are within overlapping uncertainty bands (±9, ±11, ±7); #1 is flagged preliminary; archive copy, not fetched live today.

https://web.archive.org/web/20261017054907/https://arena.ai/leaderboard/chat/text

general model leaderboard · Artificial Analysis

Artificial Analysis Intelligence Index v4.3.2 (LLM leaderboard)

published 2026-10-22 (changelog on the page lists entries through 22 Oct) · read 2026-10-22

Caveat. Mostly coding, agentic and science tasks, nothing legal; scores depend on reasoning-effort setting; the page text I read does not print an explicit as-of date other than its changelog.

https://artificialanalysis.ai/leaderboards/models

legal-specific vendor test · Harvey (benchmark) / Mercor (leaderboard)

Harvey Legal Agent Benchmark (LAB) leaderboard, hosted by Mercor

published 2026-05-06 (launch per Harvey blog coverage); leaderboard date not shown on page · read 2026-10-22

Caveat. Built by Harvey (a vendor); the leaderboard lists foundation models, not Harvey's own product; leaderboard publication date not shown and the top-5 error bars overlap heavily; the leaderboard page counts 1,749 tasks versus 'more than 1,200' reported at launch, so the set has grown.

https://www.mercor.com/apex/partner-benchmarks/harvey-lab-leaderboard/

More benchmarks and all 31 models: Arena, frontier models.

Method and cutoff

  • What a voice is. One public post by a lawyer or legal professional that states an experience or a verdict about one AI tool, read by a language model with one fixed schema (tool, task, sentiment, main complaint). Sources: Reddit, Hacker News, Bluesky, Lemmy, Mastodon, Stack Exchange; posts since Jan 1, 2023. Claude, ChatGPT and Gemini here are the tools as lawyers name them, not one model version.
  • Net and band. Net = (like − dislike) per 100 voices. The 90% band comes from resampling the same voices 4,000 times. A task or version cell is rated from 10 voices; under that it says “Few voices” with its count and no band. That is not a negative finding.
  • Lag. Lawyers keep writing about older names (GPT-4, Claude Opus 4.x) after the labs ship new models, so the record describes the families, and the newest models have few voices of their own. The “Voices naming it” table counts them and gives no score.
  • Cutoff. Data through Oct 26, 2026; the latest voice on these three is from Oct 24, 2026. Outside boards were read on 2026-10-23. The page is rebuilt with the daily refresh.
  • Not in our numbers. The scores in the last section are other people’s tests, with their own publisher, date and caveat. They are never added to, averaged with or ranked against our net. Public posts are not a survey of all lawyers. Independence policy: no vendor pays for a place or a number. Full method: how we count.

Related on AI Vortex

Use these numbers: the counts, rates and bands on this report are free to reuse under CC BY 4.0. Credit “aiforlawyers.app (AI Vortex)” with the date and a link to the page. Quotes belong to the people who wrote them: quote them with the link to the original post. Machine-readable: index.json. How to cite.