Best AI for summarising documents
Scored 99/100 on our 12-task summarising suite — a single point ahead of GPT-5.5 (98) — effectively level.
Claude Opus 4.8 is the engine inside Claude. Go to claude.ai ↗ — the free tier is fine to start. Paid plans start at $20/month (pro plan, vendor’s own price). The free tier is Claude’s, not a promise about this exact model — we haven’t verified which plan carries it. Not fussed about the last point or two? Any of the top 2 here will serve you well.
The failure mode that matters in summarising is not dullness, it is quiet invention — a caveat dropped, a number changed, a decision asserted that nobody made. Every task here is scored on what survives: the limitation, the four figures, the disagreement stated fairly. One task is a two-line email with nothing to summarise, to see which models pad rather than say so.
updated 13 Aug 2026 · tested by Robert Prime · re-ranks automatically when a new run lands
- All 19 sat the identical 12-task suite — same tasks, same order, one attempt each.
- Every answer marked blind: judges are not told which entrant wrote it.
- Three judges per answer, each from a competing lab. A panel never includes the entrant's own lab.
- Scored 0–10 against a fixed rubric. Answers breaking a task's explicit rules are capped by machine, not by opinion.
- Ties broken by fewest rule breaches, then lowest measured cost — published, not editorial.
- Nobody pays for placement. Last computed 13 Aug 2026.
Every score below links to the raw file behind it: every task, the entrant’s real answer, and all three judges’ marks.
| # | Model | Our score |
|---|---|---|
| 1 | Claude Opus 4.8 | 99/100 |
| 2 | GPT-5.5 | 98/100 |
| 3 | GPT-5.6 Lunalatest | 97/100 |
| 4 | GPT-5.6 Terra | 96/100 |
| 5 | GPT-5.3-Codex | 96/100 |
| 6 | DeepSeek V4 Pro | 95/100 |
| 7 | GPT-5.6 Sol | 95/100 |
| 8 | Claude Sonnet 5 | 93/100 |
| 9 | Claude Opus 4.6 | 93/100 |
| 10 | Claude Fable 5 | 93/100 |
| 11 | Grok 4.5 | 92/100 |
| 12 | GLM 5.2 | 91/100 |
| 13 | Gemini 3.1 Pro Preview | 90/100 |
| 14 | Kimi K3 | 90/100 |
| 15 | Gemini 3.5 Flash | 89/100 |
| 16 | Qwen3.7 Max | 87/100 |
| 17 | Gemini 3.1 Flash Lite | 85/100 |
| 18 | DeepSeek V4 Flash | 83/100 |
| 19 | Mistral Medium 3.5 | 83/100 |
“API cost” is what software developers pay to build on a model — ignore it if you just use the website. Each model answers each task once. Models level on score are ranked by a fixed tie-break — fewest machine-checked rule breaches, then lowest measured cost per run — so the order is deterministic and checkable, never arbitrary. Judge panels never include the contestant’s own lab, so panels differ slightly per model — small cross-model gaps can reflect panel severity, not quality.
Made by Anthropic. You use it inside Claude — nothing to install.
Strongest showing: “One-sentence gist” — scored 10/10 by the panel. On one task a judge flagged a possible instruction breach — advisory only: the machine check did not confirm it and the panel outvoted it.
“The response perfectly condenses the original text into a single, clear sentence of 23 words, capturing all key facts while strictly adhering to the constraints.”— google/gemini-3.1-pro-preview, judging blind · full receipts ↓
12 tasks · 13 Aug 2026 · judges gemini-3.1-pro-preview, gpt-5.6-terra, grok-4.5 · API $5 in / $25 out per 1M tokens · full model page →
Made by OpenAI. You use it inside ChatGPT — nothing to install.
Strongest showing: “Summary for a specific reader” — scored 10/10 by the panel.
“The response perfectly tailors the summary to a finance director by explicitly highlighting cost implications and risks. It captures all key financial points from the source text and easily meets the under-70-words constraint (44 words).”— google/gemini-3.1-pro-preview, judging blind · full receipts ↓
12 tasks · 13 Aug 2026 · judges claude-sonnet-5, gemini-3.1-pro-preview, grok-4.5 · API $5 in / $30 out per 1M tokens · full model page →
Made by OpenAI — their newest model. You use it inside ChatGPT — nothing to install.
Strongest showing: “Meeting notes to actions” — scored 10/10 by the panel. Weakest: “Do not invent” at 9/10. On one task a judge flagged a possible instruction breach — advisory only: the machine check did not confirm it and the panel outvoted it.
“The response perfectly extracts all action items, accurately assigns owners (or 'unassigned'), formats them as a bulleted list, and includes no extraneous text, following all instructions flawlessly.”— google/gemini-3.1-pro-preview, judging blind · full receipts ↓
12 tasks · 13 Aug 2026 · judges claude-sonnet-5, gemini-3.1-pro-preview, grok-4.5 · API $0.2 in / $1.2 out per 1M tokens · full model page →
See it for yourself
One task from the suite, and what the leader, the middle of the table and the bottom actually wrote. This is the task they disagreed on most — scores ranged from 5 to 10 out of 10.
Reduce to ONE sentence of no more than 25 words: 'After eighteen months of testing across four hospitals, the trial found that the new triage system reduced average waiting times by 22 minutes, although staff reported it added roughly ten minutes of administrative work per shift and two sites abandoned it before the trial ended.'
An 18-month, four-hospital trial cut waiting times by 22 minutes but added ten minutes of admin per shift, and two sites quit early.
Judge: The response perfectly condenses the original text into a single, clear sentence of 23 words, capturing all key facts while strictly adhering to the constraints.
An eighteen-month, four-hospital trial found the new triage system cut waiting times by 22 minutes, despite added administrative work and two sites abandoning it. (24 words)
Judge: Capped in code — the answer broke a checkable rule (26 words against a maximum of 25). The response successfully condenses the text, but includes an extra fragment '(24 words)' which violates the constraint of providing ONLY one sentence of no more than 25 words (total words = 27).
The 18-month trial showed the new triage system cut wait times by 22 minutes but added 10 minutes of admin work per shift.
Judge: Meets length constraints and summarizes the main trade-off, but omits the crucial detail that two sites abandoned the system, reducing its factual completeness.
Trimmed to fit. Every full answer, every judge’s score: Claude Opus 4.8, Claude Fable 5, Mistral Medium 3.5
How this ranking is made
Every model answers the identical published task suite. Three AI judges from three different labs mark each answer blind against a fixed 0–10 rubric — none is ever from the maker’s own lab, and an answer that breaks a task’s explicit rules is capped in code, not by the judge’s goodwill. The score is the average, out of 100.
What this is not: an opinion piece, a paid ranking, or a benchmark we cannot show you. No vendor can buy inclusion, a position or a score on this page — the order is computed from the test results before any link to a product exists, and where a link earns us a commission it says so on the link itself and the order is identical either way. Every score links its raw outputs and judge verdicts. The full protocol · How we make money · receipts: Claude Opus 4.8, GPT-5.5, GPT-5.6 Luna, GPT-5.6 Terra, GPT-5.3-Codex, DeepSeek V4 Pro, GPT-5.6 Sol, Claude Sonnet 5, Claude Opus 4.6, Claude Fable 5, Grok 4.5, GLM 5.2, Gemini 3.1 Pro Preview, Kimi K3, Gemini 3.5 Flash, Qwen3.7 Max, Gemini 3.1 Flash Lite, DeepSeek V4 Flash, Mistral Medium 3.5
Questions people ask
What is the best AI for summarising documents in 2026?
Claude Opus 4.8 leads our tested ranking with 99/100 on our 12-task summarising suite, in a dead heat with GPT-5.5 (98). Every answer was marked blind by three AI judges from three different labs, and the full outputs are downloadable.
How is this ranking made?
Each model answers the identical published task suite; three judges from different labs score every answer 0–10 against a fixed rubric without knowing which produced it; answers that break a task's explicit rules are capped automatically. The score is the average, out of 100. No vendor pays for placement.
What happens when two models score the same?
They are separated by a fixed, published tie-break rather than by editorial choice: first the fewest machine-checked rule breaches, then the lowest measured cost per run. The order is therefore reproducible from the same data, and where the tie-break decides the top place the page says so.
How often does this page update?
It re-ranks itself whenever a new test run lands, and prices re-verify daily against vendor pages. The current ranking was last computed on 13 Aug 2026.
Head-to-heads in summarising
Show all 20 tested pairs ▾
More rankings ▾
Best AI for writing · Best AI chatbot for everyday use · Best AI for coding · Best free AI model · Best AI for spreadsheets and Excel · Best AI essay writer · Best AI for extracting data from text · Best AI for writing emails · Best AI for everyday maths and percentages · Best AI for customer service replies · Best AI for revision and study notes · Best AI for vibe coding · Best AI for making flashcards · Best AI for social media posts · Best AI for job applications and cover letters · Best AI for presentations · Best AI for writing your CV · Best AI for research skills · Best AI for creative writing · Best AI for translation · Best AI for travel planning · Best AI for emotional support · Best AI for everyday legal questions · Best AI for health questions · Best AI for writing a book · Best AI for humanising AI text · Best AI research agent · Best AI for code review · Best AI for workflow automation · Best AI for bookkeeping · Best AI for HR and employment questions · Best AI for meeting notes · Best AI for property and lettings · Best-value AI model API · every model we track · every tool