Best AI for summarising documents

Strongest showing: “Meeting notes to actions” — scored 10/10 by the panel. Weakest: “Summary for a specific reader” at 9/10. The judges flagged 1 instruction breach across the run — those scores were capped automatically.

“The response perfectly extracts all action items, assigns the correct 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 ↓

Strongest showing: “Five-bullet summary” — scored 10/10 by the panel. Weakest: “One-sentence gist” at 9/10.

“Accurate, concise, meets 5-bullet/15-word constraints, though last bullet merges two facts slightly awkwardly.”— anthropic/claude-sonnet-5, judging blind · full receipts ↓

Strongest showing: “Meeting notes to actions” — scored 10/10 by the panel. Weakest: “Summary for a specific reader” at 9/10. The judges flagged 1 instruction breach across the run — those scores were capped automatically.

“Accurate, well-formatted action items with owners correctly assigned or marked unassigned; Sam's leave correctly omitted as non-action. Minor: could omit conditional caveat but still concise and useful.”— anthropic/claude-sonnet-5, judging blind · full receipts ↓

Strongest showing: “Meeting notes to actions” — scored 10/10 by the panel. Weakest: “Long to short, no loss” at 8/10. The judges flagged 1 instruction breach across the run — those scores were capped automatically.

“The response perfectly extracts all action items, assigns the correct owners (including 'unassigned'), and formats them as a bulleted list containing only the requested information.”— google/gemini-3.1-pro-preview, judging blind · full receipts ↓

Strongest showing: “Do not invent” — scored 10/10 by the panel. Weakest: “Five-bullet summary” at 8/10.

“Accurate, concise summary under 50 words, correctly follows format with 'Not stated' line, no invented facts. Very minor: could be slightly more concise but meets requirements well.”— anthropic/claude-sonnet-5, judging blind · full receipts ↓

Strongest showing: “Five-bullet summary” — scored 10/10 by the panel. Weakest: “One-sentence gist” at 8/10.

“Accurate, concise, exactly 5 bullets, all under 15 words, covers key facts clearly and readable quickly.”— anthropic/claude-sonnet-5, judging blind · full receipts ↓

Strongest showing: “Meeting notes to actions” — scored 10/10 by the panel. Weakest: “Headline and standfirst” at 8/10.

“Accurate, follows format, captures all action items with owners; minor nuance on Priya's conditional commitment omitted but acceptable.”— anthropic/claude-sonnet-5, judging blind · full receipts ↓

Strongest showing: “Do not invent” — scored 10/10 by the panel. Weakest: “Meeting notes to actions” at 6/10.

“Accurate, concise summary under 50 words, follows format exactly, plausible 'not stated' fact. Minor: could be slightly more informative but meets task well.”— anthropic/claude-sonnet-5, judging blind · full receipts ↓

Strongest showing: “Meeting notes to actions” — scored 10/10 by the panel. Weakest: “Summarise a disagreement” at 6/10. The judges flagged 1 instruction breach across the run — those scores were capped automatically.

“The response perfectly extracts the action items and their owners, correctly identifies the unassigned task, and formats them as a bulleted list exactly as requested.”— google/gemini-3.1-pro-preview, judging blind · full receipts ↓

Strongest showing: “Five-bullet summary” — scored 10/10 by the panel. Weakest: “Extract the decision” at 5/10. The judges flagged 4 instruction breaches across the run — those scores were capped automatically.

“The response perfectly follows all instructions, including the exact bullet count and word limit per bullet. It accurately and concisely captures all key information, making it highly useful for a busy professional.”— google/gemini-3.1-pro-preview, judging blind · full receipts ↓

Strongest showing: “Meeting notes to actions” — scored 10/10 by the panel. Weakest: “Five-bullet summary” at 6/10. The judges flagged 1 instruction breach across the run — those scores were capped automatically.

“Accurate, follows format, complete action items with owners correctly assigned; minor debatable inclusion but overall solid and concise.”— anthropic/claude-sonnet-5, judging blind · full receipts ↓

Strongest showing: “Do not invent” — scored 10/10 by the panel. Weakest: “Meeting notes to actions” at 5/10.

“Accurate, concise summary under 50 words, correct format, reasonable 'not stated' item. Minor: could be more concise but meets constraints well.”— anthropic/claude-sonnet-5, judging blind · full receipts ↓

Strongest showing: “Do not invent” — scored 10/10 by the panel. Weakest: “Headline and standfirst” at 8/10.

“Accurate, concise, under 50 words, follows format with 'Not stated' line and valid unmentioned fact.”— anthropic/claude-sonnet-5, judging blind · full receipts ↓

Strongest showing: “Five-bullet summary” — scored 10/10 by the panel. Weakest: “Headline and standfirst” at 8/10. The judges flagged 1 instruction breach across the run — those scores were capped automatically.

“Accurate, exactly 5 bullets, all under 15 words, clear and concise, covers all key facts.”— anthropic/claude-sonnet-5, judging blind · full receipts ↓

Strongest showing: “Do not invent” — scored 10/10 by the panel. Weakest: “Meeting notes to actions” at 6/10. The judges flagged 2 instruction breaches across the run — those scores were capped automatically.

“Accurate, concise summary under 50 words, includes required 'Not stated' line with reasonable unstated fact. Minor stylistic nitpick only.”— anthropic/claude-sonnet-5, judging blind · full receipts ↓

Strongest showing: “Meeting notes to actions” — scored 10/10 by the panel. Weakest: “Five-bullet summary” at 3/10. The judges flagged 1 instruction breach across the run — those scores were capped automatically.

“Accurate, follows format, concise; minor nuance missed (Priya's conditional 'if she has the numbers') but overall correct and useful.”— anthropic/claude-sonnet-5, judging blind · full receipts ↓

Strongest showing: “Summarise a disagreement” — scored 10/10 by the panel. Weakest: “Extract the decision” at 6/10.

“Accurate, balanced, but exceeds 80 words (approx 70... actually count: ~70 words, within limit). Clear and fair, minor verbosity.”— anthropic/claude-sonnet-5, judging blind · full receipts ↓

Strongest showing: “Five-bullet summary” — scored 10/10 by the panel. Weakest: “Headline and standfirst” at 6/10.

“Accurate, concise, follows 5-bullet format, though last bullet merges two facts, slightly overloaded.”— anthropic/claude-sonnet-5, judging blind · full receipts ↓

Strongest showing: “Meeting notes to actions” — scored 10/10 by the panel. Weakest: “Extract the decision” at 5/10. The judges flagged 2 instruction breaches across the run — those scores were capped automatically.

“Accurate, well-formatted list matching notes; correctly assigns owners/unassigned. Minor debatable point: Priya's cashflow task was conditional, but reasonably represented.”— anthropic/claude-sonnet-5, judging blind · full receipts ↓