Arena Hard Prompts

What does it measure?
The ranking recomputed on hard questions only, with easy ones removed. It shows which AI was preferred on questions needing expertise, complex reasoning and accuracy.
Who evaluates it, and how?
On Arena Intelligence, a user enters a question, two anonymous AIs answer, and the user votes for the better one. Over 82 million votes are aggregated with a statistical model (Bradley-Terry) into an Elo-style score, and this table takes the 'style-controlled' version from the official dataset, which statistically removes the bias toward long, nicely formatted answers. Prompts are rated on seven criteria — specificity, domain knowledge, reasoning steps, technical accuracy and more — and only the top third or so are counted.
Reading the score
This is a head-to-head record, not an exam score (accuracy %). It is computed from the probability of being preferred over other AIs, so there is no maximum, and when a new AI enters, existing scores shift slightly. Typical values are 1,000–1,500; a 100-point gap means the leader is preferred about 64% of the time, a 200-point gap about 76%. Correctness is not counted — a wrong but friendly answer can win — so if accuracy matters, also look at accuracy tests (GPQA, AIME, etc.). Gaps of a few dozen points are effectively the same tier even if the rank differs. An AI whose score differs a lot from the overall ranking performs differently on easy versus hard questions.

Rank

Top score1547
Models tested108
Scoring basisHigher is better
Last updated2026.11.19
Model release date
  1. 1
    Google2026.10.28
    1547
  2. 2
    Google2026.09.30
    1544
  3. 3
    Google2026.11.18
    1542
  4. 4
    Meta2026.10.12
    1531
  5. 5
    Anthropic2026.09.22
    1530
  6. 6
    Anthropic2026.06.09
    1530
  7. 7
    Anthropic2026.02.04
    1526
  8. 8
    Anthropic2026.09.01
    1522
  9. 9
    Anthropic2026.04.16
    1519
  10. 10
    Meta2026.09.02
    1518

45 of the 64 AIs released in the last 60 days have no score here yet · Microsoft-Decision-1.1, d1.2, Qwen4 Max and 42 more

It’s a preference vote, not a correctness check. A wrong but friendly answer can win.

Which questions got asked depends on who voted — they may differ from your use.

Scores labelled 'Self-reported' were published by the AI maker itself, not measured by a third party.