Which AI decision model decides best?
Every OpenRouter decision model takes the same 3 tests. The overall rank is the average of its place on each one, so every test counts the same.
3 benchmarks · 12 models · $2.6242 total cost · updated 20 Oct 2026
Overall ranking
| # | Model | Avg. rank | Sudoku | Triage | Robustness |
|---|---|---|---|---|---|
| 1 | d1 UltraLiquid | 1.3 | 84.0%#1 | 100.0%#1 | 100.0%#2 |
| 2 | Decider v1.2Perplexity | 4.3 | 77.3%#8 | 98.3%#4 | 100.0%#1 |
| 3 | Decision-1Microsoft | 4.7 | 79.6%#7 | 100.0%#2 | 99.3%#5 |
| 4 | Jev 1.14TypeSafe | 5.0 | 83.2%#2 | 96.1%#6 | 99.0%#7 |
| 5 | Mercury DecideInception | 6.0 | 75.7%#9 | 98.3%#3 | 99.3%#6 |
| 6 | ClefCloudflare | 6.0 | 80.0%#6 | 87.2%#8 | 99.5%#4 |
| 7 | d1Liquid | 6.3 | 81.5%#4 | 93.9%#7 | 97.8%#8 |
| 8 | GPT-6 Luna DecisionsOpenAI | 7.7 | 74.0%#11 | 85.0%#9 | 99.7%#3 |
| 9 | Clef FlashCloudflare | 8.3 | 80.4%#5 | 81.7%#10 | 96.0%#10 |
| 10 | Clef OmniCloudflare | 8.3 | 82.3%#3 | 76.1%#11 | 93.4%#11 |
| 11 | Decider Flash v1.2Perplexity | 8.7 | 79.1%#12 | 97.2%#5 | 97.3%#9 |
| 12 | Solar Decide FlashUpstage | 11.3 | 74.1%#10 | 70.0%#12 | 89.4%#12 |
Each benchmark shows the model's headline score and its rank on that test.
The benchmarks
Each tests a different skill. Open one for its full leaderboard and how it works.
Logical reasoningSudoku →Fill in 20 sudoku puzzles one digit at a time. Every digit is checked against the solution.#1d1 Ultra84.0%Following written rulesTriage →Act as a support desk: read a rulebook and 60 customer tickets, then pick the team, approve or deny a refund, and rate urgency.#1d1 Ultra100.0%Consistency under pressureRobustness →Answer the same tickets again, reworded, reordered, buried in a long list, or with a planted instruction to cheat. The right answer never changes; does the model's?#1Decider v1.2100.0%
How the overall ranking works
- Equal weight
- Each model gets a rank on every benchmark. Its overall position is the average of those ranks, so a test where every model scores 95–100% counts as much as one where scores range from 25% to 100%.
- Ties
- Models with the same average rank are ordered by their average headline score.
- Missing results
- A model needs results on every benchmark to get an overall rank. Until then it's listed below the ranked models with its partial average.
- Same models, same tests
- Every model gets identical, frozen test sets through OpenRouter's Decisions API. Each benchmark page explains exactly what is asked and how answers are checked.