01
Check task requirements
Models must support requirements implied by the selected task. Tool-use tasks require tool support, and long-document tasks require at least a 32K context window.
How model recommendations work.
We start with what you actually want to do, then check whether each model has enough trustworthy evidence for that job. A model cannot take first place merely because it is cautious, popular, inexpensive, or made by a familiar company.
02
Score relevant evidence
We translate technical tests into practical questions: can it answer correctly, follow your request, use tools, finish coding work, or reason through a difficult problem?
03
Account for uncertainty
Every task has essential evidence. If that evidence is missing—or a model falls below an important minimum—it cannot be recommended for that task.
Evidence used for each task.
| Use case | Benchmark evidence | Required capability |
|---|---|---|
| Everyday tasks | Overall quality 35% · GDPval-AA 20% · AA-Omniscience Accuracy 25% · IFBench 15% · AA-Omniscience Non-Hallucination Rate 5% | No additional gate |
| Taking actions for you | Agentic index 35% · Terminal-Bench Hard 30% · τ²-Bench Telecom 15% · IFBench 10% · Overall quality 10% | Tool support |
| Writing code | Coding index 40% · Terminal-Bench Hard 25% · SciCode 15% · IFBench 10% · Overall quality 10% | No additional gate |
| Math and logic | GPQA 45% · CritPt 25% · Humanity's Last Exam 15% · Overall quality 15% | No additional gate |
| Following detailed instructions | IFBench 70% · Overall quality 30% | No additional gate |
| Reading long documents | AA-LCR 50% · Context capacity 30% · Overall quality 20% | 32K+ context |
| Creating documents | IFBench 30% · GDPval-AA 30% · AA-Omniscience Accuracy 15% · Overall quality 15% · AA-LCR 10% | No additional gate |
| Answering factual questions | AA-Omniscience Accuracy 40% · GPQA 20% · Humanity's Last Exam 25% · Overall quality 15% | No additional gate |
| More reliable answers | AA-Omniscience Non-Hallucination Rate 45% · AA-Omniscience Accuracy 20% · GPQA 15% · Humanity's Last Exam 10% · Overall quality 10% | No additional gate |
Single use case.
Task scores use only evidence that actually exists. We do not fill a missing test with an average, zero, or an optimistic guess.
A model must contain every essential measurement for the task and at least 70% of its expected evidence.
Confidence = catalogue confidence × (0.35 + 0.65 × evidence coverage).
Adjusted score = task score − 15 × (1 − confidence).
Multiple use cases.
Selected use cases receive equal importance.
Their adjusted scores are combined using a geometric mean. This prevents exceptional performance in one task from hiding weak performance in another.
Final score = 70% combined score + 30% weakest selected score.
Close scores are resolved by overall intelligence, then response speed, then estimated cost per task. Speed and price cannot rescue a model that is not good enough for the selected work.
For a single task, models within one point are treated as effectively tied. We then use a practical secondary measurement for that task and explain the close alternative instead of pretending the difference is decisive.
Missing and changing data.
Missing results are never inferred from downloads, popularity, descriptions, model size, or company reputation. Models without essential evidence are omitted from that recommendation while remaining visible in the full benchmark catalogue.
Pricing, context limits, capabilities, and observed endpoint performance come from the promoted catalogue snapshot. Benchmark values retain their original source and version. Catalogue refreshes are reviewed before publication.
Rankings are decision support, not proof that one model is universally best. Workload details, prompting, provider configuration, regional availability, and application design can change real-world results.