How to Rank in the Top 100 AI Models
A practical guide for model builders, labs and open-source maintainers who want their model listed — and want to climb. Spoiler: none of it involves paying us.
Eligibility
Any AI model that is publicly available — via API, open weights, or a shipping product — is eligible. Private research previews, vapourware and models that cannot be independently tested are not considered until they can be. We list the model, not the company, so each distinct model line gets its own entry.
The five things we score
| Dimension | Weight | What moves it |
|---|---|---|
| Capability | 35% | Independent benchmarks, our own task suite, head-to-head community evaluations |
| Reliability | 20% | Consistency across repeated runs, hallucination rate, uptime and versioning discipline |
| Efficiency | 15% | Price per million tokens or per output, latency, hardware required to self-host |
| Accessibility | 15% | Open weights or API availability, licence terms, regions served, documentation quality |
| Ecosystem | 15% | Framework integrations, fine-tunes, community activity, tooling and SDKs |
Full detail is on the Methodology page.
Step-by-step: getting listed
1. Make the model testable
Give us a way to run it: a public API key tier, a Hugging Face repo, or a hosted demo. Models we cannot test cannot be ranked.
2. Publish a model card
Architecture, training data summary, context window, licence, pricing and known limitations. Clear documentation lifts the Accessibility score immediately.
3. Submit through the form below
Include links to independent evaluations if you have them. We verify everything ourselves; vendor benchmarks are treated as claims, not evidence.
4. Wait for the next edition
Rankings are refreshed on the first business day of each month. New entrants appear in the first edition after review is complete, typically 2–4 weeks.
Step-by-step: moving up
Ship measurable improvements
New checkpoints with published eval deltas are the single fastest way to move. Tell us what changed and how you measured it.
Reduce friction
Lower pricing, broader regional availability, a permissive licence or a free tier each raise Efficiency and Accessibility. These are often cheaper than a training run.
Invest in the ecosystem
First-class support in popular inference engines, agent frameworks and IDE plugins increases real-world usefulness — which is what we rank.
Be honest about limitations
Models whose documentation accurately predicts failure modes score higher on Reliability than models that over-promise.
What will not help
Sponsoring a tile, offering payment, exclusive previews for our editors, or asking nicely. Our editors are firewalled from sponsorship data and sponsorship is disclosed on every page. See our Editorial Policy.