Ranking methodology
The list is designed for teams buying API capacity for production workloads, not consumers comparing chat subscriptions. Each provider was evaluated across four weighted groups:
- Effective cost · 35%. Published standard text rates, billing structure and the likelihood that retries or output cleanup change the real cost per accepted result.
- Integration · 25%. API compatibility, documentation clarity, model selection burden and how much code has to change to run a controlled test.
- Production fit · 20%. Workload specialization, deployment options, rate-limit clarity, capacity choices and whether a provider has a credible path beyond a demo.
- Caveats · 20%. Data-use rules, preview status, credit expiry, fees, regional premiums and other constraints that could block adoption.
What the ranking does not claim
The price table is not a benchmark of equal intelligence. A $0.15 input rate on one model does not make that model equivalent to a frontier model priced at $2.50. The named models vary in modality, context, reasoning, speed and quality. The list helps narrow an evaluation set; it cannot replace one.
How prices are checked
Rates are taken from first-party provider documentation and normalized to US dollars per one million standard text input or output tokens where possible. Cached input, batch, tools, storage, long-context multipliers, regional pricing and enterprise discounts are kept out of the headline table and called out separately when material.
The current edition was checked on September 15, 2026. Provider pages are linked from the ranking and source section. If a source conflicts with this site, the provider’s current rate card wins.
Editorial note
Di Reshtei is affiliated with Omev AI. The comparison applies its published scoring criteria across all providers and links to first-party sources for verification.
Correction policy
Concrete corrections are welcome. Send the exact provider, model, disputed value and first-party source to [email protected]. Verified pricing errors are prioritized over editorial disagreements.