Best overall
Highest score of the 10 providers in this ranking (83/100). Embeddings and Command models on a direct contract.
Visit CohereModel APIs for chat, tools and embeddings. Token prices move. We rank public pricing, a free credit and self-serve access. We do not freeze a token rate into a monthly champion.
Updated Nov 2026 · 10 providers · 0 with a published monthly plan · 8 with a free tier
Three starting points from the ranking below, each for a different kind of buyer.
Best overall
Highest score of the 10 providers in this ranking (83/100). Embeddings and Command models on a direct contract.
Visit CohereBest free tier
The highest-ranked provider with a free plan or trial, after the picks above. Inexpensive tokens on popular open models.
Visit DeepInfraSorted by score out of 100. A from-price is the cheapest plan the vendor publishes, not a quote for your workload.
Embeddings and Command models on a direct contract
Inexpensive tokens on popular open models
Production traffic on a specific open model
Trying Gemini on a free tier before a paid key
Low-latency open-model inference
First-party Mistral models and a European vendor
Switching models without a new account each time
Choosing among many open models on one key
Claude models when you want a direct Anthropic contract
Frontier GPT models with the widest tooling ecosystem
You pay for tokens in and tokens out, priced per model. These are the lines that decide the invoice.
Everything you send counts: the system prompt, conversation history, retrieved documents and tool definitions.
Generated text is usually priced higher than input. Long answers and reasoning steps add up quickly.
Several providers discount repeated prompt prefixes. A stable system prompt can cost much less on the second call.
Batch endpoints trade latency for a lower price. Priority or provisioned throughput costs more for guaranteed capacity.
| Provider | Starting price | Billing model | Free tier |
|---|---|---|---|
| Cohere | Trial credits, then per token | Pay for what you use: requests, tokens, traffic or compute time. The bill moves with volume. | Yes |
| DeepInfra | Pay per token | Pay for what you use: requests, tokens, traffic or compute time. The bill moves with volume. | Yes |
| Fireworks | Pay per token | Pay for what you use: requests, tokens, traffic or compute time. The bill moves with volume. | Yes |
| Google Gemini | Free tier, then per token | Pay for what you use: requests, tokens, traffic or compute time. The bill moves with volume. | Yes |
| Groq | Free tier, then per token | Pay for what you use: requests, tokens, traffic or compute time. The bill moves with volume. | Yes |
| Mistral | Free tier, then per token | Pay for what you use: requests, tokens, traffic or compute time. The bill moves with volume. | Yes |
| OpenRouter | Pay per token, per model | Pay for what you use: requests, tokens, traffic or compute time. The bill moves with volume. | Yes |
| Together AI | Pay per token | Pay for what you use: requests, tokens, traffic or compute time. The bill moves with volume. | Yes |
| Anthropic | Pay per token | Pay for what you use: requests, tokens, traffic or compute time. The bill moves with volume. | No |
| OpenAI | Pay per token | Pay for what you use: requests, tokens, traffic or compute time. The bill moves with volume. | No |
You need the strongest general models and can pay the token rate. OpenAI, Anthropic and Google Gemini are the usual shortlist—read the data policy before production traffic.
You want lower token prices on open weights and can validate quality on your eval set. Groq, Fireworks, Together and DeepInfra fit this lane.
You want one integration and the option to switch models without rewriting clients. OpenRouter is built for that; still track per-model rates and policies.
Utilization is high and you accept ops work. Rent capacity on [GPU cloud](/gpu-cloud) instead of paying per token all day.
OpenAI, Anthropic, Google Gemini, Mistral and Cohere sell first-party models with public rate cards. Groq, Fireworks, Together and DeepInfra sell hosted open models with a per-token price you can read. OpenRouter is a router across several of them. Those are different bets: model quality, latency, data policy and the rate on the day you ship.
A ranking that crowns one API on price would be wrong next week. This list rewards a rate card you can read, a free credit when the vendor offers one, and an API key you can create yourself. For GPUs you operate, see GPU cloud. For storing embeddings or app state, pair the API with a managed database instead of bolting persistence onto the model vendor.
Read input and output prices separately. Output tokens often cost more, and tool-calling or long context can change the bill even when the headline input rate looks fine. Price a real prompt shape—system message, retrieval chunk, expected completion—not the smallest number on the marketing page.
Data policy is part of the shortlist. Retention, training use and region of processing should be on a current policy page before you send customer content. If the policy is vague, do not put private data on that key. Enterprise contracts can change defaults; self-serve keys follow the published policy.
Rate limits and model deprecations are normal. Design a fallback: a second model on the same vendor, a router like OpenRouter, or a second vendor. That is a product decision, not an afterthought when a model name disappears from the docs.
Confirm the model names in the docs match what you will call in production. Preview models and aliases change. Re-check the live rate card the day you ship, and again when you renegotiate volume—screenshots in a ticket go stale fast.
Step 1
Frontier chat, a small cheap model, or embeddings. The class decides the shortlist before the logo does.
Step 2
Output tokens often cost more. Price a real prompt, not the smallest input rate on the hero.
Step 3
Retention, training and region should be on a policy page. If they are not, do not send private data.
Step 4
Rate limits and model deprecations happen. A router or a second vendor is a product decision, not an afterthought.
Measure real traffic: average input, average output and calls per user. A guessed monthly bill is usually wrong.
Conversation length grows the input on every turn. Trim or summarise it.
Route simple tasks to a smaller model and keep the expensive one for the hard cases.
New accounts often start on low limits. Check the tier you will be on at launch.
Each provider is scored out of 100 against the other 9providers on this page, using only what is published on the vendor's site. The sponsor slot is not an input.
Full editorial policySmaller hosted open models usually have the lowest token rate. Frontier models cost more and do a different job. Compare rates on the same day, for the same prompt shape.
Token prices and model names change often. We store a public price label and refuse to invent a monthly bill from a guessed prompt.
Use an API when you want a key and a rate card. Use GPU cloud when you will run the model yourself and can keep the GPU busy.
No. Quality depends on your prompts and eval set. We rank public commercial terms: price visibility, free credit and self-serve access.
Count input and output tokens on a sample of production prompts. Multiply by the live rates. Include retries and tool calls if you use them.
It is a router across providers. You still pick models and inherit their rates and policies. Useful for fallbacks; not a substitute for reading each upstream policy.