Best LLM APIs (2026)

Model 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

Best LLM APIs right now

Three starting points from the ranking below, each for a different kind of buyer.

Best overall

Cohere

Trial credits, then per token

83

Highest score of the 10 providers in this ranking (83/100). Embeddings and Command models on a direct contract.

Visit Cohere

Best free tier

DeepInfra

Pay per token

83

The highest-ranked provider with a free plan or trial, after the picks above. Inexpensive tokens on popular open models.

Visit DeepInfra

Full LLM APIs ranking

Sorted by score out of 100. A from-price is the cheapest plan the vendor publishes, not a quote for your workload.

  1. 01
    Cohere

    Trial credits, then per token

    83

    Embeddings and Command models on a direct contract

    Free tierVisit
  2. 02
    DeepInfra

    Pay per token

    83

    Inexpensive tokens on popular open models

    Free tierVisit
  3. 03
    Fireworks

    Pay per token

    83

    Production traffic on a specific open model

    Free tierVisit
  4. 04
    Google Gemini

    Free tier, then per token

    83

    Trying Gemini on a free tier before a paid key

    Free tierVisit
  5. 05
    Groq

    Free tier, then per token

    83

    Low-latency open-model inference

    Free tierVisit
  6. 06
    Mistral

    Free tier, then per token

    83

    First-party Mistral models and a European vendor

    Free tierVisit
  7. 07
    OpenRouter

    Pay per token, per model

    83

    Switching models without a new account each time

    Free tierVisit
  8. 08
    Together AI

    Pay per token

    83

    Choosing among many open models on one key

    Free tierVisit
  9. 09
    Anthropic

    Pay per token

    73

    Claude models when you want a direct Anthropic contract

    No free tierVisit
  10. 10
    OpenAI

    Pay per token

    73

    Frontier GPT models with the widest tooling ecosystem

    No free tierVisit

How LLM APIs pricing works

You pay for tokens in and tokens out, priced per model. These are the lines that decide the invoice.

Input tokens

Everything you send counts: the system prompt, conversation history, retrieved documents and tool definitions.

Output tokens

Generated text is usually priced higher than input. Long answers and reasoning steps add up quickly.

Cached input

Several providers discount repeated prompt prefixes. A stable system prompt can cost much less on the second call.

Batch and priority tiers

Batch endpoints trade latency for a lower price. Priority or provisioned throughput costs more for guaranteed capacity.

ProviderStarting priceBilling modelFree tier
CohereTrial credits, then per tokenPay for what you use: requests, tokens, traffic or compute time. The bill moves with volume.Yes
DeepInfraPay per tokenPay for what you use: requests, tokens, traffic or compute time. The bill moves with volume.Yes
FireworksPay per tokenPay for what you use: requests, tokens, traffic or compute time. The bill moves with volume.Yes
Google GeminiFree tier, then per tokenPay for what you use: requests, tokens, traffic or compute time. The bill moves with volume.Yes
GroqFree tier, then per tokenPay for what you use: requests, tokens, traffic or compute time. The bill moves with volume.Yes
MistralFree tier, then per tokenPay for what you use: requests, tokens, traffic or compute time. The bill moves with volume.Yes
OpenRouterPay per token, per modelPay for what you use: requests, tokens, traffic or compute time. The bill moves with volume.Yes
Together AIPay per tokenPay for what you use: requests, tokens, traffic or compute time. The bill moves with volume.Yes
AnthropicPay per tokenPay for what you use: requests, tokens, traffic or compute time. The bill moves with volume.No
OpenAIPay per tokenPay for what you use: requests, tokens, traffic or compute time. The bill moves with volume.No

What fits

Frontier first-party API

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.

Hosted open-model API

You want lower token prices on open weights and can validate quality on your eval set. Groq, Fireworks, Together and DeepInfra fit this lane.

Multi-vendor router

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.

Self-hosted on GPUs

Utilization is high and you accept ops work. Rent capacity on [GPU cloud](/gpu-cloud) instead of paying per token all day.

How to choose an LLM API

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.

  1. Step 1

    Pick the model class

    Frontier chat, a small cheap model, or embeddings. The class decides the shortlist before the logo does.

  2. Step 2

    Read input and output rates

    Output tokens often cost more. Price a real prompt, not the smallest input rate on the hero.

  3. Step 3

    Check the data policy

    Retention, training and region should be on a policy page. If they are not, do not send private data.

  4. Step 4

    Put a fallback in the design

    Rate limits and model deprecations happen. A router or a second vendor is a product decision, not an afterthought.

Checklist before you buy

  • A public per-token or per-request rate
  • Input and output prices shown separately
  • A free credit or tier with the limit in writing
  • Self-serve API keys
  • A data-retention policy you can link to
  • Rate limits and model names that match the docs

Common mistakes

Estimating from a single prompt

Measure real traffic: average input, average output and calls per user. A guessed monthly bill is usually wrong.

Sending the whole history every time

Conversation length grows the input on every turn. Trim or summarise it.

Using the largest model by default

Route simple tasks to a smaller model and keep the expensive one for the hard cases.

Ignoring rate limits

New accounts often start on low limits. Check the tier you will be on at launch.

Compare LLM APIs head to head

How we rank LLM APIs

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 policy
35
Monthly floorWhere the cheapest published paid plan sits among peers in the same category. Cheaper scores higher. A vendor without a monthly floor gets the middle, so a missing price never counts as $0.
25
Public priceA rate card, plan grid or free tier you can read without talking to sales.
15
Self-serve signupYou can create an account and launch the resource yourself.
15
Documented scopeUp to three concrete, documented capabilities on the profile. Logos and ad spend are not counted.
10
Free tier or trialA bonus when the vendor publishes one. A free plan is a test, not a production price.

LLM APIs glossary

Token
A chunk of text, roughly a short word or part of a word. Prices are quoted per million tokens.
Context window
The maximum number of tokens a model can read in one request.
Prompt caching
A discount for input that repeats across calls, such as a fixed system prompt.
Batch API
Asynchronous processing of many requests at a lower price and higher latency.
Open-weight model
A model whose weights are published, so several inference providers can host it.

LLM APIs FAQ

Which LLM API is cheapest?

Smaller 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.

Why is there no monthly winner?

Token prices and model names change often. We store a public price label and refuse to invent a monthly bill from a guessed prompt.

Hosted API or our own GPU?

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.

Do you benchmark model quality?

No. Quality depends on your prompts and eval set. We rank public commercial terms: price visibility, free credit and self-serve access.

How should I price a real workload?

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.

Is OpenRouter a model provider?

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.