LLM API pricing comparison

List prices per million tokens, checked 22 Nov 2026. The sticker price is only half the bill — caching, output length, and batch discounts decide the rest.

9 modelsUSD list pricesUpdated 22 Nov 2026

Last checked 22 Nov 2026 · USD per 1M tokens

If you are building on top of a model, the subscription price is irrelevant — you pay per token. This table puts the major APIs side by side, including the cached input rate where the provider publishes one.

The table

ModelInput / 1MCached inputOutput / 1MContextSource
GPT-6 AstraOpenAI$10$1$50272K+Source
Claude Fable 5.2Anthropic$8—$401MSource
Claude Opus 5.5Anthropic$5—$251MSource
Claude Sonnet 5.5Anthropic$2—$101MSource
Gemini 3.5 ProGoogle$2—$121MSource
Gemini 3.5 FlashGoogle$1.5—$91MSource
Mistral Large 4Mistral$0.7$0.07$2.11MSource
DeepSeek V4.1 FlashDeepSeek$0.3$0.006$1.2128KSource
DeepSeek V4 ProDeepSeek$1.32$0.044$3.96128KSource

List prices, not negotiated enterprise rates. Some providers charge different rates by region, context length, or usage tier. The “source” link goes to the provider’s own pricing page.

Calculator

Estimate your monthly API bill

Enter millions of tokens per month. This uses list prices only; caching, batch, and off-peak discounts are not applied.

$140 / month

How to read the numbers

Input is what you send: prompts, documents, retrieved context, and conversation history. Output is what the model generates. Output is usually 3–5× the input price, so a workload that generates long reports costs more than one that summarizes long documents.

Cached input matters when you reuse a long system prompt or a document across many calls. If your app sends the same 50K-token context on every request, the cached rate is the difference between a viable product and a bad unit economics.

Three traps

  • Context length pricing. Some models double the input price above a threshold (for example, 272K tokens). A long-context model is not automatically cheap for long documents.
  • Reasoning tokens. Models that “think” bill the reasoning tokens as output. A hard prompt can cost several times what the visible answer suggests.
  • Retries and agents. An agent loop can make 10 calls for one user request. Multiply the token estimate by the number of turns, not the number of users.
The short version: for high-volume extraction and classification, the cheapest model that passes your eval is usually the right one. For complex reasoning, the expensive model can still be cheaper per completed task because it needs fewer retries.

Questions

Output tokens are generated one at a time, so they consume more compute. A model with a $10 input price often has a $50 output price. If your workload is output-heavy, the cheap-input model is not the cheap model.

If you send the same long prompt or document repeatedly, the provider can cache it and charge less for the cached portion. The discount varies by provider; the table shows published cached rates where available.

A rough English rule is 1 token ≈ 0.75 words, so 1,000 words ≈ 1,300 tokens. For production, measure with the provider's tokenizer rather than guessing.

They are published list prices checked on 22 November 2026. Vendors change prices and model names often. Every row links to the source page; confirm before you commit.

Several providers offer a discount for non-urgent batch jobs, often around 50%. The table does not apply batch pricing; check the provider's pricing page if your workload can wait.

Building with an API?

The advisor is for seat-priced tools. For API cost, start with the table and the calculator.

Ask the advisor