LLM API pricing compared, with a cost calculator

LLM API pricing is quoted per million input and output tokens, but those two numbers rarely predict the bill. Caching, batch discounts, long prompts, reasoning tokens and even how each model counts tokens can move the real cost by a factor of two or more. This page compares 16 current models from OpenAI, Anthropic, Google and DeepSeek, with prices taken from each provider's own pricing page on October 26, 2026, and a calculator that applies the rules they publish.

Price per million tokens

DeepSeek's prices are its peak-hour rates; off-peak rates are half. Gemini 3.1 Pro is listed by Google as a preview model. Claude Haiku 5.5 costs more on prompts over 100,000 tokens ($0.50 input and $2.50 output), and Gemini 4 Argon's prices double on December 1, 2026.

Same list price, different bill

GPT-6.1 Sol and Claude Sonnet 5.5 both list $2 per million input tokens and $10 per million output tokens. Take the calculator's example workload: 20,000 requests a month, each with 3,000 input tokens (40% of them repeated instructions read from cache) and 600 output tokens.

At list prices the two cost the same, because both read cached input at 5% of the input price ($0.10). What separates them is how many tokens the same text becomes. Anthropic's own pricing page says its tokenizer for Claude 4.7 and later models "produces approximately 30% more tokens for the same text" than its previous one. If your token counts were measured with an older Claude model, the bill grows by that much. A price per token is only comparable if the token counts are.

The 30% figure compares Anthropic's two tokenizers, not Anthropic with OpenAI or Google, and the exact change "depends on the content". For a firm number, count tokens for a sample of your real prompts with each provider's tools before you commit. For the basics of what a token is, see what is a token.

What else moves the bill

Output costs several times more than input. Across these models, output tokens cost between 3 and about 8 times as much as input tokens, and 5 times for most of them. Reasoning models also bill their "thinking" as output: Google's table says so explicitly ("including thinking tokens"). Long answers and heavy reasoning dominate the bill more than long prompts.

Caching. Repeated content (system prompts, documents, tool definitions) can be cached:

• OpenAI: cached input at 5% of the input price on GPT-6.1 Sol; writing to the cache costs $2.50 per million tokens for 30 minutes or $5 for 12 hours on that model.
• Anthropic: reading from cache costs 0.1x the input price (0.05x on Opus 5.5 and Sonnet 5.5, 0.025x on Fable 5.1); writing costs 1.25x for a 5-minute cache or 2x for a 1-hour cache. Anthropic notes that caching "pays off after one cache read" for the 5-minute duration.
• Google: reduced input price for cached content plus a storage charge per million tokens per hour. Gemini 4 Argon only caches implicitly.
• DeepSeek: a separate, much lower price for cache hits.

The calculator includes cache reads but not cache writes or storage, so it slightly flatters heavy caching.

Batch processing. OpenAI, Anthropic and Google offer 50% off for requests you can wait for; DeepSeek's pricing page lists no batch discount. Anthropic adds that batch and caching discounts can be combined. Google's batch table keeps cached input at the standard price on Gemini 3.1 Pro and lists it at $0.02 rather than half of $0.03 on Gemini 3.5 Flash-Lite; the calculator follows both.

Long prompts. Thresholds differ:

If you routinely send very long prompts, Anthropic's flat pricing changes the comparison; the calculator stops at 200,000 input tokens for that reason, and applies Haiku 5.5's higher rate above 100,000. How much text fits in a prompt in the first place is explained in what a context window is.

Where the data is processed. OpenAI adds 10% for regional processing (data residency) on models released on or after March 5, 2026. Anthropic charges 1.1x when you require US-only inference on Claude 4.6 and later models.

Time of day. DeepSeek charges half price outside its peak hours, which are 01:00 to 04:00 and 06:00 to 10:00 UTC, Monday to Friday, excluding Chinese public holidays.

Announced changes. Google lists Gemini 4 Argon at $2 input and $10 output through November 30, 2026, and $4 and $20 from December 1, 2026; and Gemini 3.8 Flash at $0.75 and $3.75 through December 31, 2026, and $1.50 and $7.50 from January 1, 2027.

Free tiers and your data

Google's Gemini API has a free tier for some models, but its pricing table says data from the free tier is "used to improve our products" while paid-tier data is not. If you send customer or company data, use a paid tier and read the provider's data terms.

API or subscription?

An API bills every token; an assistant subscription charges a flat monthly price with usage limits instead. For one person who works in the app, the subscription is usually the simpler product: see our guides to ChatGPT pricing and Claude pricing, where Claude's page also runs three sample workloads through this calculator. The API is for software that calls a model. If you would rather avoid per-token bills altogether, a model on your own hardware has no price per token, only the cost of the machine: check what fits in how much VRAM an LLM needs.

How to choose

1. Estimate your real token counts with a sample of production prompts, including the system prompt and history.
2. Run them through the calculator with your cache share; check whether batch works for your use case.
3. Shortlist two or three models across price tiers and test quality on your own tasks, starting with the cheaper ones. Anthropic's guide to building agents describes routing easy or common questions to smaller, cheaper models and hard or unusual ones to more capable models.

How we checked

Prices were read from the providers' own pages on October 26, 2026: OpenAI's API pricing page (Standard and Flex/Batch tables), Anthropic's Claude pricing documentation, Google's Gemini Developer API pricing page and DeepSeek's models and pricing page. OpenAI's long-prompt rule for GPT-6.1 Sol comes from its model page, read on October 6, 2026. The calculator's formula and the example figures are covered by automated tests. We have not compared output quality, speed or rate limits.