OpenAI API Pricing 2026: Cost Calculator & Per-Token Rates

How token pricing actually works

Every LLM API bills the same way: it counts the tokens going in, counts the tokens coming out, multiplies each by a published per-million rate, and adds them up. There is no per-request fee, no seat licence and no monthly minimum. If you send nothing, you pay nothing.

A token is a chunk of text, usually a few characters. For ordinary English prose the working rule is:

• 1 token ≈ 4 characters
• 1 token ≈ 0.75 words
• 1,000 words ≈ 1,300–1,400 tokens
• A 500-page book ≈ 150,000–200,000 tokens

Those ratios get worse for code, JSON, XML, and any language that isn't written in the Latin alphabet — 20% to 60% more tokens for the same visible content is normal. Tokenizers also differ between providers and between model generations, so the identical prompt will not produce an identical token count on GPT-5, Claude Sonnet 5 and Gemini 2.5 Flash. When you compare providers on price, you are comparing rates on slightly different token counts, which is one more reason to validate on your own traffic rather than trusting a table.

Why output tokens cost 4–8x more than input tokens

This isn't arbitrary pricing — it reflects a real difference in what the hardware has to do.

The practical consequence: the shape of your workload matters as much as your volume. Two apps that each push 100 million tokens a month can have wildly different bills.

• Summarization — huge input, tiny output. Cheap per token processed, and the ideal candidate for the batch discount.
• Classification / extraction — small input, tiny output. The cheapest shape there is; run it on a nano/lite/flash model and the cost effectively disappears.
• Chat — medium input that grows every turn, medium output. Prompt caching is the main lever.
• Content generation and coding agents — long output, often plus reasoning tokens. The most expensive shape per request by a wide margin.

LLM API pricing table: cost per 1M tokens

Standard (synchronous) tier, USD per 1,000,000 tokens. Every figure below was read off the provider's own pricing page on July 28, 2026. Models whose price could not be verified against a first-party page are not listed.

Prompt caching and batch pricing: the two biggest discounts

Prompt caching

If you send the same prefix over and over — a long system prompt, a product catalogue, a document you're answering questions about, a conversation history — the provider can keep its processed form warm and skip the work next time. You get charged a fraction of the input rate for those tokens.

• OpenAI: cached input is 10% of the standard rate across the GPT-5 family (for example GPT-5 at $1.25 → $0.125). On the older GPT-4.1 and GPT-4o generations the discount is smaller — GPT-4o cached input is $1.25 against a $2.50 base, only 50% off.
• Anthropic: a cache hit costs 10% of the base input rate. Writing to the cache costs 1.25x base for a 5-minute TTL, or 2x for a 1-hour TTL — so a 5-minute cache pays for itself after a single hit, and a 1-hour cache after two.
• Google: context caching runs at roughly 10% of the input rate, plus a storage charge billed per hour of cached content.
• xAI and DeepSeek: both publish cached input rates. DeepSeek's is the most aggressive on this page — V4 Flash cache hits are $0.0028 per million against a $0.14 miss rate.

Drag the cache slider in the calculator above to see what this does to your own numbers. On a chatbot with a large fixed system prompt, a realistic 60% hit rate typically removes over half the input line.

Batch pricing

OpenAI, Anthropic and Google all take 50% off both input and output if you submit requests asynchronously and accept results within 24 hours. There is no quality difference — it's the same model, scheduled when there's spare capacity.

This is the most under-used discount in the industry, because most teams reach for a synchronous endpoint by reflex. If a human isn't waiting on the response, batch it: nightly summarization, backfills over historical data, bulk tagging and classification, eval runs, embedding refreshes, content pipelines. Caching and batching stack, so a batched job over a cached prefix can land near a fifth of the naive sticker price.

Worked examples: what four real app shapes actually cost

Monthly cost at standard synchronous rates, with the cache hit rate noted for each shape and no batch discount applied. These are computed with the exact same maths as the calculator — load any of them as a preset above and change the numbers to match your own app.

Seven ways to cut your LLM API bill

Frequently Asked Questions