LLM pricing FAQ

For a typical request of 4,000 input and 800 output tokens, somewhere between a small fraction of a cent and about $0.264, depending entirely on the model. On Claude Opus 5 that request costs $0.040; on Claude Haiku 4.5, $0.00800; on Qwen3.7 Flash, $0.00022. Price your own numbers in the API cost calculator.

Qwen3.7 Flash from Alibaba has the lowest input rate tracked here at $0.03 per 1M tokens. On a blended 3:1 input-to-output basis, Qwen3.7 Flash comes out cheapest overall. The full ranking is here, and every model has its own page with worked cost examples.

GPT-5.5 Pro at $30.00 per 1M input tokens, which is 1000x the cheapest rate tracked. Extended-compute and frontier tiers sit well above the rest of the market; the gap between the top and bottom of the range is far wider than the gap between any two mainstream models.

That is exactly the unit every provider quotes, which is why it is the unit used across this site. It ranges from $0.03 to $30.00 for input and considerably more for output. One million tokens is roughly 750,000 words, or about 1,500 pages of prose.

About 1.3 tokens per English word, or roughly 4 characters per token. Code, non-English languages, and heavily formatted text all tokenise less efficiently, sometimes much less. The token counter estimates from a real sample rather than a rule of thumb.

Because generating them is a fundamentally different operation. Every output token requires its own forward pass through the model, one at a time, while an entire input prompt can be processed in parallel. That is why output typically costs four to six times input across almost every provider, and why capping output length is usually a bigger cost lever than shortening prompts.

A discount on input tokens the provider has processed for you recently. When a request begins with the same prefix as a previous one, that prefix is billed at a much lower rate. 27 of the 60 models tracked here publish one, at a median discount around 90%. It only applies to input, and only to the prefix, so ordering matters: anything variable must come after everything stable. Size it with the prompt caching savings calculator.

An asynchronous queue: you submit work and collect results later, usually within hours, and the provider discounts it in exchange for scheduling flexibility. 18 tracked models offer one, almost always at a flat 50% off both input and output. If nobody is waiting on the result, it is close to free money. The batch API savings calculator sizes it.

The usual culprits, in order: output length drifting upward as prompts change, system prompts accumulating instructions, retries on failed calls (which still bill for what they generated), and agent loops making more model calls per user action than intended. All four raise the bill without raising the request count, which is why per-request cost is worth tracking separately from total spend.

Several do. Some apply a surcharge above a threshold, often 200K tokens, and bill the entire request at the higher rate rather than just the excess. Others price long context the same as short. Where a tracked model has a threshold, it is noted on that model's page, and the context window cost calculator shows what filling a window actually costs.

It depends on what your workload is shaped like, and that is not a dodge. A model with a cheap input rate and expensive output is excellent value for retrieval and document analysis and poor value for drafting. Work out your input-to-output ratio first, then compare on that basis in the model switching savings calculator rather than on the headline input price.

Only if it does the job. A cheaper model that needs two attempts, or produces output someone has to fix, costs more than the model that got it right first time, and none of that correction cost shows up on the API bill. Treat the calculated saving as the budget available for evaluating the swap, not as the answer.

Yes, consistently and by a wide margin at every tier. DeepSeek, Alibaba, Zhipu, Moonshot, ByteDance and Baidu all sit well below the US price band. Price is only one input to the decision though: data residency, latency from your region, and API reliability decide whether a provider is usable at all, and none of them appear on a rate card. The price index quantifies the gap.

Only at sustained high utilisation. A rented GPU bills by the hour whether or not you send it traffic, while an API bills per token, so self-hosting wins when the GPU is busy and loses badly when it is idle. Add your own engineering time, which is usually what decides it. The GPU rental cost calculator finds the crossover point.

Four levers, roughly in order of effect: enable prompt caching and order your prompts so the stable part comes first; move anything latency-tolerant to a batch API; cap output length, since output is the expensive half; and cut the number of model calls per user action, which is the variable most often larger than people assume. Together those routinely beat what switching models would have saved, without the evaluation risk.

It was last verified on 2026-08-16, and that date is stamped on every page rather than hidden in a footer. All 60 models across 17 providers come from one dataset, so no two pages here can show conflicting numbers. See the methodology page for the process.

Yes. The full dataset is published as JSON and CSV under CC BY 4.0: use it commercially, chart it, build on it, with attribution and a link back.

All of them, with no signup and no account. They run entirely in your browser, so nothing you type is sent anywhere.

Tell us through the contact page. Providers change rates without much notice and a pricing site that looks authoritative while being wrong is worse than no pricing site, so corrections are genuinely welcome.