Reading a model price list like an engineer
The headline rate on a model page is the input price, and input is usually the cheapest thing about a request. Four other numbers decide what you actually pay.
A model's advertised price is a rate per million input tokens, and input is rarely where the money goes. Output is typically priced several times higher than input, so a request's cost is dominated by how much the model writes, not how much you send. Cached input is roughly an order of magnitude cheaper than fresh input — 0.1x the base rate on Anthropic's published multiplier{{2}} — which makes prompt structure a pricing decision. Reasoning models bill internal tokens you never see. And several providers apply a higher rate above an input threshold that can reprice the entire request rather than only the tokens above the line, published as separate short- and long-context columns.{{1}} Estimate cost as (input × input rate) + (output × output rate), check the cache and threshold terms, then measure — settled usage is the only figure that is not an assumption.
Output is the expensive half
Model pricing is quoted per million tokens, split at minimum into input and output. Those two rates are not close to each other. Output is routinely priced at three to five times input, and on some routes considerably more. A model advertised as inexpensive on the strength of its input rate can be the expensive choice for anything that writes at length.
This inverts the intuition most people bring from storage and bandwidth, where sending more is what costs. Here, a request that ships a large document and asks for a one-line answer is cheap. A request that ships two sentences and asks for a full report is not. Verbosity is the cost driver, and verbosity is something you control through the prompt and the output cap.
The practical consequence is that cost optimization mostly means asking for less writing, not sending less text. Constraining format, capping output tokens, and asking for a table instead of prose all move the number that matters.
- Estimate: (input tokens ÷ 1M × input rate) + (output tokens ÷ 1M × output rate).
- Set an output cap on every production call. It bounds both cost and latency.
- Ask for structure. A JSON object or a table costs a fraction of the same content as prose.
- Compare on your own ratio. A model's rank on cost changes completely between summarization and generation workloads.
Two line items that do not appear in the headline
Cached input. Most providers now price previously-seen input far below fresh input — by roughly an order of magnitude. Anthropic publishes the multiplier directly: cache reads bill at 0.1x the base input rate, and cache writes at 1.25x for the five-minute cache or 2x for the one-hour cache.[2] That turns prompt architecture into a cost lever: a long, stable system prompt and document preamble placed at the front of every request, with only the variable part at the end, lets the stable portion bill at the cached rate on repeat calls. The same tokens in a different order pay full price, because caches match on prefixes. Some providers also bill a one-off cache write, so the first call is dearer and the saving arrives from the second onward.
Reasoning tokens. Reasoning models generate internal tokens before producing an answer. They are billed as output and are usually not shown to you, so a short visible reply can carry a long invisible one. Where reasoning effort is configurable, it is a direct cost control; where reasoning is mandatory, every trivial request pays for it, which is a good reason not to route simple traffic to a reasoning route.
Both of these are invisible in a per-million comparison table, and both can change the ranking of two models that looked a token apart on the headline rate.
| Line item | Typical relationship to base input | What controls it |
|---|---|---|
| Input | The advertised rate | Prompt and document length |
| Output | Several times input | Response length; output cap; requested format |
| Cached input | 0.1x base input on Anthropic's published multiplier | Stable prefix; consistent prompt ordering |
| Cache write | 1.25x base input for a 5-minute cache, 2x for an hour | How often the cached prefix changes |
| Reasoning tokens | Billed at the output rate, usually unseen | Reasoning effort setting, where it is configurable |
| Threshold tier | A higher rate above an input threshold | Total input length crossing the provider's line |
The threshold that reprices the whole request
Several providers publish a second, higher rate that applies above an input threshold — commonly somewhere in the hundreds of thousands of tokens. Google splits its per-model rates at a 200,000-token prompt, listing separate figures for prompts under and over that line.[3] OpenAI's schedule makes this explicit, listing short-context and long-context input, cached input and output rates as separate columns for the same model.[1] The detail that surprises people is the scope: on these schedules the higher rate frequently applies to the request's entire input, cached input and output, not only to the tokens above the line.
That makes the threshold a cliff rather than a slope. A request at slightly under the line and a request at slightly over it can differ in cost by far more than the few thousand tokens between them. Any workload whose prompt size varies near a threshold will produce a cost distribution that looks erratic until you find the line.
It is worth knowing where the thresholds are for the routes you use, and worth deliberately staying under them where the extra context is not earning its place. This is also a case where retrieval beats long context on economics alone: a shorter, relevant prompt does not merely cost less per token, it can stay on the cheaper side of the cliff entirely.
From published rate to what you are actually charged
Platforms that resell model access add a layer between the provider's list price and your bill. That layer deserves the same scrutiny as the rates beneath it: what the markup is, whether it varies by route or routing mode, whether it is disclosed as a percentage of provider cost or folded into an opaque unit, and whether the platform's own currency has a stable relationship to money.
Where a platform prices in credits, the questions are whether credits expire, whether the credits-per-currency rate varies with purchase size, and whether the rate you were quoted at purchase is the rate applied at settlement. A published conversion table you can do arithmetic against is the minimum standard; anything less means you cannot forecast.
Finally, none of the above is a bill. Estimates are estimates. The number that counts is settled usage recorded against the completed request, and any platform serious about cost transparency shows you that per request, after the fact, next to the route that produced it.
- Is the markup published as a percentage of provider cost, and is it the same across routing modes?
- Do purchased credits expire? Do promotional grants expire on different terms?
- Does the conversion rate change with purchase size, and is that table public?
- Can you see measured usage and exact cost for an individual completed request?
Frequently asked questions
Why is output more expensive than input on LLM pricing?
Generating tokens requires a forward pass per token, while input is processed in parallel, so output is the more computationally expensive operation. Providers price accordingly — output rates are commonly several times input rates, which makes response length rather than prompt length the main cost driver for most workloads.
How do I estimate the cost of an API call?
Multiply input tokens by the input rate per million and output tokens by the output rate per million, then add them. Adjust for cached input where a stable prefix repeats, and for reasoning tokens on reasoning models, which are billed at the output rate even though you never see them.
What is prompt caching and how much does it save?
Providers cache the repeated prefix of a prompt and bill it at a much lower rate on subsequent calls, at a published multiplier — 0.1x the base input rate on Anthropic's schedule. It only applies to a stable prefix, so it rewards putting fixed instructions and documents at the front and variable content at the end. Some providers charge a one-off write fee, so savings begin from the second call.
What are context threshold tiers?
A higher rate that applies once a request's input crosses a published length threshold. The important detail is that on many schedules the higher rate applies to the entire request rather than only the tokens above the threshold, which makes it a cost cliff rather than a gradual increase.
Do reasoning tokens cost money if I never see them?
Yes. Internal reasoning tokens are billed at the output rate whether or not they are returned to you. Where reasoning effort is adjustable it is a direct cost control; where reasoning is mandatory on a route, even trivial requests pay for it, which is a reason to send simple traffic elsewhere.
Sources and evidence
Primary and authoritative sources used for factual claims. Company research and executive forecasts are labeled as such in the article.
- 1
- 2Prompt caching: cache-write and cache-read pricing multipliersAnthropic · 2026-09
- 3
- 4Kendr credit pricing, markup policy and top-up bandsKendr · 2026-09
- 5Kendr public model catalog: live per-model credit ratesKendr · 2026-09
- 6