How long-context pricing works
Processing a very long prompt costs a provider more than a short one, and some vendors pass that on with a second rate card. Each request is measured on its own: below the threshold it bills at the standard rates, above it at the long-context rates.
The important detail is what the higher rate applies to. xAI's pricing page states that past 200,000 prompt tokens the long-context rate is billed for all tokens in the request. Anthropic says the same of Claude Haiku 5.5, its one current model with a prompt-length step: a request over 100,000 tokens pays the higher prices, and the prompt length counts every input token, including cache reads and cache writes. Google publishes separate rates for prompts up to and over 200,000 tokens on its Pro models, covering input, output and caching. A threshold is therefore a cliff, not a ramp.
Who charges it, and how much
- OpenAI lists short- and long-context columns for its newest models, with the long-context card past 272K input tokens. On GPT-6 Sol the input step is $2 to $4 per million, and output goes from $10 to $15 — a smaller multiplier on the output side.
- xAI doubles its standard rates past 200,000 prompt tokens on the Grok models that have a long-context tier.
- Google steps its Pro models past 200,000 tokens; its Flash models in our index carry no step.
- Anthropic is the exception at the top end: Claude 4.6 and later models, except Haiku 5.5, include the full one-million-token window at standard pricing — Anthropic's own example is that a 900,000-token request bills at the same per-token rate as a 9,000-token one.
In our indexLong-context rates in our index
Counted at build from our price index: the published long-context input rate against the standard one among the 167 current public models we price, from 33 vendors. Each rate comes from the vendor's own price page, read between 2026-07-28 and 2026-10-09. How we collect prices.
Staying under the line
Because the step usually covers the whole request, the cheapest long prompt is one that stays just below the threshold. Retrieval helps: RAG sends the passages that matter instead of the whole corpus. So does trimming conversation history in long agent sessions — summarizing old turns or clearing stale tool results — so that a session does not drift past the line on its fortieth step.
Caching does not get you out of it. At Anthropic, cache reads and writes count toward the prompt length that decides the price, and a request over the line pays the higher rate even when part of its prompt is a cache hit. On models with no step at all, a long prompt is only as expensive as its token count.