The Price Changes Partway Through the Prompt
A per-million-token price is not one price. On two of the three large platforms it is two prices, with a size threshold between them.
This note records where those thresholds sit and what the rate becomes on the far side. All figures were read from the vendor documentation source on October 27, 2026.

The thresholds matter because they apply to the whole request, not to the tokens above the line. Crossing by one token reprices everything.
One model charges five times more past 100,000 tokens
The Anthropic pricing table lists Claude Haiku 5.5 twice. The first row is for prompts up to 100,000 tokens. The second row is for prompts over 100,000 tokens.

Below the line, input is $0.10 per million tokens and output is $0.50. Above it, input is $0.50 and output is $2.50.
The caching rates move with them. Five-minute cache writes go from $0.125 to $0.625, one-hour writes from $0.20 to $1, and cache hits from $0.01 to $0.05.
That is a clean factor of five on every line. A 101,000-token prompt to Haiku 5.5 costs roughly five times what a 99,000-token prompt costs, per the published table.
Haiku is the cheap, high-volume model in the family. The threshold is doing something deliberate: it keeps the headline rate low while pricing long documents at a different model's economics.
The OpenAI cliff sits at 272,000 input tokens
OpenAI puts the same idea in a footnote rather than a second table row. The model pages for GPT-6 Astra and GPT-6 Luna both carry the identical line.
Prompts with more than 272K input tokens are priced at 2x input and cache rates, and 1.5x output, for the full request. The phrase for the full request is the important part.
On GPT-6 Astra the standard rates are $10 input, $1 cached input, $12.50 cache writes and $50 output per million tokens, read October 27, 2026.
Apply the footnote and a 300,000-token prompt is billed at $20 input and $75 output. That is editorial arithmetic from the published multipliers, not a second table the vendor prints.
GPT-6 Luna carries the same threshold at a twentieth of the price: $0.10 input and $0.50 output, doubling and rising by half past 272K.
Both models publish a 1,050,000-token context window. So roughly three quarters of the advertised window sits on the far side of the price change.
The flagship Claude models have no cliff at all
Anthropic states the opposite policy for everything except Haiku 5.5, and states it plainly enough to quote.
Claude 4.6 and later models, apart from Haiku 5.5, include the full 1M token context window at standard pricing. A 900k-token request is billed at the same per-token rate as a 9k-token request.
Prompt caching and batch discounts apply at standard rates across the full window, per the same page. There is no long-context surcharge to plan around.
This is the clearest structural difference between the two platforms. A long-document workload is predictable on one and stepped on the other.
One Anthropic exception is worth noting. Fast mode pricing applies across the full context window, including requests over 200k input tokens, so the premium there is a mode rather than a size.
Gemini moves the whole price instead
The Gemini API pricing page takes a third approach. Gemini 4 Argon has no size threshold listed, and a dated price rise instead.
Input is $2.00 per million tokens through November 30, 2026, and $4.00 starting December 1, 2026. Output is $10.00, then $20.00. Context caching is $0.10, then $0.20.
Every published rate for the model doubles on the same day. The page prints both numbers side by side, which is more honest than most dated increases.
Two further notes sit above the table. Gemini 4 Argon is not available on the free tier, and every token the model processes is counted and billed.
That second line includes thinking tokens and, when grounding is used, internal tool system instructions and function declarations. A reasoning-heavy request bills more than its visible output suggests.
The dated step is the same device catalogued in the introductory price note, now applied to a flagship model rather than a seat plan.
The multipliers that stack on top
Size is not the only thing that moves a token rate. Each platform publishes separate multipliers that compound with it.
- OpenAI: cache writes are billed at 1.25x the uncached input rate; Batch and Flex are 50 percent of Standard; Fast mode is 2x the applicable rates
- OpenAI GPT-6 Luna: regional processing adds a 10 percent premium where available
- Anthropic: for Claude 4.6 and later, inference_geo set to us applies a 1.1x pricing multiplier, while global is standard
- Anthropic on partner clouds: regional and multi-region endpoints carry a 10 percent premium over global endpoints
- Anthropic on AWS and Microsoft Foundry: usage is rated in USD then converted to Claude Consumption Units at $0.01 per unit, so the invoice shows one line
- Google: Gemini 4 Argon supports implicit caching only, so explicit cache control is not available as a cost lever
Stacked, these change a quoted rate by more than the size threshold does in some cases. A Fast mode request on a long prompt is paying two multipliers at once.
How to price a long prompt before you send it
- Find the threshold for your exact model, not the family, because within Claude it applies to Haiku 5.5 and to nothing else
- Read whether the multiplier applies to the whole request or only to the tokens above the line; on OpenAI it is the whole request
- Check the output multiplier separately, since OpenAI raises output by 1.5x where it raises input by 2x
- Add the routing multiplier if you need a specific region, because 10 percent is applied before any size effect
- If the workload is one long document and a short question, price the cached path too, since cache hit rates are the only line that gets cheaper
- Write down the date you read the rate, because at least one flagship model has a published rate change on December 1, 2026
Bottom line
Three platforms, three shapes. Anthropic charges one rate across a million tokens on its flagship models and five times more past 100,000 on its cheapest one.
OpenAI charges double past 272,000 input tokens on both GPT-6 models read for this note. Google charges one rate until November 30, 2026 and double after it.
None of that is hidden. All of it sits below a headline figure that a spreadsheet will record as a single number.
The monthly ceilings this spend runs into are collected in the monthly cap note, and the cache rates are in the cached-input price note.
The batch route that halves these rates is in the fifty percent discount note.
For teams buying seats instead of tokens, see the Claude Team review, the Gemini for Google Workspace review and the shortlist of AI assistant plans.
What this brief prints and what it cuts is in issue zero.
Sources
- Anthropic model pricing (Claude Haiku 5.5 long context rows) - checked October 27, 2026
- OpenAI GPT-6 Astra model page - checked October 27, 2026
- OpenAI GPT-6 Luna model page - checked October 27, 2026
- Gemini Developer API pricing (Gemini 4 Argon) - checked October 27, 2026