Willison on Haiku 5.5: a cheap token is not a fixed task bill

October 7 original (about four minutes). Willison first compares Haiku 5.5's short-prompt sticker rates with GPT-6 Luna, then notes the Haiku long-prompt step and uses his token counter on the same large text: roughly 25% more tokens than Haiku 4.5. He generates an SVG under varying reasoning efforts and finds a roughly 36-fold price and 44-fold elapsed-time difference between his low and max runs; a pelican illustration says little about software correctness. He welcomes newly bundled API credit and the ability to stop requests when the balance ends.

Primary-source corrections and limits. Anthropic's model sheet puts the long-prompt price step at 100,000 tokens and calls the tokenizer increase ~30% on the same text; Willison's 25% is one sample. Luna has its own full-request step at 272,000 input tokens. Anthropic's credits rules say Claude Code and other cloud marketplaces are excluded, shared credits expire, auto-reload can draw purchased funds and exhaustion of the last balance stops API requests. This is a succinct dated operator take on pricing and bill control, not an experiment in accepted agent-coded change. The paired methods and September older-model counterexample are in cost brief and routing trial.

Hold: weekday coffee read when the four-item feed clears or a new 5.5 task-cost experiment arrives. Do not substitute the vendor launch announcement for Willison's argument.