Long context: does Haiku 5.5 hold up past its price jump?

Claude Haiku 5.5 is the cheapest current Claude model, but its price per token rises 5× once a prompt passes 100K tokens, and a coding agent that reads a real repository passes that line easily. So we built two prompts from our own code, one under the line and one over it, hid questions with exact answers at different depths, and asked all three 5.5 models.

Run October 14, 2026 · 30 calls · total spend $2.97

Results

4 lookups (find one value) and 1 two-step question (find two values and add them), graded by the number in the reply. The first call writes the prompt cache; the rest read it.

What it found

• Retrieval held past the line: Haiku found every looked-up value at both sizes, as did Sonnet and Opus.
• Haiku did not give the right two-step sum 2 times; 1 of those ran out of its output-token limit before answering. Sonnet and Opus answered it every time.
• On average, a cached Haiku call cost 12.1× more over the line for 2.5× the tokens. Over the line, Haiku still cost about 4× less than Sonnet per cached call.
• Once the prompt was cached, the average further question cost at least 12× less than the average first call, for every model and size.