Claude Haiku 4.5 vs Haiku 5.5: version, capacity & price changes
A Haiku version change affects more than the model name. The reviewed 4.5 and 5.5 entries have different context ceilings, output limits and pricing structures, so a migration needs a fresh request budget.
Pin the API version you intend to test
The 4.5 entry records claude-haiku-4-5-20251001 and its documented alias; the 5.5 entry records claude-haiku-5-5. Version identity is included in the table so that future benchmark evidence or trial receipts can be matched to the model actually used.
Do not extrapolate the base-rate reduction
Haiku 4.5 lists $1 input and $5 output per million tokens. Haiku 5.5 lists $0.10 and $0.50 at base tier, but full-request higher rates apply above 100,000 input tokens. Compare a complete request at its correct tier instead of applying the base ratio to every workload.
Bigger ceilings still need regression checks
The reviewed context/output limits move from 200,000/64,000 in 4.5 to 1,000,000/128,000 in 5.5. That is a documented capacity change. It does not prove identical behavior on existing prompts or a higher quality score; use a regression set before changing an application.
Documented specifications and standard direct API rates, manually checked 2026-10-09. Evaluation suggestions are editorial guidance; no measured comparative quality or speed is claimed.
Unknown fields are not evidence of equal capabilities. Specifications describe support, not comparative answer quality.
BENCHMARK EVIDENCE
No reliable data available
Comparable results with confirmed reuse permission have not been published here yet.
A published result must include the exact model version, benchmark version, original value and unit, test date, conditions and source. No scores or overall rankings are inferred.
Text tokens only. Output includes thinking tokens. Cache reads are a subset of input. Cache creation, tools and media fees are excluded.
KNOW YOUR API COST
A small prompt. A clearer budget.
Same token counts for both models. Standard paid text rates in USD. Output must include all billable reasoning tokens.
Cached input is included in total input. No cache creation/storage, tools, media, taxes or regional/priority/Batch fees. Tokenizers differ; equal token counts do not imply equal text length.
4 estimated credits / comparison4,000 credits across 1,000 comparisons.Reference estimate only. Playground uses verified OpenRouter charges; 1,000 credits = $1. A + B round up together once for each comparison.
See how your prompt behaves.
Choose an enabled official model pair in the Playground. Your maximum reservation is shown before running; confirmed usage determines the final charge.