Claude Haiku 5.5
High-volume & subagents. Classification, extraction, routing, summaries, subagents and browser use at very low cost.
Which reasoning effort setting?
controls how long Claude Haiku 5.5 thinks before answering. Token prices never change; higher settings spend more and raise task cost.
Start at Xhigh. It scores 34.6 against 36.5 at the top setting, for 40% less per benchmark task. Rule: cheapest setting within 2 points of the model’s best score.
At Xhigh, Haiku 5.5 thinks for about 97.3K tokens per task, then writes about 38.9K more as its answer: $0.25 per task, 5.7× Low, for +16.0 points.
Output rate: $0.50/1M at every effort level. Thinking tokens are billed as output.
| Effort | Index | Per task | Thinking | All output | × Low | Time |
|---|---|---|---|---|---|---|
| Low | 18.6 | $0.044 | 15.3K | 28.2K | 1.00× | 2.2 min |
| Medium | 24.6 | $0.092 | 32.1K | 53.3K | 2.10× | 4 min |
| High | 29.9 | $0.16 | 57.5K | 88.1K | 3.63× | 6.5 min |
| XhighStart here | 34.6 | $0.25 | 97.3K | 136K | 5.66× | 9.5 min |
| Max | 36.5 | $0.41 | 173K | 226K | 9.38× | 14.9 min |
Per task of the Artificial Analysis Intelligence Index: the average over its nine evaluations, counting input, cache, and output tokens at list prices. These are long, agent-style benchmark tasks, much longer than a chat message.
Where Claude Haiku 5.5 sits on the frontier
Intelligence Index against one 100K-input, 10K-output request.