Claude Haiku 4.5 vs Claude Haiku 5.5
Compare with other modelsHaiku 5.5 is a great step forward from Haiku 4.5. It doesn't only raise the quality score from 69 to 84; it also comes at less than a fifth of the cost. It takes big steps on several of the dimensions we test: grounding goes from 67 to 96, and tool use from 72 to 86.
But maybe the biggest improvement is the share of conversations with no unsupported claims, which goes from 19% to 62%. That is an important and substantial improvement, and it makes Haiku 5.5 much more useful for CX agents.
| Metric | Anthropic Sep 8, 2026 | Anthropic Oct 7, 2026 |
|---|---|---|
| Overview | ||
| Quality score | 69 #12 of 17 | 84 Best among selected models #4 of 17 |
| Median reply | 2.6s Served by Amazon Bedrock | 2.6s Served by Claude Platform on AWS |
| Cost per conversation | $0.043 | $0.0076 Best among selected models |
| Quality scores | ||
| Tool use | 72 | 86 Best among selected models |
| Task completion | 72 | 84 Best among selected models |
| Context retention | 82 Best among selected models | 81 |
| Grounding | 67 | 96 Best among selected models |
| Safety | 93 | 94 Best among selected models |
| Hallucinations % of conversations with no unsupported claims detected | 19% | 62% Best among selected models |
| Running stats | ||
| p95 reply | 6.4s Best among selected models | 7.2s |
| Consistency % of repeat runs ended the same way | 64% | 87% Best among selected models |
| Output per turn | 178 tokens, median | 338 tokens, median |
Reply times cover the model call only. Cost estimates use token usage at list prices.
This whole report is one Voxli workspace: simulated customers, assertion checks, and a frozen, versioned test set that reruns when new models ship.
Test set v3f-2026-09 · edition 2026-09-08