The New Haiku Finishes More Calls and Says Less
Claude Haiku 5.5 costs a tenth of Haiku 4.5 and finishes more of our simulated phone calls. It also stopped speaking before tool calls. One line in the system prompt fixes that.
Anthropic released Claude Haiku 5.5 on October 7. We ran it through the same tests as Claude Haiku 4.5 the next day. It costs a tenth as much: $0.50 per million output tokens instead of $5, and $0.10 in instead of $1.
It finishes more real calls than Haiku 4.5, and its board score fell from 89 to 73. Both are true. Much of the gap comes from one habit, and that part you can fix in the system prompt.
What got better
A real call is a full conversation: a simulated caller with a goal, a booking or order system behind real tools, and about one tool call in seven failing on purpose. A call counts only if the system ends in the right state, the caller is told the outcome, nothing false is claimed, and no machine text is read aloud.
Haiku 5.5 gained twelve points there. It also invented nothing on our out-of-policy questions, and in 180 calls it never read a tool call or JSON aloud.
Its weakest call type is the front desk, where callers ask to be routed or transferred. It finished 67% of those, against 89–92% on scheduling, orders, intake and payments.
What got worse
Before almost every tool call, Haiku 4.5 said something first: “I’ll help you with that. Let me look up your order first.” Haiku 5.5 dropped the habit and goes straight to the call. On a phone line, the caller hears nothing until the lookup comes back.
Our board score weighs speed, task completion, dead air and fabrication. Dead air went up, and first output on a full agent turn slowed from 391 ms to 579 ms at the median. Those two account for the whole drop.
Fix it in the prompt
The silence is a default, not a limit. We added one line to the same system prompt:
Before you call a tool, say one short sentence telling the caller what you are about to do, then call it.
Then we re-ran the opening turn of all 19 tasks, three times each, on both models.
With the line, Haiku 5.5 says “I’ll look up order B5678 first to check that it’s within our 30-day refund window” before the call, and goes quiet on about one tool call in ten.
If you move from 4.5 to 5.5, add the line, or anything like it, to your agent’s prompt. If your stack already plays a hold tone or a “one moment” while tools run, you may not need it.
We measured the silence only. Task and real-call scores with the line are not measured yet.