Anthropic Cut Its Cheapest Model's Price 75% and Gave It Something No Haiku Ever Had: an Effort Dial
Haiku has always been the model you reach for when a task is simple and you're running it a million times: a lookup, a short summary, a quick classification. The effort dial changes what 'simple' can mean, because the same model ID can now scale up for a harder sub-task without you switching to Sonnet or Opus and paying their price.
The benchmark jumps back that up. On Anthropic's own evaluation suite, Haiku 5.5 doesn't just get cheaper. It gets dramatically more capable at tasks Haiku 4.5 could barely do at all.
What actually got better
72.4%
on OSWorld 2.1's offline computer-use benchmark, versus 15.7% for Haiku 4.5
~75%
cheaper to run than Haiku 4.5 on average, by Anthropic's own accounting
The gap isn't universal. On GDPval-AA v2.1, Anthropic's broad knowledge-work benchmark, Haiku 5.5 scores 1620 (Elo) against Sonnet 5.5's 1840, so the flagship still wins on genuinely hard, open-ended work. Haiku's gains are concentrated in the kind of bounded, repetitive tasks it was already built for, just done far better and far cheaper now.
Five effort levels, one model ID
Haiku 5.5 ships with Low, Medium, High, Xhigh, and Max effort settings. Low behaves like a classic fast-and-cheap Haiku call. Max lets it reason longer and burn more tokens on a sub-task that's quietly harder than it looks, without switching models or rewriting the integration.
- ✦Asana reported a 30% latency reduction on task completions after switching, per Anthropic's announcement.
- ✦HubSpot said Haiku 5.5 scored 92.8% on its internal CRM task suite, the best result it has seen from any Haiku-class model.
- ✦Box reported Haiku 5.5 scored 11 points higher than Haiku 4.5 at roughly half the latency in early testing.
- ✦Rogo described it as a good fit for short, high-volume work, including lookups, subagents, and summaries, the exact use cases Haiku has always targeted.
When to actually reach for the cheap model
- 1Default to Low or Medium effort for high-volume, low-stakes calls like lookups, short summaries, and classification, where Haiku's old speed-first behavior is still the right trade.
- 2Bump individual sub-tasks to High or Xhigh effort inside a larger automated pipeline instead of routing the whole pipeline through Sonnet or Opus.
- 3Re-test any integration built on Haiku 4.5. Requests that still set temperature, top_p, top_k, or an assistant prefill can return a 400 error on Haiku 5.5; drop them during migration.
- 4Watch cost at scale, not per-call. A 75% average drop compounds fast across millions of calls, which is exactly the workload Haiku-class models are meant for.
How much cheaper is Claude Haiku 5.5 than Haiku 4.5?
Anthropic says it costs about 75% less to run on average. Standard pricing is $0.10 per million input tokens and $0.50 per million output tokens for prompts up to 100,000 tokens, rising to $0.50 and $2.50 above that.
What are the effort levels on Claude Haiku 5.5?
Low, Medium, High, Xhigh, and Max. It's the first Haiku-class model with this adjustable setting, letting a single model trade cost for intelligence on a per-request basis instead of forcing a switch to a larger model.
Is Claude Haiku 5.5 as capable as Sonnet or Opus?
No, and Anthropic doesn't claim it is. On GDPval-AA v2.1, Haiku 5.5 scores 1620 Elo versus Sonnet 5.5's 1840. Haiku's real gains are on bounded, repetitive tasks like computer-use and terminal benchmarks, where it jumped from single digits to 70%-plus.
When did Claude Haiku 5.5 launch?
October 7, 2026, per Anthropic's own announcement.
Will my existing Haiku 4.5 integration break on Haiku 5.5?
It can, if your requests still set temperature, top_p, top_k, or an assistant prefill. Those parameters can trigger a 400 error on the new model and should be dropped during migration.
Try the tools ✦
Free browser tools that never upload your files.