Anthropic released Claude Haiku 5.5 on October 7, 2026, positioning it as the fastest and most efficient model in the Claude 5.5 family, built for high-volume, cost-sensitive work. Two details matter more than the speed claims: pricing now depends on prompt length, and this is the first Haiku with effort controls for tuning cost against intelligence per task.
The pricing structure is the story
For prompts up to 100K tokens, Haiku 5.5 costs $0.10 per million input tokens and $0.50 per million output tokens. Cross 100K and both numbers jump fivefold to $0.50 input and $2.50 output. Prompt caching saves up to 90% and batch processing saves 50%, which softens the edge but does not erase it.
If you are building a routing layer, this creates a hard budget cliff rather than a smooth curve. A summarization job with a 90K-token context bills at the cheap rate; the same job at 110K tokens bills at five times the rate. Long-context workloads that drift past the boundary silently change unit economics. Worth checking whether your retrieval stack actually needs to stuff that much into the prompt — the KV-cache savings math on DeepSeek’s flash models makes a similar point: context size is a billing decision, not just an architecture one.
What the effort controls unlock
Per-task effort tuning is the other lever. Teams can dial intelligence down on easy calls and up on hard ones within the same model, which pairs naturally with routing: let a bigger model plan, hand well-defined subtasks to Haiku at low effort, and run many agents in parallel.
That pattern shows up in the customer quotes Anthropic published. Cognition reports that Devin Fusion with Haiku 5.5 as the sidekick holds a FrontierCode score of 66.2 while cutting cost and latency. Atlassian describes exactly the division of labor: a bigger model builds the deck while a Haiku subagent pulls a segment revenue line from a 10-K — accurate enough to trust, cheap enough to run constantly.
The numbers other builders measured
The quoted evals are worth reading as signals, not guarantees. HubSpot tested Haiku 5.5 on simulated CRM tasks and got 92.8% averaged over three runs — its best result on that suite — with the fastest completion and lowest false-positive rate on an audit task about stale records. Box’s early testing showed an 11-point quality gain over Haiku 4.5 at about half the latency. Another customer (the quotes are anonymized beyond company name) saw a statistically significant improvement on document Q&A, 0.84 vs. 0.76 across 400 queries on roughly 8M weekly production calls.
Anthropic also claims a significant step up over Haiku 4.5 across coding, tool use, computer use, and agents, but the supplied product page does not include the benchmark tables behind that claim.
Availability
Haiku 5.5 is selectable on Claude.ai for Free through Enterprise plans, runs natively on the Claude Platform (claude-haiku-5-5), and is on AWS, Google Cloud, and Microsoft Foundry. It is also in Claude Code.
For your own roadmap, the practical next step is small: pick one high-volume, well-scoped task — classification, form filling, a lookup subagent — and benchmark Haiku 5.5 against whatever handles it today, with effort controls set low. The tiered pricing means the test only gets expensive if your prompts are long, so measure your prompt lengths before you measure the model.