Anthropic's New Haiku 5.5 Is Its Cheapest and Fastest AI Yet
Anthropic introduced Claude Haiku 5.5 on Oct. 7, 2026, completing the first rollout of its Claude 5.5 family with a small model it describes as its cheapest, fastest, and most capable to date. Anthropic
On price, Anthropic states Haiku 5.5 costs around 75% less to run on average than Haiku 4.5. It also describes Haiku 5.5 as its fastest model to date.
The model is built for jobs that need many small, cheap answers. Anthropic lists summaries, compactions, database queries, and classification requests as the design center. In those jobs, delay per call and price per piece of text drive total cost, especially in agentic loops that make hundreds or thousands of smaller calls.
Haiku 5.5 is also the first Haiku-class model with an adjustable effort setting to trade cost against smarts, like a volume dial. Low effort fits bulk triage and compaction. Higher effort fits cases where a small-model call sits on a critical path.
Anthropic positions Haiku 5.5 as a helper for larger models, not a replacement. It states the model pairs with Opus 5.5 and Sonnet 5.5 as a subagent on coding work. That setup is now common. A large frontier model plans and reviews. A group of smaller models handles retrieval, edits, test runs, and summarization.
Measured capability
On knowledge tests, Anthropic reports Haiku 5.5 scored 1620 on GDPval-AA v2.1 and 1578 on AA-Briefcase v1.1. On Humanity's Last Exam, it reports 45.9% without tools and 57.4% with tools.
The wider point for builders here is the gap between the tool-off and tool-on scores. It shows how much of the model's useful ability depends on retrieval, code execution, and structured tool access.
On computer use and visual reasoning, Anthropic reports 72.4% on the OSWorld 2.1 offline subset and 46.4% without tools on Chartography. Anthropic
On code and terminal work, Anthropic reports 46.4% on FrontierCode 1.1 Main and 39.2% on Terminal-Bench 4.0. These tests combine long-horizon planning, shell interaction, and repository-scale edits.
In practical terms here, scores in that range point to a supporting role rather than a lead role on hard software work. They point to a model that can handle smaller, bounded subtasks reliably enough to be used at scale.
In early testing, Asana reported over 30% reduction in latency for task completions and up to 2.5x faster inference per agent turn with Haiku 5.5.
The practical consequence here for multi-turn agents is that time per turn adds up fast. A 2.5x gain changes interaction design and timeout budgets.
Pricing and rollout
Anthropic paired the release with two platform pricing changes. It is halving the price of Claude Sonnet 5.5's cache reads. After that cut, Anthropic states Sonnet 5.5 now runs around 20% cheaper on most agentic work. Read pricing matters because prompt caching, repeated system prompts, and shared prefixes make re-reading saved context as important as base input pricing.
Anthropic is also introducing a new monthly API credit for Claude Max and Team subscribers to support building agents and applications on the Claude Platform. Details on credit amounts were not included in the verified announcement, but the structure is clear. Subscription spend converts directly into platform experimentation.
The release completes a sequence Anthropic had laid out through September. Anthropic dated its Opus 5.5 announcement to Sep. 22, 2026 and its Sonnet 5.5 announcement to Sep. 28, 2026, with Haiku 5.5 dated Oct. 7, 2026. Earlier statements said Sonnet 5.5 and Haiku 5.5 would follow Opus 5.5 in the coming weeks, with many of the same improvements to performance, efficiency, and safety. Anthropic
The wider family trend is efficiency. Anthropic describes Sonnet 5.5 as running 30% faster and costing up to 30% less for most work than Sonnet 5. It states Opus 5.5 performs at the level of Claude Fable 5.1 on most work while costing 40% less to run than Opus 5. Separately, Sonnet 5.5 is priced at $2 per million input tokens and $10 per million output tokens. Reuters
The broader context here is plain for teams planning agent budgets. Frontier ability still gets attention, but running cost now sets design. When a small model cuts operating cost by three quarters while cutting latency by a third or more, engineers hand off more work. They add checks, parallel tries, and compaction passes that were previously too costly to run.
In my view, the adjustable effort setting may prove as important as the price cut. It lets one model serve two roles. Bulk classifier in one call, careful subagent in the next, without redeploying endpoints or managing separate version pins. That simplifies routing and monitoring, which are often the real limits in live agent systems.
I have seen this cycle before at home, watching my own children move from treating every search as precious to treating generation as disposable, once speed and price crossed a line. Business developers make the same shift. Cheap, fast helpers invite more use and faster feedback. The risk is sprawl. The opportunity is systems that check their own work as a matter of course.
