Claude Haiku 5.5

Claude Haiku 5.5 is Anthropic's fastest, cheapest, and most capable small model, announced on October 7, 2026. It is designed for high-volume, cost-sensitive tasks and reliably handles quick, repetitive workloads such as summaries, compactions, database queries, and classification requests. The model also pairs well with Anthropic's larger models, Opus 5.5 and Sonnet 5.5, acting as a subagent on coding work. Because it is Anthropic's fastest model to date, it is positioned for speed-sensitive workloads including live customer support and browser use, giving developers a small model they can run often rather than sparingly. The launch addresses a practical problem in production AI: many of the requests that dominate real traffic are short, repetitive, and narrow, yet teams have historically paid frontier-model prices to serve them. Anthropic notes that prompts up to 100,000 tokens make up around 90% of requests to the previous Haiku model, which means the bulk of day-to-day workloads are the kind that do not need a large model. Claude Haiku 5.5 is priced far lower than Haiku 4.5: on average, it now costs around 75% less to run. By footnote, it is priced 90% lower than Haiku 4.5 for requests up to 100,000 tokens and 50% lower for requests over 100,000 tokens. That calculation also accounts for the model's updated tokenizer, similar to those of Sonnet 5.5 and Opus 5.5, which uses slightly more tokens per task. Claude Haiku 5.5 brings substantial benchmark gains over Haiku 4.5 across knowledge work, computer use, reasoning, agentic coding, and visual reasoning. On GDPval-AA v2.1, a knowledge-work evaluation, it scores 1,620 versus 735 for Haiku 4.5, 1,437 for GPT-6 Luna, and 1,840 for Sonnet 5.5. On AA-Briefcase v1.1 it scores 1,578 against 614 for Haiku 4.5, 1,336 for GPT-6 Luna, and 1,824 for Sonnet 5.5. On the OSWorld 2.1 computer-use benchmark (offline subset) it reaches 72.4% versus 15.7% for Haiku 4.5, 48.9% for GPT-6 Luna, and 83.9% for Sonnet 5.5. On Humanity's Last Exam, a multidisciplinary reasoning test, it scores 45.9% without tools and 57.4% with tools, compared with 10.2% and 18.7% for Haiku 4.5 and 56.9% and 64.5% for Sonnet 5.5. On Terminal-Bench 4.0 agentic coding it scores 39.2% versus 0.0% for Haiku 4.5 and 70.6% for Sonnet 5.5, and on FrontierCode 1.1 (Main) it reaches 46.4%, while Chartography visual reasoning lands at 46.4% without tools versus 6.4% for Haiku 4.5. Full evaluation details are documented in the Haiku 5.5 System Card. Haiku 5.5 is the first Haiku-class model to come with an adjustable effort setting. As with Anthropic's other models, users can decide whether to optimize for cost or intelligence by choosing an effort level, with options charted across Low, Med, High, Xhigh, and Max. Anthropic publishes accuracy-versus-cost charts for three benchmarks at each setting: OSWorld 2.1 for computer use, GDPval-AA for knowledge work, and Humanity's Last Exam for multidisciplinary reasoning. OSWorld 2.1 measures how well agents can operate a real computer to finish long, multi-step tasks; GDPval-AA v2.1 evaluates agents on real-world professional work across 44 occupations; and Humanity's Last Exam tests expert-level academic knowledge and reasoning. This effort setting lets teams tune the trade-off between per-attempt cost and task accuracy without switching models or rewriting their workloads. Anthropic reports that Claude Haiku 5.5 shows major improvements across almost all of its alignment evaluations relative to Haiku 4.5, with far fewer instances of misaligned behavior and a lower willingness to cooperate with misuse; a dedicated system card describes the evaluation process and results in more detail. The model's cybersecurity safeguards are more restrictive than Haiku 4.5's but somewhat less restrictive than those applied to other recent models. In cybersecurity, they permit a wider range of defensive tasks than the safeguards for Sonnet 5.5 but still block penetration testing and other techniques more likely to be used by attackers. Its biology safeguards are the same as those for Sonnet 5, Sonnet 5.5, and Opus 5: they allow research biology questions but restrict access to requests judged likely to cause harm. Organizations working on wider-ranging biology and cyber activities can apply to the Life Sciences Verification Program and the Cyber Verification Program. The model's distinct approach is to combine a small footprint with a configurable effort dial and a clear role as a cost-efficient companion to larger Claude models. Anthropic explicitly frames Haiku 5.5 as best suited to more narrowly scoped tasks that might otherwise have been cost-prohibitive with previous versions of Claude, such as compaction, summarization, or subagent work, while Sonnet 5.5 and Opus 5.5 remain better choices for complex agentic coding tasks like those measured by Terminal-Bench 4.0. In practice this means a larger model can lead a task while Haiku 5.5 subagents handle the high-volume supporting steps. Anthropic also improved the value of its wider model range at the same time: Sonnet 5.5's cache reads were halved to $0.10 per million tokens, reducing the cost of Sonnet 5.5 on most agentic tasks by around 20%, and new monthly API credits were introduced for Max and Team subscribers. For users, the headline benefits are lower cost and lower latency at scale. The pricing table shows Haiku 5.5 at $0.01 per million tokens for cache reads on prompts up to 100,000 tokens and $0.05 over that, $0.125/$0.625 for cache writes, $0.10/$0.50 for input tokens, and $0.50/$2.50 for output tokens, against Haiku 4.5's $0.10, $1.25, $1.00, and $5.00. Early customer testing reported results consistent with the performance and cost improvements shown in the benchmarks. Asana measured over a 30% reduction in latency for task completions and up to 2.5x faster inference per agent turn. Box saw Haiku 5.5 score 11 points higher than Haiku 4.5 at about half the latency. The combination makes high-volume work affordable enough to run often rather than selectively. Anthropic and its early customers describe concrete workflows. AlphaSense's Ask in Document feature runs about 8 million calls a week in production answering very specific questions on top of one or a few documents; across 400 queries, Haiku 5.5 scored 0.84 versus 0.76 for Haiku 4.5. Box plans to use it on analytical work that runs at scale, from cost reports to financial summaries and weekly recurring reviews, across large volumes of enterprise content. HubSpot tests models on CRM tasks such as reporting on deals using simulated portals; Haiku 5.5 scored 92.8% averaged over three runs, the best result on that suite, and on a CRM audit task identifying stale but ambiguous records it was fastest to complete the task with the highest hit rate and the lowest false positive rate. Rogo uses it for quick lookups, subagents, and summaries, for example a Haiku 5.5 subagent pulling a segment revenue line from a 10-K while a bigger model builds the deck. Cognition offers Haiku 5.5 as a sidekick in Devin Fusion, holding a top-tier FrontierCode score of 66.2 while cutting cost and latency, available today in the Devin CLI with Opus 5.5 as the lead. Asana deployed it for AI Teammates use cases such as triaging bugs, setting up projects, and searching large portfolios to surface high-risk or overdue work. Claude Haiku 5.5 is available now on all platforms, including Amazon Web Services, Google Cloud, and Microsoft Azure, and developers on the Claude Platform can get started with the identifier claude-haiku-5-5; Anthropic provides a migration guide for details. For developers, Anthropic updated its Claude Python and TypeScript SDKs to add support for computer use and browser use in beta, noting that Haiku 5.5 is especially well-suited to these tasks given its combination of speed, capability, and price. Alongside the launch, Anthropic rolled out a new monthly API credit to Max and Team subscribers for use on the Claude Platform: Max 5x users receive $100 in credits per month, Max 20x users receive $200, and Team subscribers receive up to $500 pooled across their users. Credits can be used on any Claude model and are designed to let users experiment with building tools, apps, and agents that call the API. Claude Haiku 5.5's value proposition is straightforward: it is the cheapest, fastest, and most capable small model Anthropic has released, aimed squarely at high-volume, cost-sensitive work. With large benchmark gains over Haiku 4.5, an adjustable effort setting for tuning cost against intelligence, substantially lower token pricing, and availability across major cloud platforms, it gives teams a model they can run frequently for summarization, classification, lookups, subagents, customer support, and browser and computer use, while reserving larger Claude models for the most complex tasks.