Claude Haiku 5.5: Anthropic launches its fastest, cheapest small model for high-volume tasks

Anthropic has unveiled Claude Haiku 5.5, which it describes as the cheapest, fastest and most capable small model it has released to date. The company positions it explicitly for high-volume, cost-sensitive work: summarization, context compaction, database queries and classification tasks. It is also presented as the ideal complement to Opus 5.5 and Sonnet 5.5 when those act as "lead" models on coding tasks, leaving Haiku 5.5 to work as a subagent on narrowly scoped jobs. According to Anthropic, it is also its fastest model to date at standard speed (although still slower than its Opus models in fast mode), making it suitable for latency-sensitive cases such as live customer support or browser use.

On price, Anthropic says Haiku 5.5 costs on average around 75% less to run than Haiku 4.5. In concrete terms, the rate is 90% cheaper for requests of up to 100,000 tokens (which, the company says, accounted for close to 90% of requests to its previous Haiku model) and 50% cheaper for requests above that threshold. The company itself notes that this comparison takes into account that Haiku 5.5 uses an updated tokenizer (similar to that of Sonnet 5.5 and Opus 5.5), so it consumes somewhat more tokens per task than its predecessor. In the per-million-token pricing table, for prompts up to/above 100,000 tokens, Haiku 5.5 charges $0.01/$0.05 for cache reads, $0.125/$0.625 for cache writes, $0.10/$0.50 for input tokens and $0.50/$2.50 for output tokens; by comparison, Haiku 4.5 charged $0.10, $1.25, $1.00 and $5.00 respectively, and Sonnet 5.5 charges $0.10, $2.50, $2.00 and $10.00.

On performance, Anthropic compares Haiku 5.5 with Haiku 4.5, with "GPT-6 Luna" (as a competitive reference point) and with its own Sonnet 5.5 across several benchmarks. On GDPval-AA v2.1 (knowledge work across 44 professions, per Artificial Analysis) Haiku 5.5 scores 1620 points versus 735 for Haiku 4.5, 1437 for GPT-6 Luna and 1840 for Sonnet 5.5. On AA-Briefcase v1.1 it scores 1578 versus 614, 1336 and 1824 respectively. On computer use (OSWorld 2.1, offline subset) it reaches 72.4%, against 15.7% for Haiku 4.5, 48.9% for GPT-6 Luna and 83.9% for Sonnet 5.5. On Humanity's Last Exam, without tools it scores 45.9% (versus 10.2% and 56.9% for Haiku 4.5 and Sonnet 5.5; no figure is reported for GPT-6 Luna) and with tools 57.4% (versus 18.7% and 64.5%). In agentic coding, Terminal-Bench 4.0 gives 39.2% for Haiku 5.5, 0.0% for Haiku 4.5, 16.4% for GPT-6 Luna and 70.6% for Sonnet 5.5; on FrontierCode 1.1 (Main) it scores 46.4%, with no figure for Haiku 4.5, 42.4% for GPT-6 Luna and 52.1% for Sonnet 5.5 (in "Xhigh" mode). On visual reasoning (Chartography, without tools) it achieves 46.4% versus 6.4%, 29.1% and 61.6%.

One notable new addition is that Haiku 5.5 is the first model in the Haiku class with a configurable "effort" setting (Low, Med, High, Xhigh and Max levels), letting users decide whether to prioritize cost or intelligence, much as already works on other Anthropic models. The accuracy-versus-cost charts accompanying the announcement (on OSWorld, GDPval-AA and Humanity's Last Exam) show how performance varies with the chosen effort level.

Several companies offered early assessments. Asana (Aaron Vinh, Staff Software Engineer) notes that when testing Haiku 5.5 on its evaluation suite for AI Teammates — covering bug triage, project setup and searching across large portfolios to spot urgent or overdue work — they saw more than a 30% reduction in task-completion latency and up to 2.5 times faster inference per agent turn. HubSpot (Ze'ev Klapow, Distinguished Software Engineer) says Haiku 5.5 delivered the best score it has seen to date on its CRM task evaluation suite, 92.8% on average across three runs, as well as being the fastest model and the one with the best hit rate and lowest false-positive rate on an audit task to detect stale but ambiguous records. AlphaSense (Daniel Campos, Distinguished Engineer) says that in its "Ask in Document" feature — which handles some 8 million calls a week in production — a 400-query test showed a statistically significant improvement for Haiku 5.5 over Haiku 4.5 (0.84 versus 0.76). Box (Yashodha Bhavnani, VP of AI Products) reports a score 11 points higher than Haiku 4.5 with roughly half the latency, and expects to use it for analytical work at scale such as cost reports and financial summaries. Rogo (Alex Wang, Applied AI) describes it as useful for having a Haiku 5.5 subagent pull specific data (such as the revenue line by segment) from a 10-K filing while a larger model builds a presentation. Cognition (Walden Yan, co-founder and CPO) mentions that Haiku 5.5 joins as a "sidekick" option in its Devin Fusion product, achieving a top-tier score on FrontierCode (66.2) with lower cost and latency, and is already available in the Devin CLI alongside Opus 5.5 as the lead model.

On safety, Anthropic says it has observed significant improvements on nearly all of its alignment evaluations relative to Haiku 4.5, with fewer instances of misaligned behavior and less willingness to cooperate with misuse (details are referred to the model's system card). On safeguards, Haiku 5.5's cybersecurity protections are more restrictive than Haiku 4.5's, though somewhat less so than those applied to other recent company models: they allow a wider range of defensive tasks than Sonnet 5.5's safeguards, but still block penetration testing and other techniques more typical of attackers. The biology safeguards are the same as those for Sonnet 5, Sonnet 5.5 and Opus 5: they allow biological research questions but restrict requests deemed liable to cause harm. Organizations that need broader access can request it through Anthropic's life sciences and cybersecurity verification programs.

Claude Haiku 5.5 is now available across all platforms, including Amazon Web Services, Google Cloud and Microsoft Azure, and developers can start using it on the Claude Platform with the identifier claude-haiku-5-5. Alongside the launch, Anthropic is announcing other pricing and product changes: from now on, cache reads for Claude Sonnet 5.5 cost 50% less (from $0.20 to $0.10 per million tokens) which — since cache reads account for a significant share of token consumption — cuts the cost of Sonnet 5.5 by around 20% on most agentic tasks; the company itself illustrates the effect with the Terminal-Bench 4.0 benchmark. That said, Anthropic clarifies that Sonnet 5.5 and Opus 5.5 remain the best option for complex agentic coding tasks like those that benchmark measures, while Haiku 5.5 is intended for more narrowly scoped tasks that were previously prohibitively expensive, such as compaction, summarization or subagent work.

In addition, a new monthly API credit will roll out this week for all Claude Max and Team subscribers on the Claude Platform: Max 5x users will receive $100 in credits a month, Max 20x users $200, and Team subscribers up to $500, shared among their team's users. These credits, usable on any of Anthropic's models, are intended to make it easier to experiment with building tools, apps and agents that call its API. Finally, Anthropic is updating its Python and TypeScript SDKs to add beta support for computer use and browser use, features for which it considers Haiku 5.5 particularly well suited given its combination of speed, capability and price.