Anthropic launches Claude Haiku 5.5: its fastest, cheapest and most capable small model to date

Anthropic has unveiled Claude Haiku 5.5, the new version of its smallest and cheapest model, described by the company as "the cheapest, fastest and most capable small model" it has released to date. It is designed for high-volume, cost-sensitive tasks: summarization, context compaction, database queries and classification. According to Anthropic, it also works particularly well as a subagent alongside larger models such as Opus 5.5 and Sonnet 5.5 on coding tasks, and as its fastest model yet, it is well suited to latency-sensitive applications such as live customer support or browser use.

The most striking change is the price: Haiku 5.5 costs on average around 75% less to run than Haiku 4.5. In the fine print, Anthropic clarifies that the reduction is 90% for requests of up to 100,000 tokens (which accounted for 90% of requests to Haiku 4.5) and 50% for requests above that threshold, a calculation that also factors in that Haiku 5.5 uses an updated tokenizer (similar to Sonnet 5.5 and Opus 5.5) that consumes slightly more tokens per task. In concrete per-million-token figures: cache reads now cost $0.01/$0.05 (depending on the tier), cache writes $0.125/$0.625, input tokens $0.10/$0.50 and output tokens $0.50/$2.50—compared with $0.10, $1.25, $1.00 and $5.00 respectively for Haiku 4.5.

On performance, Anthropic publishes comparisons against Haiku 4.5, Sonnet 5.5 and a competing model called GPT-6 Luna across several benchmarks: on GDPval-AA v2.1 (knowledge work across 44 professions), Haiku 5.5 scores 1620 versus 735 for Haiku 4.5 and 1437 for GPT-6 Luna (Sonnet 5.5 is listed at 1840 for reference); on AA-Briefcase v1.1, 1578 versus 614; on computer use (OSWorld 2.1, offline subset), 72.4% versus 15.7% for Haiku 4.5 and 48.9% for GPT-6 Luna; on Humanity's Last Exam, 45.9% without tools and 57.4% with tools, versus 10.2%/18.7% for Haiku 4.5; on Terminal-Bench 4.0 (agentic coding), 39.2% versus 0.0% for Haiku 4.5 and 16.4% for GPT-6 Luna; on FrontierCode 1.1, 46.4%; and on visual reasoning (Chartography), 46.4% without tools versus 6.4% for Haiku 4.5.

One structural novelty is that Haiku 5.5 is the first model in the Haiku line to include a configurable "effort" setting, the same mechanism already present in other Anthropic models, which lets users decide whether to optimize for cost or for intelligence, with levels ranging from "Low" to "Max." Anthropic's own charts show how performance on OSWorld, GDPval-AA and Humanity's Last Exam varies by the chosen effort level, plotted against cost per task.

The company also includes testimonials from customers who tested the model early. Asana says it saw more than a 30% reduction in task completion latency and up to 2.5 times faster inference per agent turn in its "AI Teammates" product, in use cases such as bug triage or finding stalled work across large portfolios. HubSpot notes that Haiku 5.5 achieved the best score it has seen so far on its test suite for CRM tasks, averaging 92.8% across three runs, while also being the fastest and the best at combining hit rate with a lower false-positive rate when identifying stale records. AlphaSense, which processes some 8 million calls a week in production with its "Ask in Document" feature, reports a statistically significant improvement over Haiku 4.5 across 400 test queries (0.84 versus 0.76). Box says it scored 11 points higher than Haiku 4.5 at roughly half the latency, and is considering using it for expense reports, financial summaries and recurring weekly reviews at scale. Rogo describes it as ideal for bounded, high-volume work, such as having a Haiku 5.5 subagent pull the revenue-by-segment line from a 10-K filing while a larger model builds a presentation. Cognition, the maker of Devin, says Haiku 5.5 joins as a "sidekick" option in Devin Fusion, achieving a top-tier score of 66.2 on FrontierCode while cutting cost and latency, paired with Opus 5.5 as the lead model.

The article itself underscores an important caveat: Sonnet 5.5 and Opus 5.5 remain the better choice for complex agentic coding tasks, such as those measured by Terminal-Bench 4.0; Haiku 5.5 is meant for more bounded tasks that were previously prohibitive on cost, such as compaction, summarization or subagent work.

On safety, Anthropic reports substantial improvements across nearly all of its alignment evaluations compared with Haiku 4.5, with far fewer instances of misaligned behavior and less willingness to cooperate with misuse (details are left to the model's system card). Haiku 5.5's cybersecurity safeguards are more restrictive than Haiku 4.5's, though somewhat less so than those of other recent models: they allow a broader range of defensive tasks than Sonnet 5.5's, but still block penetration testing and attacker-associated techniques. The biology safeguards are the same as those applied to Sonnet 5, Sonnet 5.5 and Opus 5: they permit biological research questions but restrict requests deemed potentially harmful. Organizations with broader needs in these areas can apply for access to Anthropic's life sciences and cybersecurity verification programs.

Haiku 5.5 is available now on all platforms, including Amazon Web Services, Google Cloud and Microsoft Azure, as well as directly on the Claude Platform, under the model identifier claude-haiku-5-5.

Alongside the launch, Anthropic is announcing other pricing and product changes. First, it is cutting the price of Claude Sonnet 5.5 cache reads by 50%, from $0.20 to $0.10 per million tokens; because cache reads account for a significant share of token consumption in agentic tasks, this makes running Sonnet 5.5 roughly 20% cheaper in most such scenarios, with a concrete example shown on Terminal-Bench 4.0. Second, this week all Claude Max and Team subscribers will receive a new monthly API credit to use on the Claude Platform: $100 a month for Max 5x users, $200 for Max 20x, and up to $500 (pooled across users) for Team subscribers, intended to let them experiment with building their own tools, applications and agents using any of Anthropic's models. Finally, Claude's Python and TypeScript SDKs are being updated to add beta support for computer use and browser use, features for which Haiku 5.5 is particularly well suited given its combination of speed, capability and price.