On Wednesday, October 7, 2026, Anthropic introduced Claude Haiku 5.5, which it calls “the cheapest, fastest, and most capable small model” it has ever released. It completes the Claude 5.5 family, which already included Opus 5.5 and Sonnet 5.5, and it targets a fast-growing niche: repetitive, high-volume work where every fraction of a cent and every millisecond counts.
According to the company, Haiku 5.5 is designed for quick, repetitive jobs such as summaries, context compaction, database queries and classification requests. It is also pitched as a “subagent” for the larger models: while Opus 5.5 or Sonnet 5.5 lead a complex coding task, Haiku can handle narrowly scoped pieces. Anthropic adds that it is its fastest model to date at standard speed, which makes it well suited to live customer support and browser use.
The core argument is price. For requests of up to 100,000 tokens, Haiku 5.5 costs $0.10 per million input tokens and $0.50 per million output tokens, compared with $1 and $5 for Haiku 4.5. Above that threshold, the rate rises to $0.50 and $2.50. Anthropic estimates that, on average, the new model costs around 75% less to run, taking into account that 90% of requests to Haiku 4.5 were short and that the new tokenizer uses slightly more tokens per task.
The technical documentation states that the model supports a one-million-token context window, up to 128,000 output tokens and adaptive thinking through an effort parameter. It is the first Haiku-class model to offer that setting, so users can decide whether to prioritize cost or intelligence on each call. The company warns that the same text translates into roughly 30% more tokens than with Haiku 4.5.
In the tests Anthropic published, the jump from the previous generation is large. On OSWorld 2.1, which measures an agent’s ability to operate a real computer on long tasks, Haiku 5.5 reaches 72.4% on the offline subset, compared with 15.7% for Haiku 4.5. On Terminal-Bench 4.0, which covers agentic coding at the command line, it scores 39.2%, where its predecessor scored 0%. On Humanity’s Last Exam it reaches 45.9% without tools and 57.4% with them.
The company also compares it with GPT-6 Luna, OpenAI’s lightweight model. On Artificial Analysis’s GDPval-AA v2.1 professional work evaluation, Haiku 5.5 scores 1,620 points versus 1,437; on FrontierCode 1.1 it reaches 46.4% versus 42.4%. Sonnet 5.5 remains ahead on almost every test, and Anthropic acknowledges that the larger models are still the better choice for the complex agentic coding measured by Terminal-Bench, where Sonnet scores 70.6%.
Several companies that tested the model before launch shared their own data. Asana reported a reduction of more than 30% in latency for its agent tasks; HubSpot said Haiku 5.5 achieved the best score it has seen on its customer management task suite, averaging 92.8% across three runs; AlphaSense, which handles about eight million calls a week in a document question feature, saw a statistically significant improvement, from 0.76 to 0.84. Box said the model scored 11 points higher than Haiku 4.5 at about half the latency.
The announcement came with changes across the rest of the lineup. Anthropic halved the price of Sonnet 5.5 cache reads, from $0.20 to $0.10 per million tokens. Because those reads make up a large share of consumption in agent workflows, the company estimates that Sonnet 5.5 becomes around 20% cheaper on most such tasks. It is a way to compete on price without cutting the model’s headline rate.
In addition, Anthropic announced a monthly credit for its developer platform aimed at Claude Max and Team subscribers. Max 5x users will receive $100 a month, Max 20x users $200, and Team plans up to $500 pooled across their members. The idea is for people who already pay for a subscription to experiment with building their own tools, apps and agents. The company also updated its Python and TypeScript development kits to include computer use and browser use in beta.
On safety, Anthropic says Haiku 5.5 improves on 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. Its cybersecurity safeguards are more restrictive than Haiku 4.5’s, though somewhat less restrictive than those of other recent models: they permit more defensive tasks than Sonnet 5.5’s, but still block penetration testing and other techniques typical of attackers. In biology it keeps the same protections as Sonnet 5 and Opus 5.
The model is already available in the Claude app for free and paid users, in Claude Code and, for developers, under the identifier claude-haiku-5-5, as well as on Amazon Web Services, Google Cloud and Microsoft Foundry. According to the model page, it will not be retired before October 7, 2027, and its knowledge cutoff is June 2026.
The launch matters because it shows where the business is heading. As companies build systems with dozens or hundreds of agents, most of the work does not require the most powerful model, but a reliable and cheap one that can run millions of times. A 90% price cut on short requests changes the math for many projects that used to be too expensive, from document classification to automated customer support.
It also fits into increasingly tough competition in the budget segment, where OpenAI offers GPT-6 Luna and Google its Gemini lineup. The performance figures come, for now, from Anthropic itself and selected customers, so they should be checked against independent evaluations and real-world use. What is clear is that the company wants the trade-off between quality and cost to stop being a dilemma for simple tasks: with Haiku 5.5, the bet is that the small model can do much more for much less.
