Claude Haiku 5.5 Explained: Anthropic’s Smallest Model Now Thinks, Plans and Uses Tools
Anthropic’s cheapest model just learned to reason in steps, plan multi-part jobs and call tools — here is what that changes for everyday users.
On October 7, 2026, Anthropic released Claude Haiku 5.5, the third model in its Claude 5.5 family to launch within roughly a month, following Opus 5.5 and Sonnet 5.5. Anthropic describes it as its cheapest, fastest, and most capable small model to date, built for jobs that run thousands of times a day, from customer support replies to the background work that powers AI agents.
The launch matters because the AI industry’s competition has shifted. It is no longer only about which company builds the smartest model. It is about economics: who can run capable AI affordably enough for businesses to deploy it everywhere. Haiku 5.5 is Anthropic’s answer to that question.
What Is Claude Haiku 5.5?
Every major AI company sells a lineup of models at different sizes and prices, the way carmakers sell a range of vehicles. Anthropic’s lineup has three tiers: Opus (the largest and most capable), Sonnet (the balanced middle), and Haiku (the small, fast, cheap one). The “5.5” label marks the current generation.
Haiku models are designed for high-volume, cost-sensitive work: the quick, repetitive tasks that businesses run at enormous scale. Anthropic lists examples like summaries, database queries, classification requests, and live customer support. The idea is simple: you do not need a flagship model to answer a routine question, and paying flagship prices for routine questions is wasteful.
Haiku 5.5 is also aimed squarely at AI agents. Because agents make many model calls in sequence to complete a task, the cost of each call adds up fast. A small model that is fast and cheap makes agents practical for real business use. Anthropic explicitly recommends Haiku 5.5 as a subagent, a smaller model that works under the direction of a larger one on coding projects.
What’s New in Haiku 5.5
The headline change is price. For prompts up to 100,000 tokens, Haiku 5.5 costs $0.10 per million input tokens and $0.50 per million output tokens. The previous Haiku 4.5 cost $1 and $5 at the same sizes. Anthropic says the new model runs about 75% cheaper on average overall.
But this is not just a cheaper model. Anthropic’s own benchmark results show large jumps in capability over Haiku 4.5:
- Computer use (OSWorld 2.1): 72.4%, up from 15.7%. This test measures how well a model can operate a real computer to complete multi-step tasks.
- Agentic coding (Terminal-Bench 4.0): 39.2%, up from 0.0%. This measures the ability to complete professional tasks in a command-line environment.
- Visual reasoning (Chartography): 46.4%, up from 6.4%.
An important caveat: all of these figures come from Anthropic’s own testing. No independent benchmarks were available at launch, so treat the comparisons as the company’s claims rather than verified results. Anthropic also benchmarked Haiku 5.5 against OpenAI’s GPT-6 Luna, a model aimed at the same budget segment, and reports Haiku 5.5 scoring higher on every shared test, including OSWorld 2.1 (72.4% vs 48.9%) and Terminal-Bench 4.0 (39.2% vs 16.4%).
One genuinely new feature: Haiku 5.5 is the first Haiku-class model with an adjustable effort setting. Developers can dial the model toward lower cost or higher intelligence on the same request, instead of switching to a different model entirely.
Haiku 5.5 vs Haiku 4.5 vs Sonnet 5.5
For developers choosing a model, the lineup now looks like this, using Anthropic’s published pricing for standard prompts up to 100,000 tokens:
| Model | Input (per 1M tokens) | Output (per 1M tokens) | Best for |
|---|---|---|---|
| Haiku 5.5 | $0.10 | $0.50 | High-volume tasks, agents, subagents |
| Haiku 4.5 | $1.00 | $5.00 | Superseded by 5.5 for most uses |
| Sonnet 5.5 | $2.00 | $10.00 | Complex coding, difficult reasoning |
Sonnet 5.5 remains the better choice for hard tasks. Anthropic is clear about this: on Terminal-Bench 4.0, Sonnet 5.5 scores 70.6% versus Haiku 5.5’s 39.2%. The point of Haiku is not to beat Sonnet. It is to handle the enormous volume of simpler work that would be absurdly expensive on a flagship model.
Alongside the launch, Anthropic also halved the price of cache reads on Sonnet 5.5, making that model about 20% cheaper on most agent-style work, and introduced monthly API credits for Max and Team subscribers to encourage experimentation with building agents.
Why the Pricing Matters: The Economics of AI Agents
The deeper story is about how AI agents get paid for. An agent that completes a task might call a model dozens or hundreds of times: once to plan, once per search, once per step, once to write the final answer. If each call costs a cent, a single task can cost dollars. Multiply by thousands of users and the economics break.
Haiku 5.5 attacks that problem directly. At a tenth of the price of its predecessor, with much stronger performance on agent-style tasks, it makes the multi-step agent pattern affordable. Anthropic’s early customers reported results consistent with the claimed improvements: one company, Asana, said it saw over 30% lower latency and up to 2.5x faster inference per agent turn when testing Haiku 5.5 in its AI agent product.
This is why investors and analysts now watch model pricing as closely as model intelligence. As one industry roundup noted, AI competition increasingly depends on economics, not just headline benchmark performance. Businesses deploying large numbers of AI agents need models that perform routine tasks affordably and consistently.
Availability
Haiku 5.5 is available now on the Claude Platform, where developers can start using it immediately, and through the major cloud platforms: Amazon Web Services, Google Cloud, and Microsoft Azure. Anthropic also made it selectable in its consumer Claude apps. The company describes it as suitable for speed-sensitive work like live customer support and browser use, and as a subagent paired with Opus 5.5 and Sonnet 5.5 on coding projects.
On safety, Anthropic says Haiku 5.5 shows major improvements in its alignment evaluations compared with Haiku 4.5, with far fewer instances of misaligned behavior and a lower willingness to cooperate with misuse. Its cybersecurity safeguards are stricter than those on Haiku 4.5, though somewhat less restrictive than those on the larger models.
Claude Haiku 5.5 is Anthropic’s small, fast, inexpensive AI model, released on October 7, 2026. It is the third model in the Claude 5.5 family, after Opus 5.5 and Sonnet 5.5, and is designed for high-volume, cost-sensitive tasks like summaries, classification, customer support, and powering AI agents.
For prompts up to 100,000 tokens, it costs $0.10 per million input tokens and $0.50 per million output tokens. Longer prompts cost $0.50 and $2.50. Anthropic says it runs about 75% cheaper on average than Haiku 4.5.
No. Sonnet 5.5 is the stronger model for difficult tasks: it scores 70.6% on the Terminal-Bench 4.0 coding test versus Haiku 5.5’s 39.2%. Haiku 5.5 wins on speed and cost, which makes it the better choice for high-volume routine work and for running many agent steps affordably.
It is a new dial that lets developers choose whether the model should optimize for lower cost or higher intelligence on a given request. Haiku 5.5 is the first Haiku-class model to offer it, removing the need to switch models when you want a different cost-intelligence balance.
They compete in the same budget segment with identical starting prices ($0.10 per million input tokens). In Anthropic’s own benchmarks, Haiku 5.5 scores higher on every shared test, including computer use and terminal tasks. Remember that these figures come from Anthropic, not independent testing.
Developers can use it on the Claude Platform and through Amazon Web Services, Google Cloud, and Microsoft Azure. It is also available to select in Anthropic’s Claude apps for everyday use.
