On 7 October 2026 Anthropic introduced Claude Haiku 5.5, which the company calls its cheapest, fastest and most capable small model. For developers the cost side matters most: on average, running it is meant to cost around 75% less than Haiku 4.5. How that number comes about, and where it does not apply, is explained below.
What happened
Haiku 5.5 is a model for large volumes of requests where cost counts. Anthropic names quick, repetitive work such as summaries, compacting conversation history ("compaction"), database queries and classification. In coding it is meant to work as a subagent alongside Opus 5.5 and Sonnet 5.5. Because it is also the company's fastest model, it suits tasks where speed matters, for example live customer support or operating a browser. A footnote adds one limit: this applies at standard speed; in "Fast Mode" the Opus models run faster.
The model is available now, according to Anthropic on all platforms, including Amazon Web Services, Google Cloud and Microsoft Azure. On the Claude Platform the model ID is claude-haiku-5-5. For the first time a Haiku model has an adjustable effort setting: users decide whether to optimise for cost or for performance. In Anthropic's charts the levels run from Low to Max.
The costs in detail
The table below reproduces Anthropic's figures per million tokens. For Haiku 5.5 the left value applies to prompts up to 100,000 tokens, the right one to longer prompts.
| Per million tokens | Haiku 5.5 (up to / over 100,000 tokens) | Haiku 4.5 | Sonnet 5.5 |
|---|---|---|---|
| Input | $0.10 / $0.50 | $1.00 | $2.00 |
| Output | $0.50 / $2.50 | $5.00 | $10.00 |
| Cache reads | $0.01 / $0.05 | $0.10 | $0.10 |
| Cache writes | $0.125 / $0.625 | $1.25 | $2.50 |
The figure "around 75%" is an average. Anthropic explains in a footnote how it comes about: for requests up to 100,000 tokens Haiku 5.5 is 90% cheaper than Haiku 4.5, for longer requests 50%. Since, according to Anthropic, 90% of requests to Haiku 4.5 fell into the short class, the result is roughly a three-quarter drop on average. The calculation also takes into account that Haiku 5.5 uses an updated tokenizer, similar to that of Sonnet 5.5 and Opus 5.5. The model therefore needs slightly more tokens for the same work. Anthropic gives no number for this; third-party sites speak of about 30% more tokens, but that figure is not substantiated. Anyone who wants to know what changes in their own operation should recalculate the cost with their own traffic.
Performance according to Anthropic
Anthropic compares Haiku 5.5 with Haiku 4.5, with OpenAI's GPT-6 Luna and, for reference, with Sonnet 5.5. All values below come from Anthropic itself and have not been checked independently.
| Test | Haiku 5.5 | Haiku 4.5 | GPT-6 Luna | Sonnet 5.5 |
|---|---|---|---|---|
| OSWorld 2.1 (computer use, offline subset) | 72.4% | 15.7% | 48.9% | 83.9% |
| Terminal-Bench 4.0 (agentic coding) | 39.2% | 0.0% | 16.4% | 70.6% |
| Humanity’s Last Exam, no tools | 45.9% | 10.2% | – | 56.9% |
| Humanity’s Last Exam, with tools | 57.4% | 18.7% | – | 64.5% |
In knowledge work (GDPval-AA v2.1, Elo score) Haiku 5.5 reaches 1620, Haiku 4.5 735, GPT-6 Luna 1437 and Sonnet 5.5 1840. The gap to Sonnet 5.5 is therefore visible in every test, and Anthropic says so itself: Sonnet 5.5 and Opus 5.5 remain the better choice for complex agentic coding tasks such as those in Terminal-Bench 4.0. Haiku 5.5 is best suited to narrower tasks that would have been too expensive with earlier models, such as compaction, summarisation or subagent work.
Anthropic also quotes customers. Asana reports over 30% lower latency for completing tasks and up to 2.5 times faster inference per agent turn. HubSpot names the best score it has seen so far among smaller models in its simulated CRM tasks: 92.8% averaged over three runs. AlphaSense reports 0.84 against 0.76 for Haiku 4.5 on 400 queries for "Ask in Document"; that service generates about 8 million calls a week there. These are vendor statements; the customers' measurement methods are not public.
Background
Haiku is Anthropic's line of small models, meant for large volumes and cost-conscious work. Haiku 4.5 was the direct predecessor. The 5.5 family has meanwhile grown: Sonnet 5.5 and Opus 5.5 already exist; reports from third-party sites date Sonnet 5.5 to 28 September 2026.
The comparison with OpenAI is part of the story. Anthropic places Haiku 5.5 next to GPT-6 Luna, a model of OpenAI's GPT-6 line, in its performance table. VentureBeat's headline says Anthropic launches Haiku 5.5 with a 90% API cost reduction, matching GPT-6 Luna. What GPT-6 Luna costs in detail is not stated in the sources evaluated.
Two further changes come with the launch that have nothing to do with Haiku 5.5 itself but belong to the overall picture. First, Anthropic halves the cost of cache reads on Sonnet 5.5: from $0.20 to $0.10 per million tokens. Because cache reads make up a large share of token consumption in agentic work, Sonnet 5.5 becomes around 20% cheaper there, according to Anthropic. Second, Anthropic introduces a monthly API credit for Max and Team subscribers: $100 for Max 5x, $200 for Max 20x and up to $500 for Team, pooled across the team's users. It is to roll out this week and can be used on all models. For developers, the Python and TypeScript SDKs gain beta support for computer use and browser use.
- OpenAI Decisions API: you only pay for the input
- Google Cloud offers Gemini, Claude and OpenAI models
- Claude Duet: Anthropic Launches a Proactive AI Agent
- Claude rewrites the TypeScript compiler in Rust
What this means
For developers, the main change is the calculation for bulk tasks. Anyone processing many short requests, such as classification, summaries or lookups, usually stays below the 100,000-token limit and benefits from the largest reduction. Anyone who often sends very long prompts benefits less, since only 50% falls away there. The effort setting is also worth a look: it determines how much computing effort the model spends per task and, according to Anthropic's charts, affects both performance and cost.
For small businesses that have AI built into their own workflows, a cheaper small model can make applications with many calls more affordable. The AlphaSense example, where the model answers questions about individual documents, shows what kind of work is meant. Whether it pays off in a given case depends on volume and task. For demanding work, Anthropic explicitly continues to recommend the larger models.
For everyone else the announcement mostly concerns technology in the background. The sources do not say what end users will notice; the announcement is aimed at developers and at subscribers to Max and Team.
A note on interpretation: all performance and cost figures come from Anthropic or from reports repeating them. The sources evaluated contain no independent tests of Haiku 5.5. Before doing your own calculation, check the provider's official rate list.
Frequently asked questions
How much cheaper is Claude Haiku 5.5 really: 75 or 90 percent?
Both are correct; they are two different measures. Anthropic puts the average at around 75% lower running cost. For requests up to 100,000 tokens the drop is 90%, for longer requests it is 50%. On Haiku 4.5, 90% of requests fell into the short class.
What does Claude Haiku 5.5 cost per million tokens?
For prompts up to 100,000 tokens, $0.10 for input and $0.50 for output; above that, $0.50 and $2.50. For comparison, Haiku 4.5 cost $1.00 and $5.00.
Does Haiku 5.5 replace the larger models for coding?
According to Anthropic, no. For complex agentic coding tasks, Sonnet 5.5 and Opus 5.5 remain the better choice. Haiku 5.5 suits narrower tasks such as summaries, compacting context or subagent work.
Where is Claude Haiku 5.5 available?
According to Anthropic, on all platforms, including Amazon Web Services, Google Cloud and Microsoft Azure. On the Claude Platform the model ID is claude-haiku-5-5.
What does the new effort setting mean?
Haiku 5.5 is the first Haiku model with an adjustable effort setting. It lets users trade lower cost against higher performance; in Anthropic's charts the levels run from Low to Max.
