Anthropic released Claude Haiku 5.5 on October 7, 2026, a small model priced at $0.10 per million input tokens, about 90% below Haiku 4.5 for prompts up to 100,000 tokens. It is built for high-volume background work such as summaries, classification and customer support, not for chatting with directly.
Most people will never pick it from a menu. They may notice it as faster replies inside tools built on Claude, and developers will notice it on the invoice.
TL;DR: Claude Haiku 5.5 is Anthropic’s cheapest and fastest small model, at $0.10 per million input tokens for prompts up to 100,000 tokens. It scores 72.4% on OSWorld 2.1 against 15.7% for Haiku 4.5, but still trails Sonnet 5.5 on hard coding work. Anthropic positions it as a subagent and high-volume workhorse, not a replacement for its larger models.
What Claude Haiku 5.5 costs compared with Haiku 4.5
Anthropic’s launch page lists separate prices for prompts up to 100,000 tokens and prompts above that size. Below the line, input and output both cost a tenth of what Haiku 4.5 charged.
| Price per million tokens | Haiku 5.5 (up to 100k / over 100k) | 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 |
Anthropic puts the average running cost at roughly 75% lower than Haiku 4.5. The model uses an updated tokenizer that spends slightly more tokens per task, so the real saving lands a little below the sticker math. Anthropic also halved Sonnet 5.5 cache reads, from $0.20 to $0.10 per million tokens, which it says makes Sonnet 5.5 roughly 20% cheaper on most agentic work.
That is the kind of cut that turns a feature a team ran once an hour into one it can run on every request.
How far Claude Haiku 5.5 moved on benchmarks
On Anthropic’s own numbers, the jump from Haiku 4.5 is large. OSWorld 2.1, a computer-use test, went from 15.7% to 72.4% on the offline subset, and Humanity’s Last Exam went from 10.2% to 45.9% without tools and from 18.7% to 57.4% with them.
Terminal-Bench 4.0 tells the harder story. Haiku 4.5 scored 0.0%, Haiku 5.5 scores 39.2%, and Sonnet 5.5 sits at 70.6%.
These are the company’s own results, so treat them as a starting point rather than a verdict. The same table lists GPT-6 Luna at 48.9% on OSWorld, 16.4% on Terminal-Bench and 42.4% on FrontierCode 1.1, where Haiku 5.5 scores 46.4%. On those three tests the budget model now lands ahead of the listed Luna scores, and that gap matters more than the price cut.
Effort settings and the subagent role
Haiku 5.5 is the first Haiku-class model with an adjustable effort setting, running from Low and Med through High, Xhigh and Max. A support bot can run it on Low for speed, while a harder query can be pushed up without switching models.
Anthropic describes it as a subagent beside Opus 5.5 and Sonnet 5.5, handling the cheap, repetitive steps. Asana reported over 30% lower latency on task completions, and Box measured 11 points higher than Haiku 4.5 at about half the latency in early testing.
Anthropic is clear about the limit: Sonnet 5.5 and Opus 5.5 remain better for complex agentic coding.
Safeguards and subscriber credits
Anthropic says Haiku 5.5 shows major alignment improvements over Haiku 4.5. Its cybersecurity safeguards are stricter than Haiku 4.5’s but less restrictive than Sonnet 5.5’s, so defensive work is allowed while penetration testing is still blocked.
Max and Team subscribers also get a monthly API credit, rolling out this week and usable on any model: $100 for Max 5x, $200 for Max 20x, and up to $500 pooled for Team.
What this means if you only use the Claude app
The launch page does not say whether the Claude apps or free users get Haiku 5.5, so the effect for most readers is indirect for now. It will show up in third-party products that run on Claude, where cheaper inference can mean faster answers or features that were too expensive to ship before.
For developers and teams, the practical step is to test it on the narrow, high-volume jobs where Haiku 4.5 was borderline, and keep the larger models for the hard ones.
