Anthropic has released Claude Haiku 5.5, its smallest current model, and the pitch is simple: much cheaper than the Haiku it replaces, and much more capable. Anthropic says it costs around 75% less to run than Haiku 4.5 on average, and on Anthropic’s own computer-use test its score rises from 15.7% to 72.4% [1].
The launch is confirmed on Anthropic’s own page, dated 7 October 2026 [1]. The benchmark scores are Anthropic’s figures from its own evaluations. Alongside it, Anthropic halved the price of cache reads on its mid-size Claude Sonnet 5.5 model [1].
What is Haiku 5.5 for?
Haiku is Anthropic’s small, fast tier. Anthropic describes Haiku 5.5 as built for “high-volume, cost-sensitive tasks”: summaries, compaction (shrinking long conversations so they fit), database queries and classification [1].
Two uses get special emphasis:
- Sub-agent work. Anthropic says Haiku 5.5 pairs well with its larger Opus 5.5 and Sonnet 5.5 models as a “subagent” on coding work [1]. In plain terms, a big model plans the job and hands small, well-defined errands to a cheaper, faster helper.
- Speed-sensitive jobs. Anthropic calls it its fastest model to date at standard speed, and points to live customer support and browser use [1]. A footnote adds that its Opus models running in “Fast Mode” are still quicker [1].
Anthropic is also clear about where it does not fit. It says Sonnet 5.5 and Opus 5.5 remain better choices for complex agentic coding, and that Haiku 5.5 is best for narrowly scoped tasks that were previously too expensive to run on Claude [1].
How much cheaper is it?
These are Anthropic’s published prices per million tokens [1]. A token is a small chunk of text, often part of a word.
| Haiku 5.5 (prompts up to 100k tokens) | Haiku 5.5 (over 100k) | Haiku 4.5 | |
|---|---|---|---|
| Input | $0.10 | $0.50 | $1.00 |
| Output | $0.50 | $2.50 | $5.00 |
| Cache writes | $0.125 | $0.625 | $1.25 |
| Cache reads | $0.01 | $0.05 | $0.10 |
So on short and medium prompts, Haiku 5.5’s list prices are a tenth of Haiku 4.5’s. On prompts over 100,000 tokens they are half [1].
Why, then, “around 75% less” rather than 90%? Anthropic explains in a footnote. About 90% of requests to Haiku 4.5 were under 100,000 tokens, so most traffic gets the bigger cut. But Haiku 5.5 uses an updated tokeniser that turns the same work into slightly more tokens. Netting those effects out, Anthropic puts the average saving at around 75% [1]. That is Anthropic’s own estimate of typical usage; your bill will depend on your prompt lengths.
How much better is it?
Anthropic published a comparison table against Haiku 4.5, OpenAI’s GPT-6 Luna and its own Sonnet 5.5. All figures below are Anthropic’s [1]:
- Computer use (OSWorld 2.1, offline subset): Haiku 5.5 72.4%; Haiku 4.5 15.7%; GPT-6 Luna 48.9%; Sonnet 5.5 83.9%. OSWorld measures how well an agent can operate a real computer to finish long, multi-step tasks.
- Agentic coding (Terminal-Bench 4.0): Haiku 5.5 39.2%; Haiku 4.5 0.0%; GPT-6 Luna 16.4%; Sonnet 5.5 70.6%.
- Expert reasoning (Humanity’s Last Exam, no tools): Haiku 5.5 45.9%; Haiku 4.5 10.2%; Sonnet 5.5 56.9%.
- Knowledge work (GDPval-AA v2.1, a rating-style score): Haiku 5.5 1620; Haiku 4.5 735; GPT-6 Luna 1437; Sonnet 5.5 1840.
Two things stand out. The jump over Haiku 4.5 is very large on every line. And Sonnet 5.5 still leads clearly on the hardest coding test, which matches Anthropic’s own advice to keep complex coding on the bigger models [1].
Haiku 5.5 is also the first Haiku with an adjustable “effort” setting, so developers can trade cost against intelligence per request, as they already can with Anthropic’s larger models [1].
Anthropic quotes several early customers, including Asana, HubSpot, AlphaSense and Box, reporting faster or more accurate results in their own tests [1]. Those are customer testimonials selected by Anthropic, not independent benchmarks.
What about safety?
Anthropic says Haiku 5.5 shows major improvements across almost all its alignment evaluations compared with Haiku 4.5, with far fewer instances of misaligned behaviour and less willingness to cooperate with misuse [1].
Its cybersecurity safeguards are described as stricter than Haiku 4.5’s but somewhat looser than those on Anthropic’s other recent models. Anthropic says they allow a wider range of defensive security work than Sonnet 5.5’s safeguards, while still blocking penetration testing and other techniques more likely to be used by attackers [1]. Biology safeguards match those on Sonnet 5, Sonnet 5.5 and Opus 5 [1].
What else changed?
- Sonnet 5.5 cache reads halved. From $0.20 to $0.10 per million tokens. Anthropic says that, because cache reads are a large share of token use in agent work, this makes Sonnet 5.5 around 20% cheaper on most agentic tasks [1].
- Monthly API credits for subscribers. Anthropic says Max 5x users will get $100 a month, Max 20x users $200, and Team subscribers up to $500 pooled, for use on its developer platform [1].
- SDK support for computer and browser use. Anthropic is adding beta support for computer use and browser use to its Python and TypeScript SDKs [1].
Haiku 5.5 is available now on Anthropic’s platform as claude-haiku-5-5, and on Amazon Web Services, Google Cloud and Microsoft Azure [1].
Why it matters
Agents are expensive mainly because they take many small steps, and every step is a bill. A small model that can reliably click through a browser, pull one figure from a filing, or summarise a long thread at a tenth of the old list price changes which of those steps are worth automating.
It also sharpens price competition at the small end. Anthropic chose to benchmark directly against OpenAI’s GPT-6 Luna [1], the model OpenAI began rolling out to Free and Go ChatGPT users on 8 October [2].
What this does not prove
- That Haiku 5.5 beats GPT-6 Luna in real use. The comparison is Anthropic’s, on tests Anthropic selected and ran [1].
- That you will save 75%. That is Anthropic’s average estimate, based on Haiku 4.5’s traffic and a tokeniser that uses slightly more tokens [1]. Long prompts get a smaller cut.
- That it can replace bigger models for hard coding. Anthropic itself says it cannot [1].
- That the customer quotes are representative. They are testimonials Anthropic chose to publish.
- Independent safety results. The alignment claims come from Anthropic’s own evaluations; we did not review the system card for this draft.
The Bottom Line
Claude Haiku 5.5 is a confirmed, sizeable step for Anthropic’s cheapest tier: list prices a tenth of Haiku 4.5’s on most prompts, an average saving Anthropic puts at around 75%, and a leap in its own computer-use score from 15.7% to 72.4% [1]. It is aimed squarely at the helper jobs inside agents, not at replacing Sonnet or Opus. The numbers are Anthropic’s; the real test is how it holds up in other people’s workloads.
Sources
- Anthropic, “Claude Haiku 5.5” (launch page with benchmarks, pricing, safety, availability and Sonnet 5.5 cache-read price change), 7 October 2026. https://www.anthropic.com/claude-haiku-5-5
- OpenAI Developer Community, “GPT-6 and Intelligent UI in ChatGPT” (Announcements; rollout table), 7 October 2026. https://community.openai.com/t/gpt-6-and-intelligent-ui-in-chatgpt/1404139
