Anthropic: Haiku 5.5 is live, and cheaper under 100,000 tokens
On 7 October 2026, Anthropic releases Claude Haiku 5.5, model id claude-haiku-5-5. Under 100,000 tokens, input is 0.10 dollars per million, output 0.50. Above that, those prices are five times higher. The model is not at your place.
Getting started
Original AI-generated illustration · SecuFocus
At a glance
Key points
Haiku 5.5 is the small model, for volume. The price written on the page changes at 100,000 tokens. A question sent to Claude leaves for Anthropic. We did not call the API.
What changes
The page is dated 7 October 2026. The model id is claude-haiku-5-5. Anthropic says it is available on the Claude Platform, Amazon Web Services, Google Cloud and Microsoft Azure.
The table, in dollars per million tokens, splits prompts up to 100,000 tokens from those above. Input: 0.10, then 0.50. Output: 0.50, then 2.50. Cache reads: 0.01, then 0.05. Cache writes: 0.125, then 0.625. On the same page, Haiku 4.5 is 1 dollar input and 5 output, with no such threshold.
Footnote 2 says the price is 90 percent lower than Haiku 4.5 under 100,000 tokens, and 50 percent lower above that. It adds that 90 percent of Haiku 4.5 requests fell under the threshold, and that the new tokenizer uses slightly more tokens. The line “about 75 percent less to run” is Anthropic’s average, not a SecuFocus measurement.
On Terminal-Bench 4.0, the page gives Haiku 5.5 39.2 percent and Sonnet 5.5 70.6 percent. Anthropic concludes that Sonnet and Opus remain the choice for complex agentic coding. We did not rerun the test.
What leaves
Nothing in the announcement describes a weight to download. The prompt, the document or the screenshot leaves for Anthropic, or for the cloud that hosts the call.
On cyber, the page says safeguards stricter than Haiku 4.5, somewhat less strict than recent models, and more open to defensive tasks than Sonnet 5.5. It also says they still block penetration testing. In biology, the safeguards are those of Sonnet 5, Sonnet 5.5 and Opus 5: research questions allowed, requests judged harmful restricted. A vendor safeguard is not a wall on your machine.
What we did not verify
We did not call claude-haiku-5-5, read the system card, or check the scores. The API credit announced for Max and Team is described as a rollout this week, not as a payment we have seen: 100 dollars a month for Max 5x, 200 for Max 20x, up to 500 for Team, pooled. Footnote 1 says Haiku 5.5 is the fastest at standard speed, and slower than Opus models in Fast Mode.
The decision
If you pay for short, frequent calls, the figure to reread is the 100,000-token threshold: 0.10 and 0.50 below it, five times that above it, for input. If the task is a long coding agent, the page itself points back to Sonnet 5.5. If you wanted a local model, this announcement does not give you one.
Check and explore
Sources for this article
Numbers connect each reference to the passages that use it. Dates show when the documentation was consulted.
- Claude Haiku 5.5, Anthropic ↗anthropic.com ·
This article draws on the sources above. The exercises are for you to try on your devices; SecuFocus does not present them as tests carried out by its editorial team. Interfaces and features can change. Method and corrections.
Cite this article
Keep this reference with the article when you save or share it.