Anthropic cuts Haiku prices by 90% under 100,000 tokens — but the average saving is 75%
Anthropic launched Claude Haiku 5.5 on Wednesday 7 October 2026 — a small model for repetitive, high-volume tasks that the company itself says is on average around 75% cheaper to run than its predecessor Haiku 4.5.
Pricing varies with prompt length, and the most aggressive cuts sit below a threshold that covers nine out of ten requests. But the saving is not as large as the list prices suggest: a new tokenizer uses more tokens for the same job, and all the eye-catching benchmark figures were measured by the company itself.
Tiered pricing: a 90% cut where it matters most
The pricing structure splits usage in two. Below a prompt of 100,000 tokens, Haiku 5.5 costs $0.10 per million input tokens and $0.50 per million output tokens — a 90% cut compared with Haiku 4.5. Above the threshold, the price is $0.50 and $2.50, half the old rate.
According to Anthropic, nine out of ten Haiku 4.5 requests fell below the 100,000-token limit, so the cheapest tier is where actual usage lies. This is relayed through secondary sources — no primary announcement from Anthropic is present in the source material — so the figures should be read as the company's own information.
The tokenizer that eats the savings
The most important practical caveat makes the arithmetic less favourable than the list prices: Haiku 5.5 uses a new tokenizer that needs somewhat more tokens for the same job. That is why the blended saving comes to around 75%, not 90%.
For anyone planning high-volume workloads, this is decisive. The price per million tokens does not tell the whole story — the same document costs more in tokens than before, and that eats into the margin. Independent measurements of actual token consumption are not yet available.
The benchmarks: big jumps, but measured by the vendor
On Anthropic's own numbers, the leap from Haiku 4.5 is substantial: 72.4% on the computer-use benchmark OSWorld 2.1, versus 15.7% for the predecessor, and 39.2% on Terminal-Bench 4.0, where the older model scored zero. Sonnet 5.5 still leads with 70.6% on Terminal-Bench.
It should be emphasised clearly: these are vendor-run results, not independent ones. Since the model launched on 7 October and coverage appeared on 7–9 October, no independent benchmarks yet exist. HubSpot, as an early customer, reported that Haiku 5.5 scored 92.8% on the company's internal CRM test suite, measured as an average across three runs — also a customer claim, not an independent measure.
Where the model is intended to be used
Anthropic describes the model as follows, relayed by 9to5Mac (the quotes are our translations of Anthropic's English wording): "Claude Haiku 5.5 is designed for high-volume, cost-sensitive tasks," and "handles fast, repetitive work reliably (such as summarisation, compaction, database queries and classification requests)." The company adds that the model "works well alongside Opus 5.5 and Sonnet 5.5 as a subagent in coding work."
The launch is the third Claude 5.5 upgrade in roughly a month, following Opus and Sonnet, according to 9to5Mac.
Haiku 5.5 is also the first Haiku-class model with an adjustable "effort" setting, in which the user can choose between lower cost and higher intelligence, according to Cybersecuritynews. The model is available on Claude Platform, AWS, Google Cloud and Microsoft Azure.
The rest of the price pass-through
The launch forms part of a broader pricing push from Anthropic the same week:
- Sonnet 5.5 cache reading halved to $0.10 per million tokens, according to Cryptobriefing. Anthropic estimates that the change reduces costs by around 20% on most agentic workloads, where cached content makes up a large share of token consumption.
- Monthly API credits for Max and Team subscribers on Claude Platform, rolling out the same week: $100 per month for Max 5x, $200 for Max 20x, and up to $500 pooled across users for Team.
The sources disagree on one detail: Seeking Alpha describes cache-read prices for Haiku 5.5 of $0.01/$0.05, while other sources place the cut on Sonnet 5.5. The conflict is unresolved, and the first description appears unreliable.
Safety — according to the company's own description
Anthropic says Haiku 5.5 has stricter cybersecurity safeguards than Haiku 4.5, but looser than Sonnet 5.5: the model allows a wider range of defensive work, but still blocks penetration testing and techniques malicious actors would likely use. The company further states that it found "far fewer instances of misaligned behaviour, and a lower willingness to cooperate with misuse." Both claims are the company's own and not independently verified.
Headwinds and open questions
A single outlet, Geeky Gadgets (citing "Universe of AI"), claims that Haiku 5.5 in some tasks uses up to three times as many tokens as GPT-6 Luna — which could eat up the price advantage in token-intensive applications. The claim is uncorroborated by other sources and runs against the spirit of the savings picture; it should therefore not be relied upon, but it points to the same question the tokenizer caveat raises: exactly how much token consumption increases remains unclear until independent measurements exist.
The most important open questions after launch week: no independent benchmarks of the OSWorld 2.1 or Terminal-Bench results have yet appeared. There are no real production figures for token consumption. And how the effort setting affects both cost and quality in practice is not described in the source material. For high-volume users, the concrete advice for now is this: budget for around 75%, not 90%, and hold off locking in budgets until actual token usage has been measured.