Anthropic released Claude Haiku 5.5 on October 7, the first small model in the Claude 5 generation — the previous Haiku release stopped at 4.5. The new model went live the same day on the Claude Platform, and simultaneously rolled out on AWS, Google Cloud, and Microsoft Azure; Claude Code also picked it up in that day's release.
Anthropic is positioning it narrowly: high-throughput batch work, and as a sub-agent for larger models to call on. Pricing is the headline of this launch.
Two pricing tiers, split by prompt length
Haiku 5.5 splits pricing into two tiers based on the length of a single prompt, with the cutoff at 100,000 tokens:
| Per million tokens | Haiku 5.5 (≤100K) | Haiku 5.5 (>100K) | Haiku 4.5 |
|---|---|---|---|
| Input | $0.10 | $0.50 | $1.00 |
| Output | $0.50 | $2.50 | $5.00 |
| Cache read | $0.01 | $0.05 | $0.10 |
| Cache write | $0.125 | $0.625 | $1.25 |
In the short-prompt tier, every per-token price is one-tenth of Haiku 4.5's; in the long-prompt tier, it's half. Anthropic's own framing is that overall operating cost drops by about 75% on average.
The short tier's input and output pricing matches OpenAI's GPT-6 Luna, released in late September. The two companies' small models now sit on the exact same price line.
The same day, Anthropic also cut Sonnet 5.5's cache-read price from $0.20 to $0.10 per million tokens, and handed subscribers API credits: $100 a month for Max 5x, $200 a month for Max 20x, and up to $500 a month combined for a Team plan. Developer Simon Willison noted on his blog that unused credit does not roll over to the next month.
Benchmarks and reasoning tiers
Haiku 5.5 ships with reasoning built in — it can't be switched off, but can be dialed between low, medium, high, xhigh, and max, with medium as the default. In Anthropic's own results, it scored 72.4% on the OSWorld 2.1 offline subset for computer use, 39.2% on Terminal-Bench 4.0, and on Humanity's Last Exam, 45.9% without tools and 57.4% with tools.
A handful of early customers shared their own internal numbers. Box said the new model scored 11 points higher than Haiku 4.5, at roughly half the latency:
"Claude Haiku 5.5 scored 11 points higher than Haiku 4.5 at about half the latency."
Asana reported latency down more than 30%, and HubSpot said it posted the highest score the company has seen on that test suite, 92.8%. All of these figures are self-reported by the customers, and the underlying test questions haven't been made public.
The reasoning tier has a big effect on cost. Simon Willison ran the same "draw a pelican riding a bicycle" prompt at the lowest and highest settings: the low tier cost about $0.0009 and took 7 seconds, while the max tier cost about $0.034 and took 5 minutes 9 seconds — a more than 30-fold difference in cost for a single run.
Back-of-envelope: the tokenizer eats into some of the discount
There's a variable outside the price table that's easy to miss. Simon Willison pointed out that Haiku 5.5 switched tokenizers, so the same stretch of text now breaks into about 1.25 times as many tokens as it did under Haiku 4.5.
Running the numbers with that factor: the short-prompt tier's effective input cost works out to about $0.125 per million "old-style" tokens, versus Haiku 4.5's $1 — still a roughly 87% cut. The long-prompt tier looks less impressive: $0.50 times 1.25 comes to $0.625, which trims the discount down to about 37%.
Factor in reasoning, which is on by default, and output tokens will run somewhat higher than non-reasoning Haiku 4.5, by an amount that depends on the tier and the task. For teams running short-input work like benchmarking, extraction, or routing, switching models is close to a straightforward saving; for workloads that regularly feed in long documents or long context, it's worth running real traffic through the new pricing before switching over.
On safety, Anthropic says Haiku 5.5's restrictions on cybersecurity-related requests sit between Haiku 4.5 and Sonnet 5.5, while its safeguards for biological content match the Sonnet 5 series; related research requires a separate verification application.
Sources: Anthropic's Claude Haiku 5.5 product page, Simon Willison's blog, CocoLoop, SiliconANGLE; pricing figures for both tiers, cache pricing, and the Haiku 4.5 comparison were checked against Anthropic's official pricing page, and the Sonnet 5.5 cache-read price cut follows the official announcement.