Anthropic releases Claude Haiku 5.5 at 90% lower prices

Anthropic released Claude Haiku 5.5 on October 7, calling it “the cheapest, fastest, and most capable small model we’ve ever released.”

It’s built for high-volume, cost-sensitive work like summaries, database queries and classification. In coding, it works as a subagent alongside Opus 5.5 and Sonnet 5.5, and it shines in speed-sensitive tasks like live customer support and browser use.

The real headline is the price. For prompts up to 100,000 tokens, Haiku 5.5 costs $0.10 per million input tokens and $0.50 per million output tokens. That’s 90% less than Haiku 4.5 and exactly the same as OpenAI’s small model, GPT-6 Luna. Above 100,000 tokens it rises to $0.50 and $2.50, where Luna stays cheaper.

Anthropic says 90% of requests to Haiku 4.5 were under 100,000 tokens, putting the average saving at about 75%.

According to Anthropic’s table, Haiku 5.5 beats its predecessor Haiku 4.5 by a wide margin on every test. On the computer-use benchmark OSWorld 2.1 it jumps from 15.7% to 72.4%, and on the coding benchmark Terminal-Bench 4.0 from 0% to 39.2%.

It also beats OpenAI’s small model, GPT-6 Luna, on every test the two share. Its bigger sibling Sonnet 5.5 still leads on every row.

Haiku 5.5 is the first Haiku model with an adjustable effort setting. Users can pick from five levels, from low to max, to trade off cost against intelligence.

In Anthropic’s chart, its OSWorld score climbs from about 42% at the lowest level to 72.4% at the highest. GPT-6 Luna stays below Haiku 5.5 at similar costs throughout.

But the low price comes with a catch: Haiku 5.5 uses more tokens. Anthropic’s docs say its new tokenizer turns the same text into about 30% more tokens than Haiku 4.5.

Independent benchmarker Artificial Analysis found that at max effort it uses about 162,000 output tokens per task, roughly 3x GPT-6 Luna. At a lower setting, though, it matches Luna’s intelligence score with a similar number of tokens.

Anthropic draws the line itself: for complex coding work, Sonnet 5.5 and Opus 5.5 remain the better choice. Haiku 5.5 is meant for narrowly scoped tasks that used to be too costly, like summarization, context compaction and subagent work.

The same day, Sonnet 5.5’s cache read price was halved, from $0.20 to $0.10 per million tokens. Anthropic says that makes Sonnet about 20% cheaper on most agentic work.

Anthropic’s 144-page system card doesn’t hide the model’s weak spots. In the automated behavioral audit, Haiku 5.5 over-refused more than any other model tested. It used a leaked answer without telling the user 17% of the time, up from 2% for Haiku 4.5.

On the other hand, it’s the most robust Haiku model yet against prompt injection. And the card’s alignment section was reviewed by another Claude model, Mythos 5.1.

Anthropic is also giving subscribers monthly API credits: this week, Claude Max 5x subscribers get $100 a month, Max 20x gets $200, and Team accounts get up to $500. The credits work with every model. Developers also get computer use and browser use support in beta in the Python and TypeScript SDKs.