Claude Haiku 5.5 is out at $0.10 per million tokens - the effort setting decides what you actually pay

Anthropic released Claude Haiku 5.5 on 7 October, and the headline is price: $0.10 per million input tokens and $0.50 per million output tokens for prompts under 100,000 tokens. Haiku 4.5 charged $1 and $5. That is a 90% cut on the sticker, and it puts Haiku level with OpenAI's GPT-6 Luna at the base tier.

The sticker is not the bill. Anthropic's own estimate is that average workloads come out about 75% cheaper, not 90%, and independent testing suggests the real number depends heavily on how you configure it.

What actually changed

The pricing, according to VentureBeat and The Decoder's coverage of the launch:

• Under 100K tokens: $0.10 input, $0.50 output, $0.01 cache reads, $0.125 cache writes.
• Over 100K tokens: $0.50 input, $2.50 output, $0.05 cache reads, $0.625 cache writes.

Anthropic says about 90% of Haiku 4.5 requests fell under that 100K line.

Capability moved more than price did. On Anthropic's reported numbers, Haiku 5.5 scores 72.4% on the offline subset of OSWorld 2.1, a computer-use test, against 15.7% for Haiku 4.5 and 48.9% for GPT-6 Luna. On Terminal-Bench 4.0 it scores 39.2%, where Haiku 4.5 scored zero. The context window goes from 200,000 tokens to one million.

It is also the first Haiku with adjustable effort levels. Medium is the default. That detail matters more than any benchmark, for reasons below.

Two smaller changes shipped alongside it. Sonnet 5.5 cache reads were halved to $0.10 per million, which Anthropic estimates cuts most agent workloads by around 20%. And the Python and TypeScript SDKs gained beta support for computer and browser use.

Where the marketing and the measurements diverge

The benchmarks above are vendor-reported. The independent check so far comes from Artificial Analysis, as reported by The Decoder. It puts Haiku 5.5 at the top of the small-model class on its Intelligence Index, scoring 43 against 38 for GPT-6 Luna. Good news for Anthropic.

The same testing found something less flattering. At maximum effort, Haiku 5.5 used around 162,000 output tokens per task, roughly three times what Luna used. On cost per task, Luna reached a similar score for about a third of the price. At "high" effort, Haiku dropped to 38, matching Luna on both score and token use.

Effort also cuts the other way. VentureBeat notes the Terminal-Bench score is about 39% at maximum effort and about 20% at the default medium. So the impressive number and the cheap number are not the same configuration.

Add the new tokenizer, which Anthropic acknowledges uses somewhat more tokens for the same text, and the gap between "90% cheaper" and what you pay gets wider.

What it means if you run automation

Most business automation is not frontier reasoning. It is reading an inbound email and deciding where it goes. Pulling ten fields out of an invoice. Summarising a support ticket. Checking a record against a rule. These jobs run thousands of times a day, and the model cost is usually the line that decides whether the project pays for itself.

For that kind of work, Haiku 5.5 at low or medium effort is likely the most interesting change this month. Short prompts, short outputs, no long chains of thought. That is exactly the shape where the per-token price applies cleanly and the token-hungry max-effort behaviour never kicks in.

The bigger shift is computer use. A 72% OSWorld score from a model this cheap means screen-driving tasks in systems with no API, like legacy portals and old desktop tools, may now be affordable to run at volume. Treat that as something to test, not something to plan around. It is an offline subset of the benchmark, and real portals have timeouts, popups and logins that tests do not.

If you run long-document work, such as contracts, call transcripts or full case files, watch the 100K line. Above it, input costs five times more. A pipeline that stuffs whole files into the prompt because the window is now a million tokens can quietly land on the expensive tier for every call.

Who should wait

If your agent does complex multi-step work, Anthropic itself still points you to Sonnet and Opus. Haiku 5.5 at 39% on Terminal-Bench is a large jump, but Sonnet 5.5 scores 70.6% on the same test. Swapping a capable model for a cheap one in a workflow that needs judgment will save money on tokens and spend it on retries and human cleanup.

Security teams should note it ships with stricter cybersecurity restrictions than Haiku 4.5, and penetration testing is blocked under standard safeguards.

And if your current setup works and the model bill is not a meaningful cost, there is no urgency. Model swaps are never free. Prompts tuned for one model drift on another.

What to do this week

Pick your highest-volume, simplest task. Run a few hundred real inputs through Haiku 5.5 at low and medium effort, alongside whatever you use now. Record three numbers: accuracy against known-good outputs, tokens per run, and cost per run. Not cost per token. Cost per finished task.

If medium effort matches your current accuracy, switch that one task and keep the old model as a fallback. If it only matches at max effort, check the cost per task before you celebrate. That is the number that shows up on the invoice.