Anthropic just cut the price of its smallest Claude model by 90%. Claude Haiku 5.5 launched on October 7 as the cheapest and fastest model the company has ever shipped, and the benchmarks say it’s also a much bigger brain than the version it replaces. So what does half a dollar per million tokens actually buy you? Let’s look.
What Claude Haiku 5.5 actually is
Claude Haiku 5.5 is Anthropic’s smallest model. It sits at the bottom of a lineup that currently runs Haiku, Sonnet, Opus, and the research-tier Fable models, and it launched on October 7, 2026 on every major platform at once: Anthropic’s own API, Amazon Web Services, Google Cloud, and Microsoft Azure. Developers can call it with the model ID claude-haiku-5-5.
Three specs matter if you’re not a developer. The context window is 1 million tokens, the same as its bigger siblings, so it can read a small novel in one gulp. Max output is 128K tokens. And it’s the fastest Claude at standard speed, which matters more than you’d think when a job runs thousands of times a day.
There’s also a first for the Haiku family: an adjustable effort setting. The model decides on its own how much to think about each request. However, you can also force it to try harder on tricky items or keep it lazy and cheap on easy ones. More on that below.
The price cut, in real numbers
Here’s the pricing table straight from Anthropic’s launch page, because this is the part most people care about:
| Price per 1M tokens | Haiku 5.5 (prompts up to 100k) | Haiku 4.5 | Sonnet 5.5 |
|---|---|---|---|
| Input | $0.10 | $1.00 | $2.00 |
| Output | $0.50 | $5.00 | $10.00 |
| Cache reads | $0.01 | $0.10 | $0.10 |
| Cache writes | $0.125 | $1.25 | $2.50 |
For prompts longer than 100,000 tokens, input rises to $0.50 and output to $2.50. Even then it’s half the old price. Anthropic says about 90% of requests to the previous Haiku stayed under that line, so most people will live on the cheap side of the table.
For example, say you run 10,000 customer emails through the model, roughly 500 tokens each, and ask for a summary plus a category back. That’s about 5 million input tokens and 1 million output tokens. On Haiku 4.5 the bill lands near $10. On the new model, about $1. Same job, a tenth of the cost.
What the benchmarks say (in plain English)
Benchmarks are marketing until you translate them, so here’s the translation. GDPval-AA measures real professional work, the kind a freelancer or analyst would recognize. OSWorld checks whether an AI can actually operate a computer: clicking, typing, filling forms. Humanity’s Last Exam is a brutal test of expert-level knowledge. Terminal-Bench measures agentic coding, where the model completes real programming tasks in a terminal.
| Benchmark | Haiku 5.5 | Haiku 4.5 | GPT-6 Luna | Sonnet 5.5 |
|---|---|---|---|---|
| Knowledge work (GDPval-AA) | 1620 | 735 | 1437 | 1840 |
| Computer use (OSWorld 2.1) | 72.4% | 15.7% | 48.9% | 83.9% |
| Humanity’s Last Exam | 45.9% | 10.2% | not reported | 56.9% |
| Agentic coding (Terminal-Bench 4.0) | 39.2% | 0.0% | 16.4% | 70.6% |
Two numbers stand out. Claude Haiku 5.5 scored 1620 on the knowledge-work test, more than double the old Haiku and comfortably ahead of GPT-6 Luna’s 1437 from a competing lab. Meanwhile, on the computer-use test, it jumped from 15.7% to 72.4%. That’s not an upgrade. That’s a different species.
Even so, keep some perspective. On the coding test, Sonnet 5.5 still finishes 70.6% of tasks against Haiku’s 39.2%. Cheap and capable is not the same thing as frontier-grade, and pretending otherwise is how projects end up half-done.
Where it fits in real work
Anthropic positions this as the model for high-volume, repetitive jobs, and the customer examples on the launch page back that up. Box tested it for analytical work at scale, like cost reports and financial summaries, and saw scores 11 points higher than the old model at about half the latency. AlphaSense ran it against a feature that answers questions about documents, a job that does about 8 million calls a week, and it beat the previous model with statistical significance. HubSpot tried it on CRM cleanup tasks and got the best score they’d ever recorded on that suite: 92.8%, with the lowest false-positive rate of any model they tested.
For regular people, that translates into four realistic uses:
- Bulk summarizing: Condensing long reports, transcripts, or email threads.
- Sorting and labeling: Tagging support tickets, reviews, or survey answers.
- Data extraction: Pulling invoice numbers, dates, and totals out of messy text.
- The cheap subagent: Letting a bigger model delegate small lookups to it. Cognition’s Devin coding agent already works this way, with Haiku 5.5 fetching facts while Opus 5.5 leads the job.
Because it’s the fastest model in the family, live uses make sense too. Think support chat that answers in under a second, or a browser assistant that reads pages while you watch. Alongside the launch, Anthropic added beta support for computer use and browser use directly in its official SDKs, and Haiku 5.5 is the model they recommend for it.
The effort setting deserves one more look, because it changes how you pay. Set it low and the model stays fast and cheap on simple items. Crank it high and it reasons longer on the hard ones. As a result, one model can cover a spread of jobs that used to need two.
The catches worth knowing
No model launch is a free lunch. Before you move everything over, four things.
First, the tokenizer changed. The same piece of text counts as roughly 30% more tokens on Claude Haiku 5.5 than it did on Haiku 4.5. Anthropic still pegs the average savings at around 75%, but your bill won’t be exactly one tenth of the old one.
Second, that 100,000-token line matters more than it looks. Cross it and input pricing jumps fivefold to $0.50 per million. Most quick tasks never get close. Long document analysis, meanwhile, can blow past it without trying.
Third, a developer note: the model rejects requests that set temperature, top_p, or top_k, returning an error instead of a warning. If your code tuned those knobs on older models, remove them.
Fourth, and this is the big one for beginners: don’t use it for hard problems just because it’s cheap. It’s a specialist. Give it the boring 90% of work (the sorting, the summarizing, the extracting) and let the expensive models handle the rest.
The other things Anthropic announced
The Haiku launch came bundled with two changes that matter even if you never touch the small model.
Claude Sonnet 5.5 got cheaper to run. Cache reads now cost $0.10 per million tokens instead of $0.20, and since cached reads are a big share of agent workloads, Anthropic calculates the mid-tier model runs about 20% cheaper on typical agentic jobs.
Subscribers also get money to build with. Max 5x plans receive $100 in monthly API credits, Max 20x plans get $200, and Team plans pool up to $500. The credits work on any model. That quietly turns a Claude subscription into a small API budget, and we compared what each plan actually includes if you want the full breakdown.
How to try it today
Three doors, depending on your comfort level:
- The API: Sign up on the Claude Platform, then call the model ID claude-haiku-5-5. It’s pay-as-you-go, and the Batch API takes another 50% off if you can wait a few hours for results, which brings output down to $0.25 per million tokens. Our batch API guide shows how it works.
- A router like OpenRouter: One account, one bill, access to Claude and every other major model. We wrote a full beginner setup guide for OpenRouter.
- Your cloud provider: Already living in AWS, Google Cloud, or Azure? The model is available on all three, no new vendor needed.
Should you switch?
If your days involve any repeated text job, test it this week. Pick one task, run 100 items through Claude Haiku 5.5, and check two things: accuracy and cost. Then multiply by your real volume. For most people the result is a few dollars a month instead of a few dozen.
If your work is one hard problem at a time (complex code, deep analysis), nothing changes. Sonnet 5.5 and Opus 5.5 remain the right tools, and they just got cheaper to run too.
Here’s the bottom line. Anthropic didn’t make the small model smart enough to replace the big ones. It made it cheap enough to use everywhere else. That’s where most of the volume lives anyway, and now it costs a tenth of what it did last week.
