The model people talk about least, and that matters most for volume
On October 7 Anthropic released Claude Haiku 5.5. Between Opus 5.5, Sonnet 5.5 and Fable 5.2 it went almost unnoticed, which is a pity, because for many companies it is the model that moves the bill most.
The reason is simple. Most of the work a company wants to automate doesn't need deep reasoning. Classifying incoming email, extracting fields from invoices and orders, routing tickets, summarizing documents, checking that a form is complete. These are small tasks repeated thousands of times a day, and there the price per token is everything.
Haiku 5.5 has the identifier claude-haiku-5-5, is available on the Claude API, Amazon Bedrock, Google Cloud and Microsoft Foundry, and Anthropic commits not to retire it before October 7, 2027. Haiku 4.5 remains available, with no retirement date announced.
What changes from Haiku 4.5
Three things, all of them relevant.
Context goes from 200,000 tokens to one million, like the other current models, and maximum output from 64,000 to 128,000 tokens. For a small model, that means working on long documents without splitting them.
It is the first Haiku with adjustable effort, from low to max, with medium as the API default. You can decide how much reasoning to spend task by task: low for routing, higher when extraction is delicate.
And performance makes a jump rarely seen between two generations. On GDPval-AA, which measures office work, Haiku 5.5 reaches 1620 Elo against 735 for Haiku 4.5. On OSWorld, computer use, 72.4% against 15.7%. On Terminal-Bench 4.0, terminal coding, 39.2% against zero.
One technical note: it uses the tokenizer introduced with Opus 4.7, which counts slightly more tokens for the same text. Real savings are therefore a little below what the price list says.
The price, and the threshold to know
Haiku 5.5's price list has a quirk: it depends on request length.
Up to 100,000 input tokens it costs $0.10 per million input tokens and $0.50 output. Above that threshold it rises to $0.50 and $2.50. Cache reads start at $0.01. Haiku 4.5 cost $1 and $5.
In other words: compared with Haiku 4.5, savings are 90% on short requests and 50% on long ones. Anthropic estimates an average cost 75% lower.
Keep this threshold in mind when designing. A system that classifies email or extracts data from an invoice is almost always under 100,000 tokens and pays the lowest price. A system that sends Haiku entire archives in a single request pays five times as much. It often pays to split the work into smaller requests.
For comparison: Sonnet 5.5 costs $2 and $10, twenty times Haiku 5.5 on short requests. On the same day Anthropic halved Sonnet 5.5's cache read price, from $0.20 to $0.10, which the company says means about 20% less on most agentic work.
Get updates on Claude and AI for business
One email when there's something worth reading. No spam.
Evaluating Claude for your organization? Learn what it costs or which plan fits best
Where it stops
Haiku 5.5 is much better than its predecessor, but it isn't Sonnet.
On GDPval-AA Sonnet 5.5 reaches 1840 against 1620. On OSWorld 83.9% against 72.4%. On Terminal-Bench 4.0 the gap is wide: 70.6% against 39.2%. Anthropic itself says Sonnet 5.5 and Opus 5.5 remain better for complex agentic coding.
The interesting comparison is with the direct competitor. On almost every published test Haiku 5.5 beats GPT-6 Luna, OpenAI's small model: 1620 against 1437 on GDPval-AA, 72.4% against 48.9% on OSWorld.
On safety filters Haiku 5.5 is stricter than Haiku 4.5 on cybersecurity and aligned with Sonnet on biology. For a company using it on office documents and processes, the difference won't be noticeable.
Where it pays off
Anthropic's documentation describes it in one line: for high-volume, latency-sensitive tasks such as classification, extraction and routing. That's the right summary.
There are four cases where I'd use it right away. Routing incoming email and tickets, where speed matters and the task is well defined. Extracting structured data from invoices, orders and shipping documents. Summaries and context compaction for larger agents. And the subagent role: a large model decides what to do, Haiku runs the simple steps in parallel.
The customers Anthropic cites point the same way. HubSpot measured the best result it has seen from a small model in its CRM tests. AlphaSense uses it for about 8 million calls a week. Cognition pairs it with Opus 5.5 as a sidekick in its coding agent.
The principle is the usual one: mix models instead of using one for everything. With Haiku 5.5 the bottom of the ladder has become capable enough to take on a much bigger share of the work.
News for subscribers
Alongside Haiku 5.5 Anthropic introduced monthly API credits, usable on any model, for subscribers: $100 a month on Max, $200 on Ultra, up to $500 pooled on Team plans.
For a small company on a Team plan that's a concrete opportunity. With $500 a month of Haiku 5.5 on short requests you can process billions of input tokens, more than enough for a first automation project on email or documents. You can build a real prototype without opening a separate budget.
As always, going from prototype to production takes more work: error handling, checks, integration with your systems. For how to set it up there is the Claude API integration guide, and for the cost picture how much Claude costs for businesses.