Claude Haiku 5.5 costs a tenth as much up to 100,000 tokens. Keep your routine prompts under the line

On October 7, Anthropic released Claude Haiku 5.5, its small, fast model, at a price that changes the math on routine AI work. The headline is the price cut. The detail that matters more for a lean company is a single line in the price table: 100,000 tokens. Up to that line, the model is very cheap. Above it, the same work costs five times as much. If you are pricing AI work against people or vendors, start with what AI help costs, priced three ways, then come back to this line.

What changed

All prices here are company-reported, per million tokens, from Anthropic’s launch page.

• Haiku 5.5, prompts up to 100,000 tokens: $0.10 input, $0.50 output.
• Haiku 5.5, prompts over 100,000 tokens: $0.50 input, $2.50 output.
• Haiku 4.5, for comparison: $1.00 input, $5.00 output.
• Sonnet 5.5: $2.00 input, $10.00 output. Its cache reads drop from $0.20 to $0.10, which Anthropic says makes most agentic work about 20% cheaper.

Anthropic says Haiku 5.5 is 90% cheaper than Haiku 4.5 for requests up to 100,000 tokens, 50% cheaper over that line, and about 75% cheaper to run on average. It also says about 90% of requests to the previous Haiku were for prompts up to 100,000 tokens.

One caveat comes from the same page. Haiku 5.5 has an updated tokenizer that “uses slightly more tokens per task.” The same document will count as a few more tokens than it did before, so measure on the new model rather than reusing old token counts.

Anthropic names the work it has in mind: summaries, classification, database queries, live customer support, browser use and sub-agent work. It also says Max and Team subscribers will get a monthly API credit, rolling out this week: $100 for Max 5x, $200 for Max 20x, and up to $500 pooled for Team. Until it shows up in your account, treat that as an announcement.

A worked example

Here is the arithmetic on a made-up process, using list prices only. It is a calculation, not a measurement.

A business sorts and drafts first replies to 3,000 support emails a month. Each request sends the model about 8,000 tokens (instructions, a short policy sheet and the email) and gets back about 500.

That is 24 million input tokens and 1.5 million output tokens a month.

Model	Input cost	Output cost	Monthly total
Haiku 5.5, under the line	$2.40	$0.75	$3.15
Haiku 4.5	$24.00	$7.50	$31.50
Sonnet 5.5	$48.00	$15.00	$63.00

Now suppose someone “improves” the process by pasting the whole help center into every request, and each prompt grows to 120,000 tokens. Input jumps to 360 million tokens and crosses the line.

Model	Input cost	Output cost	Monthly total
Haiku 5.5, over the line	$180.00	$3.75	$183.75

The same model, the same emails, the same replies: about 58 times the bill. On this process, a habit of stuffing context costs more than any choice between models. At these prices the model fee for routine work is small. What decides the bill is how much you send it each time.

The decision rule

For any process that runs more than a few hundred times a month:

1. Measure the prompt, not the task. Log the input token count on the new model for 20 real requests. Use the largest, not the average.
2. If the largest is under about 80,000 tokens, you have room. Haiku 5.5 is a candidate, as long as its answers pass the check you already use.
3. If it is near or over 100,000, cut before you switch. Send only the relevant policy section, not the whole manual. Retrieve the three most similar past cases, not all of them. Summarize long threads once and send the summary.
4. If the work needs the long context, price it on both sides of the line and against Sonnet 5.5 before you decide. Haiku’s price advantage narrows sharply above the line.
5. Keep the check. A cheaper model is only cheaper if its output passes the same review. Count the reruns and the human fixes as part of the cost.

The 80,000 margin is our rule of thumb, not Anthropic’s. It leaves room for the updated tokenizer and for the occasional long email thread.

What to do this week

• Pull last month’s API usage and list every process that runs on a small model or should.
• For each, record the largest prompt in tokens from a sample of 20 real requests.
• Flag any process that crosses or comes within 20,000 tokens of the 100,000 line.
• Rerun one under-the-line process on Haiku 5.5 and compare cost and pass rate against what it runs on now.
• If you are on Max or Team, check whether the monthly API credit has appeared before you plan around it.

A consult block is where we pick the one or two workflows where AI saves the most time or money, and work on them with you, on your real material.