Haiku 5.5's Rate Card Hides Three Multipliers, and Your Prompt Lengths Decide the Bill
The cheapest model tier now ties on list price. Whether you save depends on reasoning effort, tokenizer inflation and how many prompts cross one cliff.
Three multipliers sit between the list price and the bill
Anthropic discloses each of these in its own fine print. The finance partner usually sees them for the first time on the invoice.
- Tokenizer inflation. Haiku 5.5 counts the same text as roughly 1.25x to 1.3x more tokens than Haiku 4.5. Net of that, Anthropic's own estimate of typical savings is about 75%, not 90%.
- The 100K cliff. Above 100,000 tokens, rates rise to $0.50/$2.50. As the tier is described, the higher rate covers the whole prompt, so a 101K prompt costs about 5.1x a 99K prompt. Because of the tokenizer, the cutoff lands at roughly 80K old Haiku 4.5 tokens.
- Reasoning that can't be switched off. Effort defaults to medium and can't be fully disabled. In Vals' testing, Haiku 5.5 took 59 reasoning steps on a legal research task. Haiku 4.5 took 17.
Anthropic says about 90% of Haiku 4.5 requests fit under 100K tokens. That is a count of requests, not dollars, and the pitch version of the number blurs the two. Take an illustration in which 90 of every 100 requests average 5K tokens and 10 average 200K. The long ones carry about 82% of input tokens and about 96% of input spend at the higher rate.
The cheapest vendor depends on prompt length
| Prompt profile | Haiku 5.5 | GPT-6 Luna | Gemini 3.5 Flash-Lite |
|---|---|---|---|
| Under 100K tokens | $0.10 / $0.50, plus tokenizer and reasoning overhead | $0.10 / $0.50; tokenizer not disclosed | $0.30 / $2.50, flat |
| Around 150K tokens | $0.50 / $2.50 | Still base rate; premium starts only above 272K, at $0.20 / $0.75 | $0.30 / $2.50, flat |
| Batch jobs | Further 50% off | Not reported | Not reported |
MarkTechPost calculated that Luna is cheaper on list price for a 150K-token prompt. Two details complicate that. Roboflow found Haiku 5.5 cheaper than Luna at high effort on vision tasks. Sonnet 5.5's halved cache-read price of $0.10 per million also equals Haiku 5.5's uncached input price. For input-heavy agent loops with a stable prompt prefix, moving up a tier can cost far less than the 20x list gap implies. Output is still 20x more expensive, so the answer turns on each workload's input-to-output ratio.
Cheap models only lower cost inside a cascade
In a cascade, a cheap model answers first and hands its uncertain cases to an expensive one. Log10's clinical tests have the cleanest public numbers. A low-cost model answered every question and sent its least-confident 20% to Claude Opus. Accuracy rose from 80.0% to 87.9%. Escalating a random 20% reached only 83.0%. The cascade cost $5.46 per 1,000 questions. Opus alone cost $25.33 and scored 94.9%. Closing that last gap works out to about $0.28 per avoided error. In any flow where a mistake costs more than that, escalation is the better buy.
Log10 also tested the cheap model as a judge of other models' answers, and it did worse than simply labeling every answer correct. Haiku 5.5's vendor-reported jump on computer-use tasks, from 15.7% to 72.4%, makes it a usable sub-agent, but the judging result rules it out as a verifier.
Graded as a judge, the cheap model lost to a baseline that marks every answer correct.
The move
Decide feature by feature from real traffic, not with a company-wide switch. Sort each flow on two axes, prompt length and the cost of a wrong answer. Short prompts at high volume where errors are cheap, such as classification, triage and support drafts, are where Haiku 5.5 or Luna win. Long-context and high-stakes flows need cost per completed task measured before anything moves.
What to do
Replay 1–2 weeks of production prompts this sprint through Haiku 4.5, Haiku 5.5 at low and medium effort, and GPT-6 Luna. Report cost per completed task, reasoning tokens, and the share of requests over 100K tokens for each feature.
Set a context budget of about 90–95K Haiku tokens per feature, with alerts, and write retrieval and chunking requirements into the PRD before any traffic moves.
Prototype a confidence cascade on your highest-volume, lowest-stakes flow this sprint. Escalate the least-confident 20% and compare it against three alternatives: the cheap model alone, a random 20% escalated, and the frontier model alone.