Anthropic cut the subagent tax. Check the tokenizer math.

Anthropic launched Claude Haiku 5.5 on Wednesday, and the headline number is a 90% price cut against Haiku 4.5. That number is true in exactly one reading and misleading in two. Here is the reading that matters for anyone running agents: the small-model tier is now a commodity, Anthropic just matched OpenAI's price to the cent, and the cheapest model in the Claude lineup is finally capable enough to be the default worker for routine agent tasks. Everything else is arithmetic you should redo yourself.

What shipped: Haiku 5.5 costs $0.10 per million input tokens and $0.50 per million output tokens for prompts under 100,000 tokens. Cache reads are $0.01, cache writes $0.125. Above 100K prompt tokens the rates quintuple to $0.50/$2.50. The model carries a 1M-token context window, is the first Haiku with an adjustable effort setting (adaptive thinking on by default, effort defaulting to medium), and is available now on the Claude Platform, AWS, Google Cloud, and Azure under the model string claude-haiku-5-5. Two companion changes shipped in the same announcement: Sonnet 5.5 cache reads dropped from $0.20 to $0.10 per million, which Anthropic says makes it about 20% cheaper on most agentic work, and Max 5x/20x and Team subscribers get $100, $200, and up to $500 a month in pooled API credits. Anthropic also put beta computer-use and browser-use toolsets into the Python and TypeScript SDKs.

Why it matters for operators. Anthropic is explicitly positioning Haiku as the subagent: the worker that retrieves the revenue figure for one slide while Opus assembles the presentation. That framing matches how serious agent deployments actually spend money. The expensive model plans; cheap models do the hundred repetitive calls. If the cheap model is good enough for retrieval, classification, compaction, and database queries, the blended cost of a multi-agent workflow collapses. Anthropic's benchmarks for Haiku 5.5 say it is: 72.4% on the offline subset of OSWorld 2.1 (Haiku 4.5 scored 15.7%), 39.2% on Terminal-Bench 4.0 (the predecessor scored zero), 46.4% on FrontierCode 1.1. Against GPT-6 Luna, secondary reporting puts Haiku 5.5 ahead at 72.4% vs 48.9% on OSWorld and 39.2% vs 16.4% on Terminal-Bench, at identical short-context list prices.

Now the arithmetic, because the vendor's numbers need their receipts.

First, the 90% cut applies only under 100,000 prompt tokens. Above the threshold the discount is 50%. Anthropic says about 90% of Haiku 4.5 requests stayed under the line, and puts the average saving at 75%. If your workloads skew long, your savings skew lower.

Second, the tokenizer. A developer test surfaced by AI Weekly's coverage found the new tokenizer counts roughly 25% more tokens than Haiku 4.5 for the same text (another writeup put it near 30%). So some of the per-token discount is eaten by per-token inflation. The real comparison is cost per unit of work, not cost per million tokens. This is the oldest trick in the pricing playbook: shrink the unit, cut the rate, keep the invoice.

Third, the comparison with Luna. Identical list prices do not mean identical bills. Luna's higher tier kicks in above 272K input tokens at $0.20/$0.75; Haiku's kicks in above 100K at $0.50/$2.50. For a 150K-token prompt, Luna is cheaper on list price. The benchmark gap is company-reported and has no named independent replication that we could find. Treat it as directional, not decisive.

What remains unknown: whether Luna, Gemini, or the open-weight field answers on price before year-end (see September's price-war roundup for how fast these cuts are landing); whether the effort knob produces predictable capability tiers or just a cost dial with vibes; and whether any independent benchmark house reproduces the agent scores.

The decision for operators is concrete. Re-run your routing thresholds this week, with real token counts from your own logs. Measure cost per completed task at each effort setting, not per million tokens. Check how many of your prompts sit above 100K, because that threshold is where Haiku's pricing stops looking like Luna's. And if you are on Max or Team, claim the monthly API credits; they are the cheapest experiment budget you will get this quarter. The small model is no longer the compromise tier. It is the volume tier, and Anthropic just priced it like one. Your job is to verify the invoice, not the press release.