Deep Dive — Haiku 5.5's fine print: same price, three times the bill

Anthropic's launch yesterday boiled down to a number a child could grasp: its small model now costs a dime a million input tokens, matching OpenAI's GPT-6 Luna to the cent, with headline cuts of up to 90 percent. We flagged one caveat when we covered Claude Haiku 5.5 arrives at up to 90% lower prices in yesterday's brief. The fuller picture — from Anthropic's own footnotes, independent benchmark runs, and a preprint on how effort settings change what you pay — is more interesting: "the price of a model" now has at least three meters, and on Haiku 5.5 they disagree by a factor of three.

What actually shipped

Claude Haiku 5.5 went live October 7, 2026, on the Claude Platform and all three major clouds. For prompts up to 100,000 tokens — the bucket that held about 90 percent of traffic to the previous Haiku — it charges $0.10 per million input tokens and $0.50 per million output tokens. Above that threshold the same rates quintuple to $0.50 and $2.50. Cache reads dropped to $0.01/$0.05 in the cheap tier. Anthropic's own footnote does the honest math: 90 percent lower than Haiku 4.5 under 100,000 tokens, 50 percent lower above it, and "around 75 percent" cheaper on average once you account for the fact that Haiku 5.5's new tokenizer — the same one Sonnet 5.5 and Opus 5.5 use — squeezes slightly more tokens out of every task.

The model arrived with two gifts attached. Sonnet 5.5 cache reads were halved from $0.20 to $0.10 per million tokens, which Anthropic says makes most agentic work roughly 20 percent cheaper, since agents reread cached context constantly. And subscribers get monthly API credits — $100 for Max 5x, $200 for Max 20x, up to $500 pooled across a Team plan, use-them-or-lose-them, claimable against any model.

The capability jump is not the interesting part

Some history first. The Haiku line is Anthropic's volume tier, and its last update, Haiku 4.5, landed October 15, 2025 at $1 per million input and $5 per million output tokens — a price that was relatively expensive even at launch and, by the time OpenAI shipped GPT-6 Luna last month, a full ten times Luna's rate. Anthropic's own API lead, Alex Albert, noted the comparison columns in the launch chart are less than a year apart. One model generation absorbed a 10x price deflation and a benchmark rewrite in the same leap; that is the pace the budget tier is now moving at.

The benchmark story is loud enough to state quickly. Computer use on OSWorld 2.1 went from 15.7 percent on Haiku 4.5 to 72.4 percent — past GPT-6 Luna's 48.9 percent. Agentic coding on Terminal-Bench 4.0 went from a flat zero to 39.2 percent, with Luna at 16.4. Humanity's Last Exam without tools rose from 10.2 to 45.9 percent. Artificial Analysis scores the top configuration at 43 on its Intelligence Index, five points above Luna and 26 points above where the last Haiku sat a year ago. Anthropic's comparison table shows Haiku 5.5 beating Luna in every tested category.

Anthropic is careful about what this is not: its own page carries Sonnet 5.5 reference numbers (83.9 percent OSWorld, 70.6 percent Terminal-Bench) and says plainly that larger models remain the right choice for complex agentic coding. Haiku 5.5 is for compaction, summaries, classification, database lookups, and subagent errands — the work Rogo's engineers describe as sending a cheap model into a 10-K to pull the revenue line a bigger model needs for its deck.

One sticker, five models

Here is where the fine print starts. Haiku 5.5 is the first Haiku with an adjustable effort setting — five tiers of reasoning, from Low to Max. Artificial Analysis measured each: Low scores 29, the default Medium scores 34, High scores 38, Xhigh 41, Max 43. That ladder spans from "noticeably worse than Luna" to "beats Luna," and every rung shares the same $0.10 sticker. Reasoning cannot be switched off entirely; the model defaults to medium.

Simon Willison's hands-on test shows what the dial costs at the extremes: a low-effort SVG generation finished in 7 seconds for 0.0936 cents; the same prompt at max effort, complete with a reasoning trace, took 5 minutes 9 seconds and cost 3.3826 cents — roughly 36 times more for one picture of a pelican. Latency and spend are now dials the caller turns, and the model name alone no longer tells you where they're set.

Three meters that disagree

Meter one is the sticker, and it's genuinely low: matching Luna exactly under 100,000 tokens, where most requests land. Meter two is the tokenizer. Willison's token counter found the same long prompt consumes about 1.25 times as many tokens on Haiku 5.5 as on Haiku 4.5 — a hidden increase Anthropic acknowledges in its footnote rather than hides, but a real one for anyone migrating line-for-line.

Meter three is the work, and this is where the gap opens. Artificial Analysis's per-task measurements, tabulated by Neowin, show Haiku 5.5 at Max effort burning an average of 162,000 output tokens and 129,000 reasoning tokens per task at an average cost of $0.21. GPT-6 Luna at its own top setting burns 50,000 and 39,000 for $0.07. Identical price per token; three times the bill per unit of work; and five points of Intelligence Index bought with the difference. Haiku 5.5 also generates tokens more than twice as fast (243 per second versus 128), yet finishes a task in slightly longer — 427 seconds versus 390 — because it writes so much more.

The prompt-length boundary compounds this. Luna's price doesn't move until 272,000 tokens, and then only to $0.20/$0.75. Haiku 5.5's cliff comes at 100,000 and jumps fivefold. Anything routinely passing six figures of context is no longer comparing a dime against a dime.

Effort is part of the contract now

An August preprint, "The Price of Thinking", makes the underlying point rigorously. The authors ran a registered paired contrast on Sonnet 5: the same model with an explicit "high" effort request versus effort left off. The explicit contract cost $0.0103 more per call on average; the accuracy difference — plus 1.3 percentage points — fell inside the confidence interval and could not be detected, though the interval still permits a gain of up to 4.67 points. Their framing matters more than the numbers: an API buyer purchases a dated contract that includes the effort term, the output rail, and the price schedule, not a model name. Even within one provider, what an omitted effort setting means differs from model to model.

Haiku 5.5 makes that academic point a daily purchasing decision. Two vendors can quote $0.10 per million tokens and sell you different amounts of thinking, different tokenizers, and different cliffs.

Why the industry is playing along

The launch landed hours after OpenAI's GPT-6 and Interactive UI push, and a day after OpenAI's GPT-6.1 pricing — a price war hardening rather than easing, with Chinese commentary framing it as the US pulling ahead on the cost-performance ratio Chinese models used to own. Anthropic timed its cheapest model against the rival's budget tier while its own IPO is reportedly planned for November; a discount on the low end reads differently on the road show than a discount on Opus.

There's a strategic reading too, and we've been here before: the market-wide price per token fell 29 percent in a single month this summer in Tokens halved in price. The bill to make them didn't, and per-token stickers were already becoming a vanity metric. The pricing innovation that matters this year is the one we traced in Nobody can price what an agent buys — agents buy tasks, not tokens, so the seller's unit and the buyer's unit have diverged. Haiku 5.5 is a clean case: the sticker is honest, the meter that counts your work says 3x, and both claims can be printed on the same page without either being false.

One more line from the announcement deserves attention because it is unusual in a pricing story: Anthropic says Haiku 5.5 shows major alignment improvements over Haiku 4.5, with fewer instances of misaligned behavior and lower willingness to cooperate with misuse. Its cybersecurity safeguards are stricter than Haiku 4.5's but looser than Sonnet 5.5's — wide enough for defensive work, still blocking penetration testing — and its biology safeguards match the Sonnet 5/5.5 and Opus 5 generation. A budget model shipping with tighter guardrails than its predecessor cuts against the usual pattern where cheap tiers get the old, permissive weights.

What to watch

First, whether OpenAI or Google answer the Luna-matching price inside days, and whether anyone answers with a cost-per-task number instead of a per-token one — that would be the first crack at pricing in the unit buyers actually feel. Second, whether the effort ladder becomes table stakes for every small model: five tiers from one endpoint is convenient for vendors because it lets them quote one price for what is effectively five products. Third, the migration math — as model releases land closer together, tokenizer churn is becoming the recurring invisible price increase, and teams that don't re-run their own token counts on every upgrade will pay it without noticing.

Do you budget your AI stack by sticker price or by cost per task? Tell us in the comments.