Claude Haiku 5.5 Pricing: A 90% Cut With a 100K Catch

“Claude api pricing” pulls 8,100 searches a month at a cost per click of $8.27, which tells you the people typing it intend to spend money. It is a commercial query, not a curiosity. And on October 7, 2026, Anthropic rewrote the answer to it.

Anthropic released Claude Haiku 5.5, calling it its cheapest, fastest, and most capable small model yet. The headline number is aggressive: $0.10 per million input tokens and $0.50 per million output tokens for prompts up to 100,000 tokens. That is roughly a 90 percent cut versus Haiku 4.5, and it lands exactly on top of OpenAI’s GPT-6 Luna pricing. On launch day the model also went live inside GitHub Copilot, which puts it in front of millions of developers who never visit a billing page.

If that were the whole story, Haiku 5.5 would be a straightforward win for anyone tracking their Claude API pricing. It is not the whole story. There is a threshold at 100,000 tokens, a new tokenizer that inflates token counts, and a migration that breaks a couple of common API patterns. This is what the cut actually means for a heavy user’s monthly bill, and how to capture it without tripping the catch.

What Claude Haiku 5.5 pricing actually is

The model ships with a two-tier price table, and the two tiers are five times apart. That gap is the entire story.

Prompt size	Input (per 1M tokens)	Output (per 1M tokens)
Up to 100,000 tokens	$0.10	$0.50
Above 100,000 tokens	$0.50	$2.50

Below the threshold, Anthropic’s small model is now a commodity-priced commodity: a tenth of what Haiku 4.5 charged, and a price point that makes high-volume background work genuinely cheap. Above the threshold, the same request costs five times more. It is still a cut versus Haiku 4.5, but a 50 percent one, not a 90 percent one.

Anthropic’s own workload estimate for the launch is about a 75 percent cost reduction. That number is the honest middle: it assumes a mixed workload where most calls stay under the line and some do not. If your calls all stay short, you beat 75 percent. If your calls routinely carry long context, you will not get anywhere near it.

The model itself is built for volume: summarization, compaction, classification, subagents, live customer support, and browser use. It carries a 1,000,000 token context window and a maximum output of 128,000 tokens, and it is served across multiple clouds through the usual Anthropic channels plus OpenRouter, which lists five upstream providers.

The 100K threshold is a routing decision, not a footnote

The pricing table reads like a detail. Operationally, it is an architecture instruction.

Most teams do not send a uniform request size. They send a mix: short classification calls at 2,000 tokens, medium summarization at 20,000, and occasional long-context jobs that carry a whole document history at 150,000. The first two categories drop into the cheap tier automatically. The third does not, and it is often the category that runs most frequently in an agent loop.

That is where the math turns. A subagent that accumulates context across turns will cross 100,000 tokens partway through a session and, from that point on, every token it sends is priced five times higher. The Blended average hides this. The line item does not. When you look at Claude API pricing for an agentic workload, the number you actually pay is set by how many of your calls sit above the threshold, not by the headline rate.

The practical move is to treat the threshold as a routing signal. Keep the cheap, repetitive steps under 100,000 tokens by resetting context between turns, summarizing before you pass state forward, and sending short briefs to subagents instead of full transcripts. Reserve the over-threshold calls for work that genuinely needs the long window, and accept the five times rate there on purpose rather than by accident.

The tokenizer changes the bill before the price does

Here is the part that gets buried under the headline. Haiku 5.5 ships with a new tokenizer that counts roughly 30 percent more tokens for the same text than the previous one.

A 30 percent token inflation against a 90 percent price cut still lands far below Haiku 4.5, so the net is still a large saving. But the comparison you run in your head, “the price went to a tenth,” is not the comparison the invoice runs. Your old prompt that measured 80,000 tokens now measures past 104,000, which means it just crossed into the expensive tier without a single character changing. A workload that saved 90 percent on paper can save noticeably less in practice purely because the ruler got longer.

This is the kind of invisible cost that separates a good Claude API pricing decision from a bad one. The rate card is public. The token count per request is yours to measure. Before you migrate, re-count your top ten prompts with the new tokenizer and see where they land relative to 100,000. If any of them now sit just above the line, you have found a routing job before it cost you anything.

What heavy users should actually do

Heavy users, the ones paying real money across Claude, GPT, and Gemini, do not need another benchmark table. They need four moves, in order.

Re-count your prompts with the new tokenizer. Do not trust the old token counts. Pull your ten highest-volume prompt shapes, run them through the Haiku 5.5 tokenizer, and note which ones now sit near or above 100,000 tokens. This is a half-day job that prevents a surprise invoice.

Route on the threshold, not on the model name. The cheap tier is not “Haiku.” It is “Haiku under 100K.” Build the split into your agent loop: short classification and compaction calls go to the cheap path, long-context reasoning goes where it needs to go and gets measured separately.

Run the migration test in a branch. The move to Haiku 5.5 breaks sampling parameters and assistant prefill. Anything that relied on those patterns will fail quietly or loudly depending on your client. Test one representative workflow end to end before you flip the default across your stack.

Price per completed task, not per million tokens. The two-tier table plus the new tokenizer makes per-token comparisons misleading. Instrument a single real workflow, count the calls and tokens it takes to finish one task, and compare that number across providers. A model that is cheaper per token can still be more expensive per finished job if it needs more turns.

Keep a second lane open. Haiku 5.5 matching GPT-6 Luna at $0.10/$0.50 is good news for buyers: two frontier labs now price the small tier identically, so switching one workflow away from the other costs you nothing in the cheap tier. Price parity plus portability is leverage when the next rate card lands.

The cross-provider picture

The Haiku 5.5 cut is not an isolated move. It is the small-model tier of a price war that has been running all year. GPT-6 Luna set $0.10 input about a month earlier. DeepSeek has been pushing off-peak discounts and forced migrations. Gemini 4 Argon ran a launch discount that quietly doubled after the window closed. Mistral’s Large 4 traded a per-token premium for an open-weight hedge.

For a heavy user, the lesson across all of them is the same: the headline rate is a marketing number, and the number you pay depends on prompt shape, token counts, cache behavior, and which tier your calls land in. Haiku 5.5 is a genuinely strong deal for short, high-volume work. It is a much smaller deal for long-context agent loops, and the 100,000-token line is what decides which one you get.

The bottom line

Claude Haiku 5.5 is the sharpest small-model price Anthropic has shipped, and at $0.10 per million input tokens it resets what “cheap” means for background work. The catch is structural, not hidden: the discount lives under 100,000 tokens, the new tokenizer inflates your counts by about 30 percent, and the 100K line is easy to cross in an agent loop without noticing.

Treat the threshold as a routing decision, re-count your prompts before you migrate, and price per completed task instead of per token. Do that, and the 90 percent cut is real money. Skip it, and you will be paying the five times tier while telling everyone you got the 90 percent one.