Claude Haiku 5.5 cuts small-model prices by up to 90 percent, and the catch is a 100,000-token line

Anthropic’s small model just got a tenth as expensive for most of the work it actually does. Claude Haiku 5.5, released October 7, lists at $0.10 per million input tokens and $0.50 per million output tokens, a 90 percent cut from Haiku 4.5. That headline holds only below a line that developers now have to design around: prompts over 100,000 tokens pay five times the rate.

What Anthropic shipped

Haiku 5.5 is available as claude-haiku-5-5 on the Claude API, Amazon Bedrock, Google Cloud, and Microsoft Foundry. It has a 1M-token context window, up to 128,000 output tokens, adaptive thinking with medium as the default effort, and a June 2026 knowledge cutoff, the same as the rest of the 5.5 family. Anthropic positions it for high-volume, latency-sensitive work: classification, extraction, summaries, routing, database queries, and sub-agent calls.

The pricing is tiered by prompt length. Up to 100,000 tokens, Haiku 5.5 costs $0.10 input and $0.50 output, with cache reads at $0.01. Above that, it costs $0.50 and $2.50. Haiku 4.5, which remains available, is $1 and $5 at any length. Anthropic says about 90 percent of requests to Haiku 4.5 fall under the 100,000-token line and puts the average saving at around 75 percent. Anthropic’s pricing docs note that Claude 4.6 and later models otherwise bill the full 1M window at one flat rate.

Anthropic also halved cache-read pricing on Claude Sonnet 5.5 to $0.10 per million tokens, which it says makes Sonnet roughly 20 percent cheaper on most agentic work.

For developers, the practical change is that the cheapest Claude tier now matches OpenAI’s cheapest current API tier on list price, which changes the default answer to which small model to route bulk work through.

The tokenizer offsets part of the cut

Haiku 4.5 used Anthropic’s older tokenizer. Haiku 5.5 moves to the newer one used since Claude Opus 4.7, which Anthropic’s pricing docs say produces approximately 30 percent more tokens for the same text, depending on content. Anthropic’s own launch post acknowledges that Haiku 5.5 “uses slightly more tokens per task.”

VentureBeat’s analysis makes the same point: token prices alone do not establish the cost of finishing a job, because token consumption, retries, and accuracy all feed the bill. A 90 percent cut on the rate card still leaves a large saving after a 30 percent token increase, but teams forecasting spend from Haiku 4.5 token counts will undercount usage.

The benchmark figures in the launch post, including Terminal-Bench and OSWorld scores, are Anthropic’s own and had no independent verification at release.

How it lines up against GPT-6 Luna

The list price matches OpenAI’s GPT-6 Luna, which its API docs put at $0.10 input and $0.50 output with a 1.05M-token context window. The structural difference is where the surcharge starts. Luna’s higher rate applies to requests above 272,000 tokens. Haiku 5.5’s applies above 100,000. Workloads that feed long documents, large retrieved contexts, or full repository slices into a cheap model will cross Anthropic’s line far sooner.

What it means for developers

Teams running Haiku 4.5 for classification, extraction, or routing have a straightforward test: switch the model ID on existing evals, compare accuracy, and measure tokens per completed task rather than tokens per request. For most short-prompt workloads the result should be a substantially smaller bill.

The design question is prompt length. Pipelines that currently pass 120,000 tokens into a small model may now be cheaper if they trim context, chunk documents, or lean on prompt caching to stay under 100,000. Agent frameworks that hand sub-tasks to a small model should check where their context accumulates, because a long-running sub-agent can drift across the threshold mid-session.

The small-model tier is no longer a rounding error in agent budgets, and the pricing now rewards teams that control how much context they send.

What to watch

The open questions are whether independent evaluations confirm Haiku 5.5’s quality claims against Haiku 4.5 and Luna, how often real agent workloads cross the 100,000-token line, and whether other providers adopt length-tiered pricing for their cheapest models. Stackmaven’s follow-up will revisit around January 8.

Sources cited

1. Claude Haiku 5.5 (Anthropic) www.anthropic.com
2. Pricing (Claude API docs) platform.claude.com
3. Models overview (Claude API docs) platform.claude.com
4. Anthropic launches Claude Haiku 5.5 with 90% API price reduction, matching GPT-6 Luna (VentureBeat) venturebeat.com
5. GPT-6 Luna model reference (OpenAI API docs) developers.openai.com