The Price Cut, Token by Token
Anthropic released Claude Haiku 5.5 on Wednesday, 7 October 2026, at $0.10 per million input tokens and $0.50 per million output tokens for prompts up to 100,000 tokens. That is a 90 percent cut from Haiku 4.5's $1 and $5, and it exactly matches the pricing of OpenAI's GPT-6 Luna, which shipped last month. The model is available now on AWS, Google Cloud, and Azure, according to the-decoder.com.
The predecessor was expensive for its class. Haiku 4.5 shipped almost a year earlier at $1 per million input and $5 per million output, which simonwillison.net describes as relatively expensive even at launch and a full 10x the price of GPT-6 Luna. Anthropic's new model erases that gap in a single step, dropping to parity with a competitor that had been undercutting it by an order of magnitude.
The parity holds only below 100,000 tokens. Beyond that threshold Haiku 5.5's price increases 5x to $0.50 input and $2.50 output. Luna also raises prices for long prompts, at 272,000 tokens, but only to $0.20 and $0.75. A developer running 150,000-token prompts therefore pays 2.5x Luna's input rate on Haiku. The headline match is therefore a match for short-context workloads specifically, not across the board.
Benchmarks: Where Haiku 5.5 Beats Luna and Where It Doesn't
On benchmarks Anthropic claims two wins over Luna. Haiku 5.5 scored 72.4 percent on the OSWorld 2.1 offline subset and 39.2 percent on Terminal-Bench 4.0, per decrypt.co, ahead of OpenAI's budget model on both. Those results map onto the use cases Anthropic assigns the model: live customer support, operating a web browser for the user, querying databases, and summarizing documents at high volume.
The ceiling is Sonnet. Haiku 5.5's 39.2 percent on Terminal-Bench 4.0 sits well behind Sonnet 5.5's 70.6 percent on the same coding test, per decrypt.co. That gap frames the product: Haiku is built and priced for throughput on routine agentic chores, not for the hardest coding runs. Buyers choosing between the two Anthropic models are trading a roughly 31-point benchmark deficit for a cheaper tier.
The cut lands inside a wider pricing fight. finance.biggo.com reports OpenAI and Anthropic are escalating a pricing war as both prepare for IPOs, with OpenAI expanding GPT-6 access to free users alongside Anthropic's 90 percent cut. the-decoder.com reads the release as proof the AI pricing arms race is far from over. Where margin survives that competition is the open question. Both labs now price their smallest models at the same rates.
The Hidden Costs: Tokenizer, Long Context, and the Subscriber Credits
The sticker price conceals a tokenizer change. Haiku 5.5 uses a new, less generous tokenizer than Haiku 4.5: simonwillison.net's Claude Token Counter tool shows the same long prompt consuming around 1.25x as many tokens under the new model. Because billing is per token, that multiplier functions as a hidden price increase layered on top of the headline cut, and it applies at every context length, including the cheap tier.
Long-context economics diverge sharply. Above 100,000 tokens Haiku 5.5 costs $0.50 input and $2.50 output, while Luna above 272,000 tokens costs $0.20 and $0.75. simonwillison.net's verdict is direct: if workloads fit within 100,000 tokens, Haiku matches Luna's price and reports higher benchmark scores; above that threshold, Luna looks like a much better deal. The tokenizer change raises effective long-context costs further. Developers should measure actual prompt lengths before assuming parity.
Two adjacent changes shipped in the same week. Anthropic halved Sonnet 5.5's cache-read price to $0.10 per million tokens, cutting the cost of reused context for agent workloads. It is also adding monthly API credits for Max and Team subscribers, per decrypt.co, a mechanism that converts subscription revenue into API spend. the-decoder.com confirms the credits roll out this week. Both moves bind existing customers closer to the paid API.
Sources
- simonwillison.net — Claude Haiku 5.5 Matches GPT-6 Luna's Pricing - and Anthropic Starts Paying Subscribers…
- decrypt.co — Anthropic Launches Haiku 5.5: Its Cheapest and Fastest Claude Model Yet (2026-10-07)
- the-decoder.com — Claude Haiku 5.5 arrives with massive price cuts proving the AI pricing arms race is far… (2026-10-07)
- biggo.com — Nvidia and Micron Set to Dominate S&P 500 Earnings as AI Spending Boom Drives 29.5%… (2026-10-08)
- New Horizon daily digest, 2026-10-08 (topic provenance)