VendorsOCT 09, 2026

Anthropic cuts Haiku 5.5 to $0.10 per million tokens, matching OpenAI's GPT-6 Luna

The 90% price cut on sub-100K-token requests resets the cost floor for the small-model work that powers SaaS tools small businesses rely on to find and win customers.

BY GRETA REINHART · ENTERPRISE SAAS

Anthropic launched Claude Haiku 5.5 on Oct. 7 with a 90% price cut on sub-100,000-token requests, dropping input to $0.10 and output to $0.50 per million tokens and matching OpenAI’s GPT-6 Luna at the base rate. For requests above the 100,000-token threshold, Haiku 5.5 runs $0.50 input and $2.50 output per million, a 50% cut from the prior generation. Haiku 4.5 had been $1.00 input and $5.00 output.

The two-tier structure matters because Anthropic says roughly 90% of Haiku 4.5 traffic sat under the shorter tier. After accounting for an updated tokenizer, the company puts average workload savings at about 75%.

Cache reads moved too. Haiku 5.5 cache reads are $0.01 per million tokens on the sub-100K tier and $0.05 above it, down from a flat $0.10 on Haiku 4.5, which puts the small tier at parity with GPT-6 Luna’s $0.01. Sonnet 5.5 cache reads were halved to $0.10 per million tokens, which Anthropic estimates trims typical agent workloads by roughly 20%.

This is the third Claude 5.5 model in a month, per Reuters, and arrives with Anthropic’s planned IPO expected before Thanksgiving, according to Yahoo Finance. The sequencing is its own tell: a pricing floor reset is the kind of story a company wants in the air during a roadshow, especially one running parallel to this week’s OpenAI revenue reset and Anthropic’s recent Claude for Startups expansion.

For small-business SaaS, Haiku-class work is the quiet machinery of customer acquisition: summaries, classification, database queries, extraction. Lead scoring, CRM tagging, support triage, and the first pass on outbound copy all sit here. A 75% cut at the model layer gives vendors three choices in Q4 repricing. They can pass savings through as sticker cuts. They can raise usage ceilings on existing plans. Or they can turn on accuracy-sensitive features they’d throttled for unit economics.

Two caveats. The 90% headline only applies under 100,000 tokens. And Anthropic’s benchmarks are vendor-reported: Haiku 5.5 posts a Terminal-Bench 4.0 score of 39.2% at maximum effort, but roughly 20% at default. Buyers evaluating a repriced tool in November should ask which setting the vendor is actually running.

Sources

About the author

Greta Reinhart

ENTERPRISE SAAS

Greta Reinhart tracks the enterprise software stack from San Francisco — data platforms, AI bundling, seat pricing, and channel checks across the largest SaaS vendors. She files on go-to-market shifts, packaging changes, and quarterly enterprise reads.