Anthropic Ships Claude Haiku 5.5, Its Cheapest Model Yet, and Cuts Sonnet 5.5's Cache Price the Same Day

Anthropic released Claude Haiku 5.5 on Oct. 7, calling it "the cheapest, fastest, and most capable small model we've ever released." The launch closes out a three-model refresh that began with Claude Opus 5.5 on Sept. 22 and continued with Claude Sonnet 5.5 on Sept. 28. Anthropic folded a second announcement into the same release: a same-day price cut to Sonnet 5.5's cache-read rate.

The cheapest, fastest, and most capable small model we've ever released.

For prompts under 100,000 tokens, Haiku 5.5 costs $0.10 per million input tokens and $0.50 per million output tokens -- what Anthropic calls a 90% cut from Haiku 4.5's $1.00/$5.00 rate. Above that threshold, pricing rises and the cut shrinks to 50%. Anthropic's separate headline claim -- that the model "costs around 75% less to run" on average -- blends both tiers across the company's own internal usage mix, which is a different number from either tier cut alone.

That $0.10/$0.50 entry rate is, per The New Stack, "in line with" OpenAI's cheapest model, GPT-6 Luna -- one more sign that the bottom tier of frontier-model pricing has become a genuine price-matching race rather than a one-lab discount.

The Sonnet 5.5 change is narrower and easy to miss inside a small-model launch post: cache-read tokens drop from $0.20 to $0.10 per million, a 50% cut applying only to the share of a request served from cache. Because cache reads make up a large share of token volume in most agentic work, Anthropic says the change lowers the cost of running Sonnet 5.5 on typical agentic tasks by roughly 20% -- Sonnet 5.5's input ($2.00), output ($10.00), and cache-write rates are unchanged.

Pricing is where Haiku 5.5 is most directly comparable to what came before it and to OpenAI's equivalent tier:

On capability, Anthropic's own benchmark tables put Haiku 5.5 well ahead of its predecessor and roughly in range of GPT-6 Luna on one widely-used agentic-computer-use test, OSWorld 2.1: 72.4% for Haiku 5.5, against 15.7% for Haiku 4.5, 48.9% for GPT-6 Luna, and 83.9% for Sonnet 5.5 -- still well behind Sonnet 5.5 on anything requiring sustained agentic work. Anthropic also publishes results on GDPval-AA, Terminal-Bench 4.0, and Humanity's Last Exam showing the same broad pattern.

Every one of those numbers, for every model, comes from Anthropic's own evaluation harness, scored by Anthropic, against baselines Anthropic chose. No independent evaluator has checked Haiku 5.5 yet. That's a different position than Sonnet 5.5, which already carries a score of 56 on the Artificial Analysis Intelligence Index -- the one number on the Scoreboard that never comes from a vendor's own test suite. Anthropic's own charts also break each benchmark out by how much the model is allowed to "think": the same Haiku 5.5, run at Low, Medium, High, Xhigh, or Max effort, climbs steadily up every one of them -- a reminder that a single benchmark number for this model is already a choice of setting, not a fixed fact about the model itself.

Haiku 5.5 is also the first Haiku-class model with an adjustable effort setting -- Low, Medium, High, Xhigh, or Max -- letting a developer trade cost for intelligence on the same model rather than switching to a larger one. Anthropic positions it as a subagent that pairs with Opus 5.5 and Sonnet 5.5 on coding work, and for narrow, high-volume jobs: summarization, classification, database queries, compaction. (An effort dial changes what a sticker price actually means -- two customers running the same model at different settings pay meaningfully different amounts per task, which is part of why Anthropic's own 75% figure is an average rather than a single number.)

The release lands inside a broader price war -- OpenAI, Anthropic, and Google have each cut prices on at least one model tier since late July -- and it comes as Anthropic prepares for a reported IPO, per Reuters' review of the company's S-1 filing. Cheap small models are also how a lab without consumer-chatbot scale competes for the highest-volume, lowest-margin slice of enterprise AI spending -- the queries too numerous and too routine to justify a flagship model's price. The New Stack frames the same pressure from the buyer's side: enterprise customers are pulling back after a year of heavy, open-ended spending on frontier models -- what the piece calls the "tokenmaxxing" era -- and weighing open-weight alternatives from both U.S. and Chinese developers against anything a closed lab charges.

Anthropic's model lineup is now four names deep, and Haiku 5.5 slots into the bottom of it: Haiku for fast, cheap, high-volume work; Sonnet and Opus for heavier coding and enterprise use; and Fable at the top for extended, multi-day workloads. A fifth name, Mythos, is a reduced-safeguard variant restricted to vetted members of Anthropic's Cyber Verification Program, which the company expanded into three access tiers on Oct. 6 -- a reminder that "Claude" is no longer one model with one set of guardrails, but a family whose access and restrictions vary by name as much as by price.

Haiku 5.5 is live now on the Claude Platform, Amazon Web Services, Google Cloud, and Microsoft Azure. Anthropic's own launch post includes endorsements from early customers -- Asana calls it "a noticeably snappier experience," HubSpot says it "got the best score we've seen on this suite yet, at 92.8% averaged over three runs," and Cognition says the model "joins the sidekick lineup in Devin Fusion as an excellent option." Those are Anthropic's chosen quotes from its own announcement, not independently selected case studies.