Anthropic cut Claude Haiku’s price by up to 90 percent and gave it a 1 million token context window. The cheap-model tier of the price war just got a lot more capable.

Anthropic released Claude Haiku 5.5 on 7 October 2026, its smallest current model and the direct successor to Haiku 4.5. The headline change is price: Anthropic set input at 0.10 dollars per million tokens and output at 0.50 dollars per million tokens for prompts up to 100,000 tokens, rising to 0.50 and 2.50 dollars per million above that length. Haiku 4.5 charged a flat 1.00 dollars input and 5.00 dollars output per million tokens, so the new short-prompt rate is a 90 percent cut and the long-prompt rate is a 50 percent cut.

The model’s context window grew from 200,000 to 1 million tokens, matching the rest of the current Claude lineup (Claude Sonnet 5 and Claude Opus 5.5 both already sit at 1 million), and maximum output rose to 128,000 tokens, according to Anthropic’s own model-comparison table. Anthropic describes Haiku 5.5 as its first Haiku-class model with an adjustable effort setting, which lets a developer trade latency against intelligence per request rather than per model choice.

Why it matters

Haiku is the model tier most high-volume production systems actually run on: classification, extraction, routing, subagents, and summarisation, where cost per call compounds fast. A 50 to 90 percent price cut combined with a 5x larger context window changes what is economical to automate with a small model, not just what is cheap to prototype with one.

The cut also lands squarely inside the ongoing price war: Anthropic’s own materials position the sub-100k-token rate as matching OpenAI’s GPT-6 Luna pricing, which shipped the same week. Buyers evaluating small models now have at least two frontier labs pricing at the same point, which should put further downward pressure on the tier.

The pricing is split by prompt length, not flat

Anthropic’s two-tier structure means the advertised 90 percent discount applies only below the 100,000-token threshold. A workload that regularly sends prompts longer than 100,000 tokens pays 0.50 dollars input and 2.50 dollars output per million tokens instead, a 50 percent cut from Haiku 4.5 rather than 90 percent. Cache reads follow the same split: 0.01 dollars per million under the threshold, 0.05 dollars above it, against Haiku 4.5’s flat 0.10 dollars. Teams that previously treated Haiku pricing as a single number will need to model which side of the 100,000-token line their typical prompt falls on.

Anthropic made one further pricing change the same day, unrelated to Haiku itself: cache reads on Claude Sonnet 5.5 dropped from 0.20 to 0.10 dollars per million tokens, which Anthropic says cuts the cost of agentic tasks that repeatedly re-read a cached prompt by roughly 20 percent.

The benchmark numbers are vendor-reported and self-selected

Anthropic reports Haiku 5.5 scoring 72.4 percent on OSWorld 2.1, a benchmark that tests whether an agent can operate a real computer desktop to complete long, multi-step tasks, and 39.2 percent on Terminal-Bench 4.0, which measures multi-step professional tasks completed inside a command-line interface. Anthropic also cites 45.9 percent on Humanity’s Last Exam without tools, a test of expert-level academic knowledge and reasoning, and a GDPval-AA v2.1 score of 1620 Elo, which Anthropic frames as a measure of real-world professional work across 44 occupations.

All four figures come from Anthropic’s own release materials, with no independent benchmark run cited alongside them. They describe Haiku 5.5’s performance on Anthropic’s chosen test suite, not a third-party audit, and should be read as Anthropic’s characterisation of its own model rather than a settled external result.

What this means for buyers

Anyone running high-volume Haiku-tier workloads should re-run their own cost model against the two-tier pricing rather than assuming a flat percentage saving, since the real discount depends on typical prompt length. The larger context window also removes a reason some teams previously routed long-context work to Sonnet-tier models purely for context size; that trade-off is now closer to a pure cost-versus-latency decision. Anyone deciding between Claude and OpenAI’s small-model tier should compare like for like: Anthropic’s own framing against GPT-6 Luna only holds below the 100,000-token threshold.

What would change our reading

An independent, third-party run of OSWorld, Terminal-Bench, or GDPval-AA on Haiku 5.5 would either confirm or complicate Anthropic’s self-reported scores; until one appears, the benchmark claims stand as vendor-reported only. Early real-world reports already flag that Haiku 5.5 can consume more tokens per task than Haiku 4.5 despite its lower per-token price, which would change the actual cost-per-task picture; a wider sample of production cost data over the coming weeks would settle whether net costs fall as much as the headline per-token cut suggests.