Claude Haiku 5.5 arrives with massive price cuts proving the AI pricing arms race is far from over

What happened

Anthropic has released Claude Haiku 5.5, a new entry in its lower-cost model tier that posts a sharp jump in computer-use capability alongside steep reductions in token pricing.

On the OSWorld benchmark, which tests a model's ability to operate a computer interface to complete tasks, Haiku 5.5 scores 72.4 percent, up from 15.7 percent for its predecessor — a substantial leap in a single generation. Anthropic has paired this with token prices cut by as much as 90 percent compared with the prior Haiku model, positioning it as a cheaper option for high-volume or latency-sensitive workloads.

The release also introduces a new tokenizer, which changes how text and instructions are broken down for processing. According to the reporting, this tokenizer causes the model to consume more tokens per task than before, meaning some of the headline price reduction is offset in practice by higher token consumption for equivalent work.

Why it matters

This release is best read as a technology and cost-economics story rather than a customer-experience one. The scale of the OSWorld improvement suggests Anthropic is pushing its smaller, cheaper models toward genuine agentic competence — the ability to navigate interfaces and complete multi-step tasks autonomously — rather than treating "lite" tiers as stripped-down chat assistants. That matters for any organisation building automation, back-office agents or AI-assisted interfaces on top of foundation models, since it widens the range of tasks that can be delegated to a lower-cost tier.

At the same time, the tokenizer trade-off is a useful reminder that headline price cuts in the AI market don't always translate one-to-one into lower real-world spend. Teams evaluating model switches need to benchmark actual token consumption for their own workloads, not just list prices, before assuming savings.

By the numbers

• 72.4 percent — Claude Haiku 5.5's score on the OSWorld computer-use benchmark.
• 15.7 percent — the predecessor model's score on the same benchmark.
• Up to 90 percent — the reduction in token pricing claimed for Haiku 5.5 versus the prior Haiku model.

The Renascence take

The interesting signal here isn't the discount — it's what a cheap model being this competent at computer use implies for how work gets automated.

Most coverage of AI pricing wars treats the headline percentage cut as the story, but the real operational question is cost-per-completed-task, and this release shows why that distinction matters. A model that's 90 percent cheaper per token but consumes more tokens to finish the same job may deliver far less net savings than advertised — and teams that provision agentic workflows on sticker price alone will discover the gap at invoice time, not at launch. The deeper shift worth watching is that "budget" model tiers are now capable enough to run real interface-driving agents, which changes the calculus for where organisations automate first: not just chatbots and copilots, but the fiddly, repetitive screen-based tasks that have resisted automation until now. Leaders evaluating this generation of models should pilot against their own task mix and measure total token spend end-to-end, rather than comparing list prices across vendors.