Anthropic Priced a Capable AI at Ten Cents a Million Tokens. The Budget Tier Just Became the Agent Tier.

On October 7, 2026, Anthropic shipped Claude Haiku 5.5 at a headline rate of $0.10 per million input tokens and $0.50 per million output tokens. For prompts under 100,000 tokens, that is a 90% cut from the $1 and $5 rates of its predecessor, Haiku 4.5, and it lands at exactly the price OpenAI charges for GPT-6 Luna. It shipped a materially different one: a 1-million-token context window, native computer-use support, and the first adjustable effort setting in the Haiku line. The cheapest tier on the company's rate card is no longer a place for simple classification. It is now a place where agents can run.

That combination, a headline price and a capability jump, is why the release is being read as a pricing war rather than a product update. In the past month, OpenAI rolled out the GPT-6 family, DeepSeek shipped a 552-billion-parameter open-weight model under an MIT license, and Mistral put its trillion-parameter Large 4 into preview at roughly half its intended price. Haiku 5.5 is Anthropic's answer to all of them at once, and it collapses the industry's price floor to a level where the bill for running millions of AI calls a month starts to look like a rounding error.

The fine print on the ten-cent headline

The headline number is real, but it is not the whole price list. Haiku 5.5's $0.10 rate applies only to prompts up to 100,000 tokens. Cross that threshold and the rate jumps to $0.50 per million input tokens and $2.50 per million output tokens, five times higher, though still 50% below Haiku 4.5's rates. Anthropic's own release notes say the model costs about 75% less to run on average than its predecessor, not the 90% the headline rate suggests, and that gap has a documented explanation: a newer tokenizer counts the same text as roughly 30% more tokens, and about 10% of Haiku 4.5 traffic historically ran on longer prompts priced at the higher tier.

None of that undermines the release; it contextualizes it. A 90% cut on the short prompts that dominate real Haiku traffic is the largest single-generation cost drop any major lab has shipped in a tier like this. But it matters because the pricing war is now being fought on rate cards, and rate cards are becoming two-tier documents. Builders who quote the $0.10 figure in a budget spreadsheet without modeling their prompt lengths and token counts under the new tokenizer will rediscover the difference when the invoice arrives.

The headline rate is $0.10 a million tokens, but the strategic shift is bigger: Anthropic is selling capability that used to require a premium model at a price that used to buy a summarizer.

The benchmark jump matters more than the price

Ask Anthropic's own engineers what the launch is about and the price is the second slide. The first is OSWorld 2.1, a benchmark that tests whether a model can operate a real computer desktop: opening applications, filling forms, navigating file systems, chaining multi-step actions without a human correcting it. Haiku 4.5 scored 15.7% on the offline subset. Haiku 5.5 scored 72.4%, a 4.6-fold improvement in one generation, on scores Anthropic itself published. Terminal-Bench 4.0, which measures command-line tasks, went from 0% to 39.2%. On Anthropic's own tables, Haiku 5.5 also outscores GPT-6 Luna on every benchmark the two share, including computer-use and terminal work, with the narrowest gap on pure coding.

A score in the teens means a model fails most multi-step desktop tasks outright. A score in the seventies means it completes most of them. That is the difference between a demo model and a deployable one, and it is the kind of leap that used to arrive in flagship tiers, not the budget line. Haiku 5.5 also ships with an effort setting, a dial that lets developers choose how much thinking the model does per task.

The honest caveat: Sonnet 5.5 and Opus 5.5 remain far ahead on deep reasoning. Anthropic's own numbers put Sonnet 5.5 at 70.6% on Terminal-Bench 4.0, nearly double Haiku 5.5, and the company steers complex agentic coding to the larger models. This is not a flagship in disguise. It is a specialist that got good at exactly the repetitive, high-volume work that dominates real API bills.

Why every lab is cutting at once

Haiku 5.5's price was built to match GPT-6 Luna exactly, and the matching is the tell. When two closed-weight labs converge on $0.10 per million tokens for their budget tiers within two weeks of each other, price is no longer a differentiator and the competition moves back to capability and latency. That is terrain Anthropic wants to compete on, because Haiku 5.5's OSWorld jump gives it something to say there.

The third competitor in the room is the one that does not charge at all. DeepSeek's V4.1 Flash, released September 10 with a 1-million-token context window and an MIT license, offers hosted access at roughly $0.30 per million input tokens, but anyone willing to self-host pays only compute. Open weights put a ceiling on what anyone can charge for a capable small model, because a customer can always walk away and run the model itself. That ceiling has been falling all year, and Haiku 5.5's 90% cut reads as Anthropic's response to it.

Anthropic is not only cutting Haiku. The company also halved Sonnet 5.5's cache-read price, from $0.20 to $0.10 per million tokens, which it estimates will cut the cost of most agentic workloads on Sonnet by about 20%. And it added monthly API credits to its subscription tiers: $100 for Max 5x, $200 for Max 20x, and up to $500 pooled for Team accounts.

What this means for builders

The deployment pattern Anthropic is selling is routing: Haiku 5.5 takes the repetitive subtasks inside larger workflows, Sonnet or Opus handles the hard parts, and the blend of the two prices the job correctly. Early customer feedback cited in the release coverage points in that direction, with Asana reporting over a 30% reduction in task-completion latency against the model it had been using, though that is one customer's experience, not a guarantee.

The pattern works, but only for teams that measure instead of assuming. Haiku 5.5 changes the tokenizer, the thinking controls, the sampling parameters, and the computer-use tooling relative to 4.5, so migration is engineering work, not a rate-card swap. The new model can divide the same text into more billable tokens. Higher effort settings can sharply increase reasoning use. And the cheapest tier is bounded: anything above 100,000 tokens per prompt reprices at five times the rate.

The deeper read on the pricing war is that inference is becoming a commodity input: supply expanding fast, marginal cost falling, differentiation moving to tooling, latency, and reliability. That is good news for anyone building on top, because for two years AI inference was the most expensive input no one could route around. At $0.10 per million tokens, a customer-support bot processing 50 million input tokens and 10 million output tokens a month costs roughly $10 before volume discounts. The bottleneck has moved from the model bill to the quality of the workflow around it. That is the real story of Haiku 5.5.