Anthropic Just Made Its Cheapest AI Ever, Right Before Its IPO

Anthropic has launched the cheapest AI in its history, a small model built for high-volume work that costs about seventy-five percent less to run than its predecessor. The release landed in October 2026 as the third model in the company's current family in about a month, and it arrives just as Anthropic prepares to pitch investors ahead of a planned public listing.

The new Haiku model is priced at ten cents per million input tokens and fifty cents per million output tokens for prompts under one hundred thousand tokens, with rates rising for longer prompts, according to Reuters. The wire service also notes it is the first small model from the company with built-in safeguards for a narrow set of high-risk cybersecurity requests, while everyday tasks remain unaffected.

The price cut is not subtle. It is a deliberate move in an industry-wide race to the bottom on model pricing, and it signals how aggressively Anthropic wants to own the cheap-and-fast tier of the market before its shares start trading. For developers running agents at scale, the sticker price is now nearly an afterthought.

What the cheapest AI actually costs

The pricing has two tiers. Short prompts — anything under the one-hundred-thousand-token threshold — get the headline rate, while longer prompts cost more. Compared with the previous Haiku generation, the new short-prompt rates are roughly ninety percent lower per token.

In short, it is the cheapest AI Anthropic has ever put on the price sheet. As reported by VentureBeat, the short-prompt pricing matches Luna's headline rate and puts direct pressure on other budget options, including Flash-Lite from Google and Grok from xAI. The publication frames the launch as a matching move: Anthropic is not undercutting the market so much as refusing to be undercut.

The company also added an adjustable effort setting, a first for its small-model line, letting developers trade cheaper answers for smarter ones on a per-task basis. And the model is available free to use for consumers, as reported by Notebookcheck, which widens the funnel of users who will meet the cheapest AI before they ever pay for it.

Behind the bragging rights

Anthropic's own numbers tell a dramatic story. On OSWorld, a test of whether an agent can operate a real computer, the new Haiku scores about seventy-two percent, up from under sixteen percent for its predecessor and ahead of Luna at about forty-nine percent, according to the company's published results. On Terminal-Bench, a test of multi-step work at the command line, it jumps from zero to nearly thirty-nine percent, again ahead of Luna's roughly sixteen percent.

An independent check tells a more complicated story. Artificial Analysis tested the model and found the performance gains real but smaller than the headline figures suggest — and the independent testing outfit found a catch that matters for budgets. The new Haiku uses roughly three and a half times more output and reasoning tokens than Luna to reach similar results, which pushes the estimated cost of a completed task to about twenty-one cents, versus roughly seven cents for its rival. Those findings were summarized as reported by Edufabtech.

In other words, it remains the cheapest AI per token the company has ever offered, and the per-token price is the cheapest AI has ever been at Anthropic, but the per-task price depends on how chatty the model gets. Developers optimizing for completed work, not token volume, will need to measure their actual workloads rather than trusting the sticker.

Why the timing is about the IPO

The cheapest AI is also a marketing weapon aimed at a very specific audience: the investors Anthropic is about to meet. The company is expected to begin IPO investor meetings in mid-October 2026, and arriving with the cheapest frontier-class small model on the market is a growth story told in one slide. The launch completes a three-model rollout that began with the lineup's larger models in late September 2026 — a full product shelf, refreshed just in time for the roadshow.

The enterprise push is running in parallel. Anthropic has announced two nine-figure initiatives this season: the Genesis Mission, aimed at power grids, and the Claude Frontier Academy, aimed at bank back offices, as reported by Morningstar. Cheap models win developer mindshare; enterprise deals win the revenue multiples that public-market investors price.

There is a trust angle in the launch too. The new safeguards for high-risk cybersecurity requests — the first of their kind in the small-model line — give Anthropic an answer for safety-conscious enterprise buyers and regulators alike. It is a small technical detail with large political value in a year when frontier labs are under unusual scrutiny.

What the cheapest AI means for developers

For builders, the practical upshot is a new default for agent workloads. Anthropic pitches the model as a helper that its larger models can hand smaller pieces of a coding job to, making it the cheap muscle in a multi-model team. Summarization, classification, database queries, live support, voice agents, and browser tasks are the named targets — the unglamorous high-volume work where a ninety-percent price cut changes the math on what is worth automating.

The catch is measurement. With independent tests showing the model spends tokens more freely than its rivals, the winning strategy is to benchmark cost per completed task on real workloads, not per million tokens. Price wars are won on receipts, and even the cheapest AI per token still sends an invoice — sometimes a bigger one than a pricier rival.

For more on how the industry's pricing race is reshaping what developers pay, see our AI News coverage, including how free-tier users are getting the weakest models elsewhere in the market.