US Labs Are Giving the Model Away Now Too. OpenAI Is Paying With Ads. Anthropic Is Paying With the IPO.

In July we wrote that China's AI labs were giving the model away, and that it was a go-to-market, not a philosophy. We ended with the question worth carrying: which labs are willing to give theirs away, and what they expect to collect in return. This week the two best-funded US labs answered, with two different collection plans.

## What Anthropic gave away in three days

Claude Haiku 5.5 shipped on October 7 with a 1 million token context window and, by Anthropic's own figure, prices around 75 to 90 percent lower than Haiku 4.5 for most workloads. The Indian Express noted that it completes the Claude 5.5 family weeks before the company's expected IPO.

The day before, Anthropic expanded Claude for Startups: a free year of the Claude Team plan for up to five seats plus around $1,000 of API credits, live globally. The day after, the White House listed Anthropic's $150 million over three years of Claude access for Genesis Mission projects, inside $2.4 billion of compute tools and credits from eleven partners.

A price cut, a free year, a government grant in kind. Three ways of putting the model in more hands for less money, in three days.

## What it is being paid with

A draft prospectus seen by The Inference asks for a valuation of $1.8 trillion to $2 trillion and discloses at least $518 billion of long-term compute and equipment commitments. A Chinese outlet, citing TMTPost and Reuters, adds the figures the first report left out: about $4.6 billion of 2025 revenue, a $42 billion net loss, and a $100 billion raise.

Divide the two reported numbers and the commitments are more than a hundred times the revenue. That is the company that just cut its small-model price by up to 90 percent. The cut is not funded by margin. It is funded by the listing, and the listing is priced on volume. Every startup on a free seat and every agency on Genesis credits is a line in the growth chart the $2 trillion ask depends on.

## OpenAI chose the other collector

On October 5 OpenAI announced labeled display ads alongside image generation in ChatGPT, with US tests later this month and attribution and incrementality measurement through AppsFlyer, Triple Whale and DoubleVerify. Measurement partners are the tell: advertisers move budget on incrementality, and a lab building that pipe is building an ad business. Since August the ads have run in India on the Free and Go tiers, with paid plans ad-free. GPT-6 started rolling out to ChatGPT on October 7, paid plans first and then free tiers.

So the free model at OpenAI is paid for by the user's attention. The cheap model at Anthropic is paid for by the next shareholder. Both are the July play: make a good-enough model nearly free and commoditise what the other side charges for. In July we said we could not verify the Chinese labs' money. Here the money is in a prospectus.

## Who the cut is aimed at

In August we argued that cheaper inference does not automatically squeeze the frontier labs. It squeezes whoever sells last year's model at a lower price, the category we told you to prefer in a slowdown. A 75 to 90 percent cut on the small model is a frontier lab moving into the price band that Chinese open weights and API resellers occupied. That band got more crowded the same week: Mistral Large 4, a 1 trillion parameter mixture of experts with 49 billion active, will release its weights this month.

## Where the money that still wants a margin went

Arena raised a $200 million Series B at a $3.1 billion valuation, co-led by Lightspeed and Khosla, and launched an Alignment Index for how closely agents follow human intent. Nous Research raised $90 million at about $1.5 billion to turn an open-source agent into an enterprise assistant. Clockwork.io raised $31 million to stop GPU clusters wasting hours.

In August we wrote that generating the answer got cheap and testing it did not, and in September that a fifth of rounds were going to companies that police other AI. This week $200 million went to grading the model, in the week the model's price fell by three quarters.

## Hedges

The 75 to 90 percent figure is Anthropic's, for "most workloads", and a per-token cut says nothing about revenue if volume rises with the million-token context. The prospectus numbers are second hand, from a draft; the revenue and loss figures reach us through a Chinese outlet relaying TMTPost and Reuters, and none of it is in a public filing. The ads are a US test. A free year capped at five seats and $1,000 of credits is marketing, not margin.

## What to do with this

If you buy inference, reprice your stack this month. The frontier labs' small models now sit where you were using open weights or a Chinese API, so ask the question we set in August: cost per completed task. The headline moved 75 to 90 percent. Your task count did not.

When the filing is public, divide the compute commitments by the revenue. The reported figures are $518 billion and $4.6 billion, more than a hundred to one. If the public numbers hold that ratio, the price cut is a volume bet financed by the listing, and the question for anyone buying at $2 trillion is which of OpenAI's two collectors Anthropic adopts next. OpenAI, Anthropic, Google and Arena are on our tracker, and the two-exchanges piece from Wednesday has the line to divide by on the Chinese side.