Model

Harvey rebuilt on Moonshot's Kimi K3 and its margins turned positive again.

Usage fees stacked on base subscriptions pushed Harvey's gross margin from about 50 percent to minus 50 percent before an August rebuild on Moonshot's Kimi K3 fixed it. The cheapest compliant seats for that Chinese model are American inference hosts running chips export controls keep from China.

Vincent JiangVincent Jiangtwo-minute read
Share
Moonshot AI founder Yang Zhilin. Tingshu Wang/Reuters
1 / 6Slide 1 of 6
Moonshot AI founder Yang Zhilin. Tingshu Wang/Reuters

A $15.6 billion startup losing money on every query

Harvey$15.5B — Harvey, private, latest valuation $15.5B, the legal-AI company valued at $15.6 billion, watched its gross margin slide from about 50 percent at the start of the year to minus 50 percent by June, after a March update to its agents sent usage soaring, according to a person familiar with the numbers 1. Token consumption is up twentyfold this year, the company says 1. The rented models did the damage: OpenAI$1.18T — OpenAI, private, latest valuation $1.18T and Anthropic$350B — Anthropic, private, latest valuation $350B now bill enterprises for usage on top of base subscription fees, pricing that had already burned Uber's+0.34% — Uber, up 0.34 percent today full-year AI budget by April 1.

The escape route runs through Beijing

Harvey's August fix is a model post-trained on Kimi K3, the 2.8-trillion-parameter open-weight release from China's Moonshot AI, and its margins turned positive again 1. Abridge$6.69B — Abridge, private, latest valuation $6.69B now builds on Nvidia's−4.48% — Nvidia, down 4.48 percent today open weights, Decagon~$1.5B — Decagon, private, latest valuation ~$1.5B routes 80 percent of customer queries through its own models, and Ramp, which raised $750 million in June, is weighing training for the first time 1. The quiet collectors sit below the headlines. Modal$87M — Modal, private, latest valuation $87M, Fireworks AI~$1.1B — Fireworks AI, private, latest valuation ~$1.1B and Baseten~$1.4B — Baseten, private, latest valuation ~$1.4B host Kimi K3 at roughly $3 per million input tokens and $15 per million output, on Nvidia GB300 and AMD−5.50% — AMD, down 5.50 percent today MI350X hardware that export controls keep from Chinese firms, sold with zero-data-retention terms and US endpoints 2. No public account names Harvey's host; the layer collects whoever defects.

US-hosted Kimi K3 undercuts Anthropic's flagships on every token

  • Input
  • Output
$0/M tok$20/M tok$40/M tok$60/M tokClaude Fable 5.1Claude Opus 5Kimi K3, US-hosted$15/M tok
Data
InputOutput
Claude Fable 5.1$10/M tok$50/M tok
Claude Opus 5$5/M tok$25/M tok
Kimi K3, US-hosted$3/M tok$15/M tok
List API prices per million tokens, September 2026. Anthropic first-party list pricing; Kimi K3 shared-endpoint pricing across Modal, Fireworks AI and Baseten.2,3

"Cosplay," until it saves the margin

Menlo Ventures partner Matt Kraning calls most custom-model talk "a lot of cosplay," and Salespeak abandoned its own model-building effort after seeing no major advantage 1. US security agencies accuse Moonshot of distilling American models, and Rogo~$1B — Rogo, private, latest valuation ~$1B has had to walk large customers past their hesitation about Chinese weights 1. Even Anthropic's own investor-forum slide concedes Harvey still needs Opus for its hardest tasks 1. The frontier stays rented; the volume work leaks.

Cut off the week of the summit

OpenAI will end SpaceX−1.87% — SpaceX, down 1.87 percent today-owned Cursor's model access on 12 November, citing past contract violations by Elon Musk's companies, weeks after investors floated a $1.2 trillion valuation ahead of an IPO expected next year 45. Nothing argues for hedging harder than a supplier who can quit. Xi Jinping meets Donald Trump in Washington this week with artificial intelligence on the agenda and Sam Altman and Jensen Huang at the state dinner 6. Watch whether the compliant American wrapper around Beijing's biggest open model survives the summit.

Get Model by email, free

The AI labs, their models, and the business behind them. In our daily email, every weekday.

Loading UBER

Reader comments

0 comments

    Model

    Meta's Q3 earnings fell to $3.83 against $5.46 expected, on a $9.89 billion teen-harm legal charge

    Excluding the charge Meta beat, and Businessweek's settlement pricing implies only about half the pending teen-harm claims are covered, with 2,000 school districts unpriced. The reserve note in January's filing decides.

    Model

    Nvidia shows how to run Z.AI's GLM-5.3 on $190,000 of Nvidia hardware, with no per-token bill to Z.AI

    The tutorial walks enterprises through serving GLM-5.3 on two DGX Stations, roughly $190,000 at listed reseller prices, with no per-token bill to Z.AI. For a stock carrying 118.6 times sales, that reads as a price list for everything Z.AI sells.

    Model

    Anthropic's Claude models tie at 58 on a coding index Cognition partly built, days before its Nasdaq IPO

    A third of the refreshed index runs on FrontierCode, a benchmark built by Cognition, whose Devin agent announced one of the two tied winners, Claude Fable 5.2, on the day of its release. Anthropic's model lists on Nasdaq on November 19; Cognition's Devin has offers from $70 billion on the table.

    Model

    Anthropic's Fable 5.2 passes 41.3% of real freelance jobs before its Nasdaq debut

    The one independent test of finished freelance work puts Anthropic's Fable 5.2 first at 41.3%, and the model still fails most of it, days before a Nasdaq debut priced at up to $1.4 trillion.

    Model

    Hedra cut Krea 3's edit time from 22.9 to 3.3 seconds on AMD, and the app doesn't own the speed

    The eight-second edit Krea advertises is manufactured in the serving stack, and a partner's measurement shows the app does not own the speed. Inference efficiency accrues to the optimizer and the chip vendor, which caps a per-image app's margin and hands AMD a demand channel beyond hyperscalers.

    Model

    xAI ships AI teammates at $30-$200 a month, ten days after Musk said serving them loses money

    The published card sets a chatbot price against an always-on machine the CEO says loses money on every subscriber; SpaceX shareholders fund both sides of that trade in one stock.

    Subscribe to TheInference for free