Published · 2h agoLeadership2 min read
Anthropic and OpenAI launch cheaper models as startups move to open weights
Anthropic and OpenAI have released lower-cost models as startups such as Harvey, whose gross margin hit -50% by June, shift work to cheaper open models. Their low tiers now compete with startups training their own models, the route that returned Harvey to positive margins.
Context for builders, not their beat.See today for builders
What happened
- Anthropic and OpenAI have both released lower-cost models as they face growing competition from cheaper Chinese models.
- Harvey's gross margin fell from about 50% at the start of the year to -50% by June after a March agent update lifted usage, according to a person familiar with the matter.
- An August Harvey model built on Moonshot AI's Kimi K3, plus other changes to its AI usage, returned its gross margins to positive, people familiar with the efforts said.
- Abridge is building a clinical foundation model trained on Nvidia's open models, one of several startups building their own models.
Compiled by The Board RoomSomething wrong?How this is made
Why it matters
- exposure Both labs are absorbing this pressure while preparing for IPOs, so each startup that leaves takes revenue out of the figure prospective investors will price.
- constraint A lower price addresses cost only; startups using open models as a hedge against frontier models being slowed or restricted still carry that risk at any discount.
- precedent Harvey's recovery on Kimi K3 gives other application companies a worked example of switching, and any lab discount now has to beat that result.
The coverage of these launches does not state per-token prices, so the low tier's input price can only be judged against what a customer would do instead. Bloomberg reports that Moonshot AI's Kimi K3 can perform close to Anthropic's best offerings at a fraction of the cost [4]. According to the same report, open-weight models are typically cheaper than proprietary US models and let a firm train custom models on its own data [9].
Harvey's numbers show how large the cost pressure is. The legal startup, valued at $15.6 billion, saw its gross margin move 100 percentage points between the start of the year and June, and the fall came as customer usage rose [7][1][3]. At -50%, it was spending about $1.50 to serve each $1 of revenue [3].
Timing decides how much traffic a cheaper tier can reach. Decagon said it already routes 80% of queries through its own models, so a lower Anthropic or OpenAI price is competing for the remaining 20% [6][2]. A startup that trains its own model this quarter will weigh any later lab price against work it has already paid for. Bloomberg reports that Ramp and Rogo are exploring training their own models for the first time, so for them the new prices arrive before that decision is made [12]. The same report says investors including Sequoia Capital and General Catalyst are backing the move to open models [11].
What to watch
- Published per-token rates for the new Anthropic and OpenAI low tiers, set beside hosted Kimi K3 pricing, would show whether the cost gap Bloomberg describes has narrowed.
- Any disclosure from Harvey or Decagon that traffic has moved back onto proprietary lab models.
- IPO filings from either lab that break out revenue from startup customers.
Claim ledger
Ranked by verification strength, evidence, and original report placement.
- [1]
Anthropic and OpenAI have both released lower-cost models.
- [2]
Anthropic and OpenAI face growing competition from cheaper Chinese models.
- [3]
After a March update to its AI agents, Harvey's customer usage spiked, but its gross margins dropped sharply, from about 50% at the beginning of the year to -50% by June.
- [4]
Harvey released a model of its own in August, powered by China-based Moonshot AI's Kimi K3, which can perform close to Anthropic's best offerings at a fraction of the cost.
- [5]
Harvey's Kimi K3-based model launch, plus other tweaks to its AI usage, made Harvey's gross margins positive again; the company declined to comment on specific financials.
- [6]
AI customer support startup Decagon said it now flows 80% of queries through its own models.
- [8]
The startup push toward open-weight AI risks cutting into OpenAI's and Anthropic's revenue as both gear up for highly anticipated initial public offerings.
- [9]
Open-weight offerings are typically cheaper than proprietary technology from US AI developers and let firms create custom models with their own data.
- [11]
Investors like Sequoia Capital and General Catalyst are backing the trend toward open-weight AI.
- [12]
Fintech startups including Ramp and Rogo are exploring training their own models for the first time.
- [13]
For some startups, open-weight AI is becoming a hedge against a future in which the most advanced models could be slowed or restricted.
- [14]
Healthtech startup Abridge recently announced it was building a custom foundation model for clinical settings, trained on Nvidia's open models.
- [d1]
Harvey's gross margin moved about 100 percentage points between the start of the year and June.
Derived - [d2]
About 20% of Decagon's queries still run outside its own models.
Derived - [d3]
At a -50% gross margin, Harvey was spending about $1.50 to serve each $1 of revenue.
Derived
Sources & coverage · 2 publishers
The reporting this story was synthesized from, earliest first. Every link goes to the original.
- finance.yahoo.com3d agoOpenAI, Anthropic Costs Push More Startups to Build Off Cheaper Open Models
- technologyreview.comyesterdayThe Download: India's smart glasses menace and AI's trillion-dollar gamble
Additional citations
- MIT Technology Review, The Download, citing CNBC
- Bloomberg, according to a person familiar with the matter
- Bloomberg
- Bloomberg, according to people familiar with the efforts
- Bloomberg, citing Decagon