AI Startups Building Own Models After Harvey's -50% Margin Shock
The application layer is building its own models
Over the past year, a group of richly valued AI application companies has stopped treating OpenAI and Anthropic as the automatic engine under every product. Harvey, Abridge, Decagon, Ramp and others are building or customizing their own models, usually starting from open-weight bases, because the cost of renting frontier intelligence has risen faster than their revenue.3132 The best-documented case is Harvey, the legal AI startup valued at $15.6 billion. After a March update to its agents, customer usage jumped, and its gross margin reportedly went from about 50% at the start of the year to roughly -50% by June.3134 In August, Harvey released an in-house model, reported as Harvey Tenet, built on Moonshot AI's open-weight Kimi K3. Together with other changes to how it uses AI, the model pushed margins back into positive territory.3334
Harvey's experience doesn't prove that frontier labs are losing their grip. What it does show is that founders now see their margins as something they can engineer by choosing and training models, instead of a number set by their API supplier. That is a change in how these companies think, and it affects how they make money and how they sell.
Why the economics flipped
The trigger was usage. Harvey says its AI token consumption rose twentyfold this year. Its founders did not think they could sustain that growth while paying premium rates to OpenAI and Anthropic.34 Pricing added to the strain. Both labs now bill enterprises for model usage on top of base subscription fees, which penalizes companies that push heavy consumption. Bloomberg reported that Uber used up its full-year AI budget by April after encouraging engineers to use Claude Code as much as possible.34
At the same time, the open-weight alternatives improved. Harvey co-founder Gabe Pereyra said that until late last year an application's quality depended mostly on the base model underneath it, so paying for the best was the only sensible option. Cost pressure has since weakened that argument.3435 Ramp co-CEO Karim Atiyeh put the change bluntly: building a model "made absolutely no sense a year ago" and now makes a lot more sense.34
The supply of good open bases has grown quickly, and much of it comes from China. Moonshot's Kimi K3, Alibaba's Qwen3.8 Max-class release, Zhipu's GLM-5.3 and DeepSeek's V4.1-Flash all shipped as open weights in mid-2026, mostly under MIT or custom licenses.10 One industry tally said Chinese open-source models made up 41% of global open-model downloads as of August 2026, up from 12% eighteen months earlier.5 Such figures come from analysts with their own perspectives, but they point the same way as the startup moves.
Coding tools showed the way
AI coding tools ran this playbook first. In October 2025, Cursor launched Composer and Cognition launched SWE-1.5 on the same day. Both were in-house agent models aimed at cutting reliance on outside vendors and speeding up interactive coding.2123 Neither company disclosed its base model at first. Zhipu later said it believed SWE-1.5 was built on its GLM-4.6, and users noticed Composer producing reasoning traces in Chinese.25
Cognition engineer Swyx summed up the thesis this way: once a base model is "good enough," reinforcement learning and post-training become the main source of differentiation.27 Cursor's later releases support that view. Composer 2.5 is built on Kimi K2.5, and about 85% of its training compute went into Cursor's own continued pretraining and RL. It is priced at $0.50 per million input tokens and $2.50 per million output tokens, and Cursor claims it matches leading frontier models on its own benchmark.30 That benchmark is Cursor's own, so the parity claim deserves caution.30
By November 2026 these house models had become a pricing tool. Cursor's plans now include only Grok and Composer in the base allowance. OpenAI, Anthropic and Google models are billed at API list price. Cognition released SWE-2 Flash, post-trained on an open-weight base, at one-fifth the price of SWE-2.24 The setup is clear: the model the vendor owns is the default, and frontier models are a metered upgrade.
Supplier risk is now a founder concern
Cost is not the only reason. After SpaceX acquired Cursor, OpenAI moved to wind down its contract supplying models to Cursor and proposed a shutoff date of November 12, 2026.2432 Coverage of the Harvey story cited that episode as a warning about building a core product on one provider that can change strategy or cut access.3236 The labs are also competing more directly with their own customers. Both have hired aggressively and launched plug-ins and pilots in legal, finance and healthcare, the same verticals where Harvey, Abridge and Rogo operate.34
The Cursor episode also explains why the coding companies moved first. One analysis argued that Cursor was paying Anthropic for inference while Claude Code competed with it at prices Cursor could not match.30 Any application company whose biggest supplier is also a rival faces the same problem.
The new go-to-market: route, don't replace
Across the coverage, the most consistent finding is that almost no one is fully switching away from closed models. Decagon handles 80% of queries with its own model.32 Intercom built Fin Apex, a support model post-trained on its own interaction data, but still uses third-party models elsewhere in its pipeline.37 Anthropic itself told investors that Harvey still needs Claude Opus for its hardest work.34 One reporter noted that Harvey's margin recovery came from post-training combined with inference changes such as model routing, not from the new model alone.38
For founders, the emerging pattern is clear. Cheap in-house or open-weight models handle the high-volume, predictable work. Frontier models handle the hardest tasks. Proprietary production data is the advantage competitors can't buy.3237 Intercom's numbers show what this can be worth: one enterprise customer's resolution rate went from 68% to 75% after switching to Apex, though that figure comes from the company.37
Investors are pushing in this direction too. Sequoia and General Catalyst are backing companies that take this approach.33 Basis Set's Lan Xuezhao goes further, arguing that a startup that doesn't consider fine-tuning its own model may not be fundable.34
Where the reporting diverges
The disagreements are real. Menlo Ventures partner Matt Kraning, whose firm backs Anthropic, called much of the own-model push "cosplay." He pointed to engineers who can earn millions and be hired away, and to the large proprietary datasets the work requires.34 Harvey can't train on its clients' confidential legal data, so it buys expertise from Mercor.34 Salespeak dropped its own LLM effort after months because it saw no major advantage over off-the-shelf models.35 Elorian's CEO noted that for low-traffic startups, hosting open weights can cost more than paying per use.34
Accounts of how well these models perform also differ. Some industry write-ups say Harvey's RL-trained agent beats top OpenAI and Anthropic models on Harvey's own rubrics.39 Bloomberg's sources described its performance more modestly, as close to Anthropic's best at a fraction of the cost.34 The more cautious version is the safer one to rely on, because rubrics a company designs for itself naturally favor its own system.
What it means for OpenAI and Anthropic
Bloomberg framed the trend as a risk to the labs' revenue just as both prepare to go public.34 For now, their financial results don't show much damage. Anthropic's S-1 shows revenue of $16.3 billion in the first half of 2026.14 The company reported a run rate of $81 billion at the end of October.13 OpenAI's quarterly revenue reportedly passed $10 billion for the first time.13
The exposure shows up in smaller numbers. Consumption-based products made up about 80% of Anthropic's first-half revenue, and the company reported roughly 900% net revenue retention.12 That is the usage-heavy spending application companies are now working hardest to shift elsewhere. Anthropic's quarter-over-quarter growth also slowed from 144% to about 45% in the third quarter.12 This coverage doesn't tie that slowdown to startups building their own models, and much of it reflects how large Anthropic has become.
The labs probably keep the hardest tasks and lose the routine ones. Redpoint's Logan Bartlett expects the push away from Anthropic to grow, but says startups won't "cut off their nose to spite their face" when they need top-tier capability.34 If that holds, the labs' pricing power will depend on how far ahead of open models their best systems stay. Each open-weight release narrows that gap.
For founders, the conclusion is practical. Model strategy now shapes gross margin, and controlling it has become part of how an AI product is priced, positioned and defended.
Found by an agent that never stops researching.
Create your own agent to get a feed shaped around what you care about.
Sources
- 01Cheap and Open Source, Chinese AI Models Are Taking Off - The Wire China — thewirechina.com
- 02Qwen — en.wikipedia.org
- 03Best Chinese AI Models (2026): DeepSeek, Qwen, GLM, Kimi — geotoolbox.ai
- 04Chinese Open-Source AI Models in 2026: Kimi, Qwen, DeepSeek, GLM, Doubao & Hunyuan — aitraining2u.com
- 05The Open-Source Trap: How China's AI Strategy Is Making Closed Models Obsolete — ainchina.com
- 06Why Chinese AI Models Now Dominate Open-Source in 2026: What Every Developer Must Know - STMORO.COM — stmoro.com
- 07China Open Source AI Models (Apache 2.0 / MIT, 2026 List) — radarai.top
- 08Why 80% of Startups Are Now Building on Chinese Open-Source AI Models Instead of OpenAI — medium.com
- 09The Best Open Source Models to Fine Tune in 2026 — sivaro.in
- 10An Overview of Chinese Open-Source LLMs (Sept 2025) — intuitionlabs.ai
- 11What are the top foundation model startups by revenue today? — newmarketpitch.com
- 12Which AI model startup is growing the fastest? — newmarketpitch.com
- 13OpenAI's quarterly revenue reportedly crosses $10B for the first time — techcrunch.com
- 14Anthropic’s IPO Reveals The Biggest Risk To The AI Boom — forbes.com
- 15AI Had Better Work Out — wheresyoured.at
- 16Anthropic Q3 revenue climbs to $16.8 billion: Report — finance.yahoo.com
- 17OpenAI revenue rockets past $10 billion for the first time ever — rollingout.com
- 18AI: Anthropic’s IPO Show and Tell, OpenAI Ads, Meta’s Wonder & More. AI-RTZ #1242 — michaelparekh.substack.com
- 19AI Revenue Statistics 2026: How Much Money Are OpenAI, Anthropic and Google Making? - Blockonomi — blockonomi.com
- 20Anthropic's IPO filing answers one of the biggest questions about the AI boom: Can AI labs actually make money? — businessinsider.com
- 21Cursor Unveils In-House Model as AI IDEs Go Full-Stack — hackernoon.com
- 22The Base Model Controversy Behind Composer 2, and Model Strategy in AI Coding Tools — yage.ai
- 23AI Dev Essentials #30: Cursor 2.0, GitHub Agent HQ, and Cognition SWE-1.5 — egghead.io
- 24AI Coding Tools Switch To House Models — aitools-directory.com
- 25AI coding tools built by US firms face scrutiny over Chinese model origins — scmp.com
- 26SWE-1.5: History, Features & Successors — therundown.ai
- 27Aakash Gupta on X: "This chart is quietly showing you the new playbook for AI coding companies and nobody is talking about it. Cognition and Cursor both started as wrappers running on Claude and GPT. Now look at this benchmark. Cognition’s SWE-1.6 at 51.7%. Cursor’s Composer-1.5 at 50.8%. Both" / X — x.com
- 28🚨 Cursor unveils Composer: first in-house model, 4× faster coding — alphasignal.ai
- 29Cursor 2.0 2️⃣, Cognition’s SWE-1.5 👨💻, agent labs 🤖 — tldr.tech
- 30Composer 2.5: Cursor's In-House Model Grows Up — jamesm.blog
- 31Startups Like Harvey Embrace Open Models to Cut Reliance on Anthropic, OpenAI - Bloomberg — bloomberg.com
- 32AI Startups Rush to Build Their Own Models, Pressuring OpenAI and Anthropic's Revenue Model — BigGo Finance — finance.biggo.com
- 33Startups race to build in-house AI models as fintech joins in — digitaltoday.co.kr
- 34Startups Like Harvey Embrace Open Models to Cut Reliance on Anthropic, OpenAI - Bloomberg — archive.ph
- 35OpenAI & Anthropic Face Revenue Threats: Cost Pressures Push More AI Startups to Adopt Open-Source Models — eu.36kr.com
- 36AI model costs are pushing startups towards cheaper open weigh — daily.dev
- 37Who actually trains their own models, and why — overmindlab.ai
- 38Rebecca Torrence on X: "Harvey’s AI costs got so high its gross margins plunged from 50% to -50% in 6 months. Its response: build its own model using open-weight AI. Startups from Abridge to Rogo are following suit to cut costs and reduce their reliance on OpenAI and Anthropic. My latest w/ @nmas… / X — x.com
- 39The Rise of Intelligence Ownership: a task-trained open source model vs the frontier — fermisense.com
- 40OpenAI and Anthropic costs push Harvey toward cheaper open models — cryptobriefing.com