OpenAI and Anthropic Face Revenue Threats: More AI Startups Are Turning to Open-Source Models Under Cost Pressure

Legal AI unicorn Harvey once saw its gross margin plummet to -50% this year, which serves as a microcosm — as API fees from OpenAI and Anthropic keep rising, a growing number of AI application startups are embracing open-weight models to regain control over costs.

This trend is accelerating. Startups spanning legal, healthcare, financial and customer service tracks including Harvey, Abridge, Decagon and Ramp have successively announced plans to develop or customize their own models, and some enterprises have migrated 80% of their traffic to self-owned models.

Leading investment institutions such as Sequoia Capital and General Catalyst are also fueling this trend behind the scenes.

This move poses direct pressure on OpenAI and Anthropic. The two companies are each preparing for their highly anticipated IPOs, and the outflow of application-layer customers could erode their most important revenue source.

01

Gross Margin Plunge Forces Harvey to Restructure Its Model Strategy

Harvey's situation best illustrates the problem.

This legal AI startup with a valuation of 15.6 billion US dollars has long used OpenAI's GPT-4 as the core of its products. According to Bloomberg, after the company's AI agent completed a major update in March this year, user usage surged sharply, followed by a cliff-like drop in gross margin — from around 50% at the beginning of the year to -50% in June.

Harvey's AI token usage has increased twentyfold this year.

Gabe Pereyra, co-founder and President of Harvey, said, "Previously, the performance of AI applications was highly dependent on the capabilities of underlying models, so paying for the best models from OpenAI and Anthropic was a necessary choice, and the significance of self-training models was relatively limited. But this logic has started to loosen under the heavy pressure of costs."

In August this year, Harvey released its self-owned model, which is supported by Kimi K3 from Moonshot AI. The performance of this model is close to the best product under Anthropic, but the cost is only a fraction of it. Combined with adjustments to other AI usage strategies, Harvey's gross margin has turned positive again.

02

The Wave of Using Open-Source Models Sweeps Multiple Vertical Tracks

Harvey's transformation is not an isolated case, and this trend has spread across multiple industries.

Medical technology startup Abridge recently announced that it will build a foundational model for clinical scenarios based on NVIDIA's open-source model.

AI customer service startup Decagon said that 80% of its query requests are now processed by its self-owned model.

In the fintech sector, Ramp and Rogo are exploring the possibility of training their own models for the first time.

In the programming tool sector, Cursor, which has been acquired by SpaceX, and Cognition, with a valuation of 48 billion US dollars, are among the earliest AI application companies to release customized models.

Karim Atiyeh, co-CEO of Ramp, said that previously, training a dedicated model was completely uneconomical, but after the company completed a $750 million financing in June this year and the performance of open-weight systems has improved significantly, this logic is reversing.

Dr. Lan Xuezhao, founder and Managing Partner of San Francisco-based venture capital firm Basis Set, took a tougher stance on this:

"If you don't optimize costs and fine-tune your own model, you are by definition inefficient. If a company doesn't consider building its own model, it may not be able to get financing at all."

03

Dual Pressure from Closed-Source Giants: Price Hikes and Invasion

What drives this transformation is not just the cost itself, but also the increasingly aggressive business strategies of OpenAI and Anthropic.

Both companies have recently shifted to charging enterprise users extra for model usage, which, combined with the previous subscription base fee, directly punishes the "token-maximizing" usage pattern. It is reported that after Uber encouraged its engineers to maximize the use of Anthropic's Claude Code, it exhausted its annual AI budget as early as April.

At the same time, OpenAI and Anthropic have both been actively recruiting talent in the core tracks of startups including legal, financial and healthcare sectors this year, launching plugins and piloting industry applications, which directly forms a competitive relationship with their customer groups.

Startups' dependence on model vendors also brings another risk: being cut off from access. Less than a week after SpaceX completed the acquisition of Cursor, OpenAI announced that it would suspend model access to this programming tool startup, on the grounds that Elon Musk's companies had violated the terms of service in the past.

It is against this backdrop that an investor forum held at Anthropic's office specifically discussed the trend of startups building their own models. Anthropic stated that Harvey still needs to rely on Anthropic's most powerful Claude Opus model to handle the most complex tasks.

04

Practical Challenges on the Open-Source Path

Shifting to open-weight models does not come without costs, and this path is also full of obstacles.

Talent is the primary bottleneck. Matt Kraning, Partner at Menlo Ventures, an investor of Anthropic, pointed out that professional engineers qualified for model fine-tuning work can get salaries of millions of US dollars, and they are extremely easy to be poached by large organizations such as OpenAI and Anthropic.

Data is another threshold. For enterprises to self-train models, a large amount of proprietary data is required. Since Harvey cannot access customers' sensitive legal documents, it has to purchase training data from AI data provider Mercor.

The infrastructure cost of open-source models cannot be ignored either. Andrew Dai, CEO of computer vision AI startup Elorian, pointed out, "The expense of downloading open-weight models and managing computing infrastructure on your own is quite considerable. For early-stage companies with low traffic, paying for closed-source models on a pay-as-you-go basis may be more cost-effective instead."

In addition, there are indeed enterprises that failed after trying. Startup Salespeak previously announced a plan to build a large language model on its own, but abandoned this direction after several months of exploration. Its co-founder and CEO Omer Gotlieb said that compared with the off-the-shelf models of Anthropic and OpenAI, it failed to see "significant advantages".

05

Dependence Remains, Game Continues

Despite the surging wave of self-development, most startups do not expect to completely get rid of OpenAI and Anthropic.

Logan Bartlett, Managing Director of Redpoint Ventures, who holds investments in Anthropic, Abridge and Ramp at the same time, predicts that the industry's motivation to reduce dependence on Anthropic's models will continue to grow, but he admits that startups will still pay for the best performance when necessary. "They won't cut off their own limbs just out of pique," he said.

Anthropic's presentation materials also confirm this reality — Harvey still needs Claude Opus to handle the most complex tasks, which shows that at this stage, the mixed usage strategy of open-source and closed-source models may be the optimal solution for most startups.

This game around AI cost and control is essentially a process where application-layer startups seek to renegotiate prices with model-layer giants. As the performance of open-source models continues to catch up with closed-source products, this tension will continuously test the interest boundaries of both sides as OpenAI and Anthropic move towards their IPOs.

This article is from the WeChat official account "Wall Street News Max", author: Yang Chen, published with authorization from 36Kr.