Harvey Co-founder Gabe Pereyra: Turned Gross Margin from -50% to Positive in One Quarter
Gabe Pereyra, co-founder and president of legal AI company Harvey, stated that with user consumption doubling month by month and the continued provision of the strongest models, Harvey turned its gross margin from -50% to positive within a single quarter. He mentioned that the company did not force customers to switch to usage-based billing early to protect profits, nor did it cut costs by using inferior models, but instead advanced the billing model transformation at a pace acceptable to customers, even if it harmed its own profits.
According to Pereyra, the turnaround was achieved through product-level optimizations: routing different tasks to different models, improving the operational framework (harness), and reducing inference costs through post-training to provide cutting-edge intelligence at lower prices. Meanwhile, Harvey built cost control tools for clients, including usage dashboards, case-based cost aggregation, spending limits, and return on investment (ROI) reports. He stated that the company's philosophy is to do the best for customers, even if the process is painful, harms profits, or invites criticism from competitors, social media, and the press.
This statement responds to external doubts about Harvey's gross margin. Reports indicate that Harvey's gross margin was about 50% in early 2026 but fell to -50% by June. The decline was triggered by an AI update launched in March: the legal AI agent can continuously perform multi-step tasks, consuming far more tokens than ordinary Q&A, leading to a 20-fold increase in annual token usage. Meanwhile, Harvey rents models from OpenAI and Anthropic and pays per token, while revenue only doubled during the same period, increasing from about $190 million to over $400 million, with cost growth far outpacing revenue growth.
A key step in reversing the gross margin was changing the underlying model. In August, Harvey launched its own model "Harvey Tenet," based on the open-source weight model Kimi K3 released by Moonshot AI in July, and completed post-training in collaboration with the inference platform Fireworks AI. Harvey's engineering blog stated that the token price of the open-source weight model is lower: Kimi K3's token price is about 40% lower than Claude Opus 5, and the cost per query has decreased by 90%, reducing the cost of completing a task from $52 when using Opus to $30. Harvey's previous products were mainly customized based on closed-source models from Anthropic, OpenAI, and Google, and the OpenAI startup fund was one of Harvey's earliest investors.
The capital market has not cooled due to fluctuations in gross margin. On September 9, Harvey completed a $550 million financing round led by Lightspeed Venture Partners and Diffusion, with a valuation of about $15.5 billion, up about 41% from the $11 billion valuation in March; Sapphire Ventures and Whale Rock participated for the first time, while Sequoia, Kleiner Perkins, Andreessen Horowitz, Coatue, Goldman Sachs Alternatives, and Singapore's sovereign fund GIC continued to invest. Harvey's annual recurring revenue (ARR) reached $400 million, quadrupling from $100 million in August 2025 in just 13 months; the company serves 80% of the top 100 law firms in the U.S. (Am Law 100), five of the top 10 Fortune 500 companies, and over 3,000 institutions across 60 countries.
Market mechanisms indicate that this is a supply chain shift driven by a cost structure crisis, not a single event. The sellers are Anthropic and OpenAI: Harvey has shifted most of its inference demand away from their closed-source models, resulting in reduced token revenue from high-usage application clients. The buyers are application-layer companies like Harvey: they are turning to open-source weight models, combined with self-trained models, to lock in costs. In terms of capital flow, inference budgets are shifting from closed-source frontier labs to open-source model developers and inference hosting platforms. Beneficiaries include open-source and inference infrastructure companies like Moonshot AI and Fireworks AI, as well as law firm clients who can obtain lower prices and cost transparency; the pressured parties are closed-source model vendors charging per token, and AI application companies that lack model optimization capabilities and still rely entirely on renting cutting-edge models. Companies like Abridge, Decagon, and Ramp are also reportedly shifting to open-source weight models.
Source: Public Information
ABAB AI Insight
Harvey was founded in 2022 by former O'Melveny & Myers litigation attorney Winston Weinberg and former Google DeepMind and Meta AI researcher Gabe Pereyra. The OpenAI startup fund was an early investor, allowing Harvey early access to GPT-4, which became a key advantage in its entry into the legal AI space. In February 2023, Harvey partnered with the UK law firm Allen & Overy to open its product to over 3,500 lawyers, and later collaborated with PwC, opening up the market for large law firms and the Big Four accounting firms. This path has allowed Harvey to reach an ARR of $400 million in just over three years, but it has also created risks: both product capabilities and cost structures are highly dependent on a few closed-source model suppliers.
In terms of capital strategy, Harvey is transitioning from being a "packaging layer for models" to becoming a "vertical platform with its own models." The first step is to exchange high-valuation financing for cash and time: raising funds at a $110 billion valuation in March and then at a $155 billion valuation in September, providing a buffer during the period of negative gross margin. The second step is to build Harvey Tenet using an open-source foundation and post-training, shifting the largest cost item from "renting" to "owning." The third step is to transform the billing model: moving from seat-based charges to usage-based charges, providing case-based cost aggregation and spending limits, transparently passing token costs to clients while reducing their resistance to "usage-based billing."
A direct analogy is the AI programming tool Cursor. From June to July 2025, Cursor changed its pricing to usage-based due to soaring costs of cutting-edge models, leading to strong user dissatisfaction; the company subsequently publicly apologized and refunded users. In October, Cursor launched its self-developed programming model Composer to reduce reliance on model suppliers like Anthropic. The competition in the legal AI space is also intensifying: Thomson Reuters acquired Casetext for $650 million in 2023 and launched CoCounsel; Sweden's Legora is rapidly rising and directly competing with Harvey. Currently, vertical AI applications are transitioning from an "expansion phase relying on cutting-edge models" to a "self-built model cost control" phase, where those who can turn model costs from external rents into internal assets will be able to maintain gross margins during scale expansion.
Structural judgment: This represents a transfer of pricing power, with the pricing power of inference costs shifting from closed-source frontier labs to application-layer companies. The mechanism is that intelligent agents cause the token consumption of individual tasks to increase exponentially, while application companies charge clients based on seats and pay model vendors based on tokens; the greater the usage, the greater the losses. At this point, frontier labs hold complete pricing power. The emergence of open-source weight models breaks this dependency: models like Kimi K3 are gradually approaching the capabilities of closed-source frontier levels but at much lower prices, allowing application companies to conduct self-training and choose inference platforms, reducing unit task costs by 40% to 90%. Once application companies can freely switch and route between multiple models, the pricing power of closed-source model vendors will be weakened, while application companies will build new moats through "owning models + client data + workflows." Harvey's turnaround to a positive gross margin in a single quarter is a landmark case of this power shift.
ABAB News · Cognitive Law
The more successful the usage, the more expensive the rented intelligence.
The moat of application companies starts with owning clients and ends with owning models.
When open-source catches up to closed-source, pricing power slips away from those selling shovels.
There are several figures and statements that suggest rechecking:
Financing figures: The valuation in this round varies by source, with both $15.5 billion and $15.6 billion mentioned; the text uses $15.5 billion.
Single-source cost data: The claims that Kimi K3 is about 40% cheaper than Claude Opus 5, that the cost per query has decreased by 90%, and that the cost of a single task has dropped from $52 to $30 come from only one secondary source. Additionally, Gabe Pereyra's original post only mentioned routing, operational framework improvements, and post-training, without naming Kimi K3; the point about "changing to Kimi K3 as the foundation" comes from reports by Bloomberg and other media.
Content written based on background knowledge without verification: The AI interpretation includes the backgrounds of the two founders, the collaborations with Allen & Overy and PwC, the pricing turmoil of Cursor and the Composer model, the amount Thomson Reuters paid for Casetext, and the situation of Legora.