Author: Daniel Barabander

Translated by: Deep Tide TechFlow

Deep Tide Guide: Three years ago, Cursor was a VS Code plugin running on the OpenAI API; today it has launched its own model, beating Claude Opus 4.6 on key benchmarks at just one-tenth the cost.

This article systematically addresses one of the most important strategic questions on the internet: when should one open an API, and when should one keep it closed? The conclusion serves as a warning for everyone working on platforms.

The full text is as follows:

Co-authored with Elijah Fox (@PossibltyResult).

In early March, Cursor released Composer 2—a proprietary programming model built on an open-source foundational model, which defeated Claude Opus 4.6 on key benchmarks at just one-tenth the price. Three years ago, Cursor was a fully operational VS Code branch running on the OpenAI API.

The journey of Cursor from a dependent client to a real competitor is a microcosm of the most important strategic question on the internet: when should a company open its capabilities via an API, and when should it remain closed?

We have developed a framework to answer this question, which relies on two things. First: does opening the API erode your moat? If so, can you find your moat elsewhere?

Whenever a company opens its intellectual property through an API, it faces the risk of eroding its moat through demand aggregation. In simple terms: competitors can use this intellectual property to guide their early-stage products, and once they have accumulated enough demand, they can cut off the API through vertical integration. Netflix did just that: it first licensed film and television content, and once it had a large enough user base to dilute enormous fixed costs, it produced "House of Cards."

But the truly dangerous scenario is when the output of the API can directly be used as input, compounding the quality of competing products. This poses a double threat because competitors can both use the API to guide and aggregate demand and directly improve their own production processes. This is exactly what is happening in the AI field. Although OpenAI and Anthropic explicitly prohibit companies that access their API from using the output to train competing models, they cannot prevent companies like Cursor from using cutting-edge models to guide workflows needed to collect proprietary product data and improve their own models over time.

This seems to be precisely what is happening behind Composer 2. Cursor has gathered enough demand using foundational models like Claude and GPT, achieving an annual revenue of about $2 billion, and then leveraged the open-source foundational model Kimi K2.5, along with data obtained through continuous pre-training and reinforcement learning from its IDE, to build a cutting-edge programming model.

When this output/input dynamic is present, API providers have only two choices: either shut down the API to stem the bleeding or remain open and find supplementary assets that leverage their own moat.

Twitter is a classic example of the first path. It was originally known for having a generous, freely accessible API—at its peak, developers could pull 500,000 tweets per month for free. However, Twitter closed most of its endpoints because the API revealed its moat: the proprietary social graph. Today, the API is effectively closed: access is strictly rate-limited, priced prohibitively at a meaningful scale, and building serious products requires tightly controlled B2B integration.

The second path is to keep the API open and supplement it with an alternative source of power. No industry understands this better than crypto—where APIs are mandated to be open, and the only way to survive is to find a moat elsewhere.

The lending protocol Morpho provides a representative case. This protocol was born from connecting to the open APIs of Aave and Compound to build optimizing products on top of them. It then used the output of these protocols—its aggregated liquidity—as the input to guide its own platform. Thus, it can be seen that Cursor and Morpho share a similar path in using APIs to build competitive products.

However, the truly interesting dynamic is what Morpho did next. Since Morpho itself is an open API, it needed to find a moat to compensate for the lack of switching costs. It thus decided to make the protocol as aggregate-friendly as possible, instead establishing a moat through other means—such as the Lindy effect and network effects generated by deep liquidity from diverse lenders and borrowers.

Using this framework to project forward, we can make a prediction: over time, foundational model companies are likely to choose the first path, gradually restricting API access to their cutting-edge models.

To believe in the second path, one must think that models like Opus and GPT are already powerful and trusted enough to remain open, allowing competing models to use their outputs as inputs, yet third parties will not leave. This means that model companies are betting on other sources of power: the Lindy effect (if they believe users do not want to build trust in new models), developer network effects (if they believe users will build tightly dependent ecosystems on their API openness), or economies of scale (if they believe maximizing API call volumes can dilute the fixed costs of training cutting-edge models).

However, current evidence points in the opposite direction. The dynamic of "the hottest models of the month" remains strong, with users migrating unabashedly to the best current model—as seen in the recent surge in Claude use following the release of Opus 4.5. At the model level, there are no clear signs of developer network effects either—interoperability between APIs is increasing rather than decreasing, and the surrounding tool ecosystem is actively combating lock-in, intentionally making vendor switching easy. Meanwhile, the current scale economy during the training phase is no longer sufficient as a moat, because distillation techniques enable competitors to train comparably performing models at much lower costs. Without alternative sources of power, foundational AI companies are likely to reserve limited access only for enthusiasts, focusing instead on B2B deployments with strict usage controls and monitoring. More and more often, the winning choice will be to refuse to play this game.

This is a concerning outcome because the explosion of current consumer-grade AI products is built on these model providers. It also opens a door to a reverse positioning: if leading labs increasingly restrict access, choosing competitors with weaker moats but a strong commitment to remain open has value to offer.

Thanks to @systematicls (@openforage) and @AlexanderLong (@Pluralis) for their thoughtful feedback on this article.