Open Models Shift Power to IDEs

Open weight coding models shift power from the lab to the application layer. When a product like Cursor can build its own coding workflow on Moonshot models and serve Kimi K3 through outside inference partners, it can choose the cheapest serving path, tune behavior for its own editing loop, and keep billing, usage data, and developer trust inside Cursor instead of handing them to OpenAI through either the API or Codex.1 2 3 4

• This changes who owns the user. OpenAI sells raw model access through its API and also sells Codex as a full product, but Cursor packages model access inside its IDE, so the developer opens Cursor, accepts Cursor pricing, and forms habits around Cursor workflows rather than around OpenAI directly.3 4 5
• Open weights also change the cost equation. Moonshot distributes Kimi through providers such as Fireworks AI, Together AI, and AWS Bedrock, which lets application companies shop for infrastructure, avoid a single lab’s markup, and swap providers without rebuilding the customer facing product.1 6 7
• The broader pattern is specialization. Cognition has already built SWE-2 by post training Kimi K3 for software engineering, showing how coding products can start with an open model, add their own reinforcement learning and tool use, and turn the foundation model into a workflow specific engine.8 1

The next step is a coding market where frontier labs supply ingredients, while IDEs and agent products capture most of the workflow value. That pushes OpenAI to keep improving both model quality and product depth, because winning the base model alone no longer guarantees winning developer spend or developer mindshare.3 4 5