Cursor lost OpenAI's models, and Kilo is poaching its customers for free

The divorce

The most important product event in coding assistants this month was not a launch — it was a breakup. On October 12, Cursor moved its cheapest plan, the $8 Plus tier, to Grok-only models, Kilo reports, and on November 12 OpenAI will stop providing its models to Cursor entirely. Requests to third-party models in Cursor's Premium, Super, and Ultra plans are now billed at API list price, and after the cutoff, personal OpenAI API keys in Cursor will work only for local Chat and Agent. For a product whose value was an opinionated bundle of the best models in a good editor, the bundle just lost its anchor tenant. m:…6782

The poacher's playbook

Kilo AI moved within days. It is offering free enterprise access for up to one year to companies currently paying for Cursor, positioning its platform as the answer to 'model freedom' — access to over 500 closed and open-weight models through the Kilo Gateway, with Auto-Efficient routing that sends prompts to the cheapest model proven capable of the task. Kilo's pitch leans on usage data showing developers reach for three or more model providers up to 70% of the time when balancing cost and performance, and it claims Anaconda's engineers have already migrated from Cursor. The offer is a bet that model-switching pain is now lower than vendor-lock pain — and that a contract year is long enough for the habit to stick. m:…6782

The open-weight counterweight

Meanwhile the supply side is diversifying fast. Baseten announced day-0 support for Reflection's Beam, a 501-billion-parameter mixture-of-experts model with only 23 billion active parameters per token and a 1-million-token context window, trained with reinforcement learning on longer rollouts and a length penalty. Beam scored 80.9% on SWE-bench Verified, 90.5% on GPQA Diamond, and 78.7% on the MCP Atlas benchmark. The point is not any single score; it is that a credible open-weight coding model now ships with day-one inference infrastructure, adjustable reasoning effort, and agentic benchmark coverage. The frontier is no longer the only address for coding capability. m:…1322

And NVIDIA is attacking the other end of the stack. RTX Spark consolidates the fragmented AI development workflow — today split across laptops, cloud model APIs, and cloud VMs — onto a single machine built around the GB10 system-on-a-chip with large unified memory. Code, test, and deploy in one place. If the coding assistant's moat was the model bundle and the cloud, NVIDIA is betting some of that value moves to local hardware that can run the whole agent loop. m:…5199

What the war is really about

Strip away the announcements and the war has one axis: who controls model supply to the developer's editor. Cursor built on borrowed models and is learning what that costs. Kilo is selling freedom from exactly that dependence, with a free year to make the switching math easy. Baseten and Reflection are making open weights a first-class, day-one option. NVIDIA is saying the machine under the desk can host more of the loop. This suggests the durable moat in developer tools is not the IDE, the keybindings, or even the agent UX — it is guaranteed, multi-source model supply with sane routing economics. Whoever owns that owns the developer's next decade, and the current scramble says nobody owns it yet. m:…6782 m:…1322