OpenAI is cutting Cursor off on 12 November. Here's what actually breaks.

OpenAI said on 28 August that it's ending the agreement that puts GPT models inside Cursor, effective 12 November 2026. The reason given isn't commercial. SpaceX owns Cursor now, and OpenAI says it can't be confident a Musk-owned company will stay inside its terms of service.

"we cannot be confident that SpaceX will use our technology within our terms of service, based on our experience with Elon Musk's companies violating contracts."

SpaceX closed its $60 billion all-stock acquisition of Anysphere, Cursor's parent, earlier this month. That tripped a change-of-control clause, and OpenAI used it — giving what it describes as the maximum notice the contract allows. Which works out to about ten weeks.

What actually breaks

Less than the headline suggests. Cursor keeps working. Claude, Gemini, Grok and Cursor's own Composer models all stay. Anthropic's Tom Brown replied within hours that Anthropic will keep adding compute for Claude inside Cursor. One lab's supply cut is another lab's distribution win, and Anthropic was not subtle about noticing.

The part worth planning around is quieter. Pasting your own OpenAI key into Cursor doesn't restore what you had. It reportedly covers local Chat and Agent in the desktop app — and not Tab and autocomplete, not Auto mode, not cloud agents, not Automations, not the CLI, not the API or SDK. So if your team leans on Auto to pick the model, or runs long jobs through cloud agents, bring-your-own-key is a patch and not a migration.

The lesson isn't really about Cursor

It's that model supply is a political dependency now. Nobody at Cursor shipped anything wrong here. OpenAI's stated reason is who owns them, not what they built. You can be a model-citizen customer and lose access anyway, on someone else's timeline, for reasons that never appear in your changelog.

Which makes the boring defence the right one: keep an abstraction between your code and the model. Vercel put Cursor behind the AI SDK's harness layer on 27 August via @ai-sdk/harness-cursor, using the same HarnessAgent interface as every other harness. Gateways do the same job for plain model calls — the model becomes a string you resolve at runtime instead of a provider SDK you import.

import { generateText } from 'ai'

// With a gateway, the model is a slug resolved at runtime.
// Switching providers is an env change, not a refactor.
const { text } = await generateText({
  model: process.env.MODEL_ID!,
  prompt: 'Summarise this diff.',
})

There's nothing clever there, and that's the whole point. The teams who spend November calmly are the ones for whom changing providers is an environment variable and a deploy.

What I'd actually do

• If Cursor plus GPT is just your editor, there's no emergency — but stop treating Auto as load-bearing. Pin a default model explicitly this week and see whether anyone notices. If nobody does, you're already migrated.
• If anything programmatic runs through Cursor — CI review bots, codegen scripts, the CLI — that's the piece with no key-swap fallback. Move it behind a gateway or the AI SDK in September, not in the second week of November.
• Don't change editors over this. Cursor with Claude is still Cursor, and Anthropic just said it's leaning in. Swapping an editor costs a week of everyone's muscle memory; swapping a default model costs a dropdown.

Also this week

• Tencent open-sourced Hy4 Preview on 28 August: a 770B-total, 49B-active MoE with a context window over 1M tokens, at $0.834 per million input tokens and $2.501 per million output. It scored 2.99/4.00 against GLM-5.3 at 2.92 and Kimi K3 at 2.94 — but that's Tencent's own blind eval (163 experts, 203 engineering tasks), and a 0.05 margin on a vendor's own scorecard isn't a result. It's on Vercel's AI Gateway already if you want to judge for yourself.
• Debian settled its generative-AI question. The winning option neither endorses nor prohibits LLM-assisted contributions, and holds contributors fully responsible for understanding, reviewing, testing and where appropriate modifying anything they submit. The stricter anti-AI options lost to "None of the Above". Useful precedent if you're drafting your own policy.
• Lemmalog treats long-running agent investigations as incremental program analysis: facts go into a Datalog engine with provenance, so when an observation changes, the conclusions that depended on it get retracted instead of the model re-deriving everything from a transcript. Reported 0.463 F1 on LongMemEval and 0.533 on LoCoMo, with 38x smaller context. Best context-engineering read of the week.
• Neon shipped agent-first CLI commands: neon mcp installs its MCP server into an agent with auth and tool-scoping, neon skills drops skills into a repo, neon plugins publishes to agent marketplaces. Tool-scoping is the interesting bit — limiting which tools an agent can even see.
• Cloudflare AI Search added GLM-5.3 Flash at 1,048,576 tokens, days after adding six more text models from DeepSeek, OpenAI, Qwen and Moonshot. You can now pick the generation model inside AI Search rather than bolting a provider on the side.
• Claude Code shipped six releases in four days: a restricted execution mode that strips command-execution tools, hook events that can gate model switching, and a warning for overly permissive permission configs. Same direction as the Neon change — the agent's blast radius is becoming something you configure instead of something you accept.
• Dan Luu on bug blindness: most people have stopped noticing defects entirely, and the blindness runs strongest inside the teams shipping the thing. Not a release, but it'll ruin your week in a useful way.