Supermaven

Nothing to pick. It was sunset in November 2025 and VS Code users were told to move to Cursor, where the completion model now lives as Cursor Tab. If you already have it on Neovim or JetBrains, nobody will make you remove it, and the completions still arrive, but nobody will improve them either, and the day they stop there will be no announcement. The daily trait that decides it is that there is no daily anymore.

Everyone else: pick Cursor for the model that replaced it, Tabnine for a completion vendor that still sells to you, and Zed if you want fast completion inside an editor with a future.

The risk is that the product is shut. The sunset post drops agent conversations outright, calling them a newer, infrequently used addition, and keeps autocomplete inference running for the foreseeable future, a phrase with no end date, no contract and no one obliged to honor it. A tool that can be switched off by a blog post is not a dependency; it is a courtesy.

Anyone still typing against it should have the replacement installed before the courtesy ends. What it did right: prorated refunds went out the same day as the announcement, which is more than most sunsets manage.

The architecture is worth recording. 1. A proprietary low-latency completion model with a claimed 1M-token context window, self-reported and never measured on a published harness, so the number is a specification rather than a result. 2. A chat panel backed by GPT-4o and Claude 3.5 Sonnet, so the in-house model never handled conversation; the proprietary part covered autocomplete only. 3. No verification loop, since completion has nothing to verify against.

No benchmark was published while it was live. The observation: the design survived the model change by being absorbed into its acquirer's product, which is one way to be reproducible.

A founder-led completion company with a fast model, bought by Anysphere in November 2024, retired a year later once the model had been integrated into the acquirer's editor. Customers were a liability to be refunded, not an asset to be kept, which tells you what was bought: the team and the weights, not the revenue. That is the acqui-hire pattern in its cleanest form.

No pivot, no acquirer to name; the acquirer already happened and the product was the wrapper. Position: closed, and a case study in roadmaps following cap tables.

The demo was fast completion on a large repository. Procurement is closed: Team was $10 per user, $600 a month for sixty, but there are no new subscriptions, no contract, no SSO, no audit log and no vendor to sign an MSA. Seats still running sit on inference the acquirer can withdraw without notice, which is a dependency with no owner on our side or theirs.

Any engineer still using it is on a tool with no support path, so the migration is a ticket, not a decision. Not yet, and not ever.

Nothing to own: a proprietary completion model behind an API, no bring-your-own key, no local weights, no MCP, no source for the part that mattered. What survives is supermaven-nvim on GitHub, a Lua plugin I can read and patch that still fetches completions for existing users, and reading it tells you exactly how thin the client was.

When that endpoint goes dark the plugin is a client for nothing. A fork of the client cannot fork the model, which is the whole lesson: forkability is only real for the layer that does the work.