Ten days ago, on April 29, the Zed editor reached version 1.0. The team had been working toward that milestone for five years. The piece I wrote that day, Zed Is 1.0 — and the Electron Era Just Ended, was about why the foundation of the editor was the news: a native, GPU-accelerated, Rust-built code editor with no Chromium underneath, ready for the developers who passed on it during the long preview.
This piece is about what happened next.
In the ten days from April 29 through May 8, Zed shipped four stable releases after 1.0, posted four blog entries, launched a paid Business plan, opened a public conversation about why the team is investing in AI at all, and released a new edit-prediction model that uses about a third as many tokens as the one it replaced. None of those things were on the launch slide for 1.0. All of them landed in the time it takes most software teams to argue about a sprint goal.
The cadence is the story. The features are how you read it.
Ten days, six shipping events
Here is the calendar, in the order things actually happened. The dates are pulled from Zed's own stable-release page and the team's blog.
| Date | Release / event | What shipped | Why a normal user notices |
|---|---|---|---|
| Apr 29 | 1.0.0 | Five years of work declared stable | Foundation is no longer marked "preview" |
| May 4 | 1.0.1 | Agent edit-apply fix | Agentic code edits stop silently failing |
| May 5 | Blog: We're Not Building AI Features for the Money | Philosophy post on why AI is in Zed | Counter-narrative to vendor-AI hype |
| May 6 | 1.1.5 + Zed for Business | Panel layout switcher (classic / agentic), git graph view, split diff in agent and file diff panels, LSP code lens, Helix amp jump navigation, DeepSeek V4-Pro/Flash and OpenCode Go provider support — plus a $30-per-seat Business plan with org-wide AI controls | The editor for teams now exists, and the headline interaction surface changed |
| May 6 | 1.1.6 | Windows ACP-launch fix, Linux inotify event-queue overflow fix | Zed actually works on Windows and on busy Linux trees |
| May 8 | 1.1.7 + Zeta2.1 | zeta2 prompt-format fix, filesystem-error CPU regression fix, Helix-motion panic fix, markdown-preview reload — plus a new edit-prediction model with 67% fewer output tokens and 28% lower median latency | Suggestions feel snappier and the editor stops eating CPU on a broken symlink |
A quick reading of that table is enough to see the pattern. May 6 is the loud day: a feature release, a Business plan, and a same-day patch chasing the bugs the feature release surfaced on Windows and Linux. May 8 is the quieter substantive day: a small bugfix release in the foreground and a new AI model in the background, shipped together because the model and the editor have to land at the same time for either to work.
There are also two version numbers that did not happen. Zed went from 1.0.1 to 1.1.5 without 1.1.0 through 1.1.4 ever being promoted to the stable channel. Those numbers existed; they were preview-channel cuts, real builds with real changes, that the team chose not to push to every user. The decision to skip them is its own piece of information about the cadence: Zed runs a fast preview channel and a careful stable channel, and lets the users who like the cliff edge ride the preview while the rest get a smaller number of stable promotions.
The May 6 story: the day Zed became a business
The most significant thing that happened in the ten-day window did not have a version number. On May 6, Zed Industries announced Zed for Business — a $30-per-seat-per-month plan aimed at teams that want central control over the AI defaults their engineers can flip.
The shape of the plan is worth reading carefully. Companies can bring their own API keys from Anthropic, OpenAI, Google, or AWS without an additional Zed markup, or use Zed-hosted AI billed at provider cost plus 10%. Prompt sharing and edit-prediction training are off by default at the organisation level, and individual engineers cannot override that setting. Administrators can disable Zed-hosted models, edit predictions, and collaboration features for the whole organisation, and set spend limits on tokens.
That last detail is the one that makes the Business plan more than a SKU. The privacy guarantees normal users have always had on Zed — no prompt storage by default, no training on your code by default — are now enforceable as policy. A security team can lock them on. The individual engineer cannot opt back into "share my prompts" by accident on a Tuesday afternoon. That is not the same product as "Zed with AI features turned on" — it is a meaningfully different artefact aimed at a different buyer.
The same day, release 1.1.5 added the panel layout switcher between a classic IDE arrangement and an agentic one. The two layouts are both first-class. You pick which one matches the work you are doing in the moment. A debugger session in classic; a multi-agent refactor in agentic. The editor stops insisting that there is one right way to lay out the screen.
Then, while the new layout was rolling out, Windows users on certain configurations could not launch their Agent Client Protocol agents at all, and Linux users on busy trees were hitting inotify event-queue overflows. Version 1.1.6 shipped the same day to fix both. The polite reading is "they caught the Windows and Linux bugs in their dogfooding within hours and pushed a fix that afternoon." The honest reading is the same.
The May 8 story: a smaller, faster brain
Zeta is the model that powers Zed's edit prediction — the inline ghost-text suggestions that appear as you type, that you accept with Tab. It is not the agent. It is the smaller, lower-latency thing that runs continuously in the background, trying to keep up with where your cursor is going.
On May 8, Zed posted Zeta2.1. The numbers are the headline:
- Output tokens dropped from about 270 to about 90 — a 67% reduction.
- Median latency dropped from 189 ms to 136 ms — about 28% faster.
- Acceptance rate improved by 0.51%; explicit-rejection rate fell by 4.10%.
- Infrastructure footprint dropped by roughly 30% — fewer servers carrying the same traffic.
The technical change is a new prompt format the team calls Multi-Region. The previous version had the model output a large region around your cursor with the model's edits applied; the new one only outputs the slice of code that actually changed. The model has the same amount of context going in. It says less coming out. Less to generate, less to send over the wire, less to render on screen.
For someone using the editor, the practical consequence is: suggestions feel slightly snappier, and the model says yes more often when you accept the suggestion. The deeper consequence is in the model's open-weight release on Hugging Face, trained on opt-in open-source data. The model that ships in your editor is the same model anyone can download, inspect, and run independently. That is an unusual posture for a feature in a 2026 IDE. It is also the posture the Zed team has been talking about for several years.
The same day, version 1.1.7 closed out the small bugs in the foreground — including a fix for local Zeta2 edit predictions, which had been using the wrong prompt format.
The Zed Guild
The piece that shows up in most release notes only as credits at the bottom, rather than as a headline item, and that I think matters more than most of what does get the headline framing, is the Zed Guild.
The Guild is a twelve-week cohort program for outside contributors. Selected applicants pair with a Zed engineer for the duration of the cohort and ship features into the actual repository. The first cohort has finished. The page that describes the program is, by 2026 standards, almost embarrassingly low on marketing copy: a paragraph of program description, a wall of GitHub avatars from cohort members, and a closed application window.
The reason this matters for an article about ten days of shipping is that ten days of shipping at this density is not something an in-house team produces by itself. The 1.1.5 release notes credit a long list of community changes alongside the marquee features. The Guild is one of the legible mechanisms by which that list gets longer. It is also a quietly important answer to the question every editor that wants to outlast its founders eventually has to answer: who else cares about this codebase enough to keep it healthy when the founders eventually move on? Atom, the editor that taught the Zed founders what they did not want to build again, was killed by its corporate owner in 2022. The Guild is a slow, careful bet on building a constituency that does not depend on a single corporate owner staying willing.
Why the cadence is the actual story
The reason the calendar matters is that the dominant editor-category competition in 2026 is between three different theories of what an editor is for, and the theories ship on three different clocks.
VS Code wins on inertia. The integrated bet is that once you and your team have invested in extensions, settings, and muscle memory, you do not switch even when something better appears. Microsoft can ship at any pace because the customer's switching cost is already doing the work.
JetBrains wins on completeness. The integrated bet is that once you need refactoring, database tooling, and language intelligence at IntelliJ depth, you accept the long startup time and the heavy memory footprint, because nothing else covers the same ground. JetBrains can ship a major IDE on a multi-month rhythm because the alternative is not catching up.
Zed is trying to win on momentum. The integrated bet is that an editor that meaningfully improves every two weeks pulls users toward it the way a static editor cannot pull them, because the gap between what your editor was a month ago and what it is today is large enough to feel. That bet only works if the cadence is sustainable. The ten-day window between 1.0 and 1.1.7 is the first public proof that the cadence is real on the stable channel, not just in preview, and not just for one release. Five stable releases, four blog posts, a Business plan, a new model, and an open community program — that is what the bet looks like when it lands.
It is too early to call the bet won. Six months of shipping at this density would be the harder test. Three months of shipping at this density while the team stops running on launch adrenaline would be the test after that. What we have today is ten days of evidence that the cadence is being delivered.
What to do with this
If you have been waiting on the original should I switch question, my answer is the same as it was ten days ago: do not switch your daily-driver editor in the middle of a project, but it is now a reasonable thing to put on the side project you start next month.
The thing that has shifted is the watch interval. The right amount of attention to give Zed in May 2026 is roughly check the changelog once a fortnight and see whether anything you would actually use has landed. That is not how I have ever paid attention to an editor in twenty years of doing this work. It is the right amount of attention to pay to this one.
Ten days. Five stable releases. Four blog posts. A Business plan. A smaller, faster AI model. A community program that finished its first cohort. None of it was on the launch slide.
The launch was the foundation. The foundation is now visibly carrying weight. The next month or two will tell us whether the people on top of it can keep stacking, or whether the rhythm slows and the pattern shifts to the one we have all seen before. For now, the rhythm has not slowed. That alone is worth saying out loud.
Top comments (0)