Zed Review 2026: Pricing, AI Agents, Pros and Cons

Zed is an open source code editor from the team behind Atom and Tree-sitter, and it now sells an AI layer on top. The pitch has two halves. The first is a native editor with multiplayer editing built in; the second is an Agent Panel that runs Zed's own agent or third-party agents side by side. The editor half is free, and the AI half is metered by the token. That makes Zed's bill very different from the flat AI subscriptions most rivals sell, and it rewards readers who know their own usage.

This review is for developers choosing between Zed and Cursor, VS Code with GitHub Copilot, or a terminal agent like Claude Code. It covers what Zed does, what each plan costs once tokens are counted, and where the 10% markup bites. It also shows when bringing your own API key is the cheaper route. Devshot runs no test lab, so every figure here comes from Zed's pricing page and documentation, not from a trial account.

Zed calls itself the world's fastest IDE, open source, with AI optional. The source code is licensed mainly under GPL-3.0-or-later, with Apache-2.0 components where marked, per the Zed repository. It runs on macOS, Windows and Linux, and the editor needs no account: you sign in only for hosted AI and collaboration.

Three things set it apart from a VS Code fork. Collaboration is native, with channels, shared projects and voice chat in a Collaboration Panel, so two developers can edit the same files with live cursors. The Agent Panel hosts Zed's own agent plus external agents connected through the Agent Client Protocol, including Claude Agent, Codex, OpenCode, Copilot and Cursor's agent. And edit predictions come from Zeta, an open source model Zed develops, with GitHub Copilot, Codestral, Mercury Coder or a local model as alternatives.

Zed Industries also launched Delta in 2026, a separate multiplayer environment for coding with agents. It shares the Pro trial balance with Zed, but this review covers the editor.

Zed gives you three ways to pay for AI, and only one of them goes through Zed. Bring your own key: on any plan you can connect Anthropic, OpenAI, Google AI, Amazon Bedrock, Mistral, DeepSeek, OpenRouter, Ollama, LM Studio and other providers. The provider then bills you directly, at its own rates. External agents: Claude Agent, Codex and the other agents in the ACP Registry own their login and billing, and Zed charges nothing for them, per Zed's external agents docs. Zed-hosted models: on Pro, Student or Business you use Zed's hosted copies of Claude, GPT and Gemini models, billed per token at the provider's price plus 10%.

The first two routes mean an existing Claude or ChatGPT subscription works inside Zed at no added cost. Parallel Agents then lets you run several agent threads and terminal threads at once from a Threads Sidebar, each with its own context window and history. Threads can run in separate Git worktrees, so two agents can work on one repository without touching each other's files. Zed's own advice is to start a new thread for each task: every prompt, file and reply in a thread grows the context, and on hosted models that growth is the token bill.

Zed's real price is the plan fee plus tokens, and the spend limit decides the ceiling. Pro's included token credit resets on your billing date. Past that, overage is charged at the end of the month or sooner as it builds up. The default spend limit lets extra tokens match the plan fee, so an untouched Pro account tops out at $20 a month, per Zed's plans and usage docs. Set the limit to zero and Pro costs exactly the plan fee, with hosted AI paused once the credit runs out.

Our arithmetic on a heavy month shows what the markup costs. Say one developer sends 10 million input tokens and 1 million output tokens to Claude Sonnet 5, ignoring cache discounts. At Zed's rates of $2.20 and $11 per million, listed on Zed's hosted models page, that is $33 of usage. The included credit absorbs part of it, leaving $28 of overage and a $38 month, which also needs a spend limit raised above the default. The same tokens on your own Anthropic key cost $8 less in total on the free Personal plan, since you pay only the provider prices Zed lists.

So Pro never wins on tokens alone. The fee buys only half its value in credit, and every hosted token carries 10% on top. At every volume we can compute, Pro costs more than the same usage on your own key. What the fee buys instead is Zeta edit predictions, which the Free plan does not include, plus one bill and higher rate limits on hosted models. Light users get the most from it: a cheap model like GPT-6 Luna at $0.11 per million input tokens stretches the included credit a long way.

Business and Student change the shape. Business seats include no token allotment, so a five-seat team pays $150 a month in seats on our arithmetic. Hosted usage comes on top at the same plus-10% rates, or runs through the team's own provider contracts. There is no minimum seat count and no free trial, order forms start at 25 seats, and SSO is not available yet (SAML and SCIM are planned). Verified university students get a free year with a monthly token credit, capped at that credit and without the Opus, Fable or GPT-6 Astra models.

Methodology: we read Zed's pricing page, its plans and usage docs and its hosted models page on October 19, 2026. Competitor prices come from Cursor's pricing page, GitHub Copilot plans and Claude's pricing page, read the same day. Totals and break-even points are our arithmetic on those list prices; nothing here comes from a benchmark or a usage test.

Zed fits a developer who already pays an AI provider and wants a faster editor to spend it in. With an Anthropic or OpenAI key, a Claude subscription or a Codex login, the free Personal plan gives you the editor, the Agent Panel and external agents at no extra cost. That combination is hard to match, because Cursor and Copilot both want a subscription before their agents do real work.

Pro makes sense for a light AI user who wants tab-style edit predictions and a hosted model without managing keys. Keep the spend limit at zero and the bill is fixed. A heavy agent user is better served elsewhere. The markup and the half-value credit make Pro the expensive way to buy a lot of tokens, while a flat plan such as Claude Code on Claude Pro caps the bill with usage limits instead.

Teams should read Business as an admin layer, not an AI bundle. It suits a company that wants to restrict which hosted models engineers use and lock data sharing, but a team that needs SSO today has to wait. Zed is also a weaker pick if your workflow depends on the VS Code extension catalog or on the deep language tooling JetBrains IDEs ship. Our best code editors roundup weighs those trade-offs.