Zed vs Cursor: AI features, API access and real cost structure

Zed and Cursor are different kinds of purchase: Zed is a free, open-source editor that sells hosted AI on top of it, while Cursor is a proprietary VS Code-derived editor whose subscription is the product. Zed's editor costs $0 and even its free plan can drive your own API keys and external agents; Cursor's paid ladder starts at $20 per month and its documented bring-your-own-key support reaches five named vendors, never Tab completion. Pick Zed if you want to pay for tokens instead of a seat; pick Cursor if the VS Code-shaped workflow and its first-party models are what you are buying.

One ownership change belongs at the top, because most comparisons predate it. Cursor announced on August 14, 2026 that it "has officially been acquired by SpaceX." The legal entity in Cursor's footer is still Anysphere, Inc. and the product still ships under the Cursor name. The post announces no change to prices, plans or data handling; what it does say about the product is that the deal brings more compute, that this "means we can provide customers with more capable models at lower cost," and that Grok 4.6 is an early result. That is a statement of intent, not a published price. No change to any plan price or policy was observed on the pages checked for this guide, so treat the acquisition as a fact about ownership and a forward-looking claim about models, nothing more.

Who should pick which

The execution model differs more than the feature lists suggest. Zed talks to model endpoints from your machine — its local-model documentation configures api_url as http://localhost:11434, which only works if the request originates on your laptop. Cursor's privacy documentation says Cursor sends prompts and code context to language model providers, and that Cloud Agents are the only feature that requires Cursor to store code — encrypted, and deleted after the agent completes; Cursor's API-keys page adds that all requests are routed through Cursor's servers for final prompt building, including requests made with your own key. If your endpoint is internal or unreachable from the public internet, that difference decides the question on its own.

Migration cost is asymmetric too. Moving to Cursor means importing VS Code settings into a familiar editor. Moving to Zed means a different editor with a different extension story and a settings.json block per AI surface. Budget an afternoon, keep your current setup installed, and judge on a real task rather than on a first impression.

Three AI surfaces in Zed, and where Cursor puts its limits

Zed splits its AI into three independently configured things, and a comparison that treats "Zed AI" as one feature will be wrong about two of them.

Two things follow. First, Zed's custom provider covers only Zed-owned AI features — its documentation names the Zed Agent, the inline assistant, git commit generation, thread summaries and similar Zed-owned features. An API key set there does not configure Claude Agent or Codex running inside Zed, and Zed's external-agent documentation says a Cursor subscription does not configure Zed's provider settings either. Second, the marketplace row corrects a widely repeated claim. Zed's own comparison page, updated May 1, 2026, says Cursor "directly inherits VS Code UI, extensions, and keybindings" and has "near-full VS Code extension compatibility", while Cursor's extensions documentation says it serves Open VSX through its own proxy and that not every Microsoft Marketplace extension is listed there. Trust the vendor's documentation about its own product.

What each one costs

Two inconsistencies in Zed's own material are worth knowing before you budget. Zed's pricing page names the free tier "Personal" and quotes "2,000 accepted edit predictions", while the docs plan table calls it "Free" and records the same quota only as "Limited"; the docs table also lists a Student plan the pricing page never shows. A third page, Zed's edit-prediction documentation, puts the number at "2,000 Zeta predictions per month" — the same figure, counted per month and without the word "accepted". Confirm the quota in your own account before you budget around it. No annual option appears on Zed's pricing page.

The prices above come from Cursor's plans and pricing help page; the quoted plan descriptions come from Cursor's marketing pricing page, which splits the same ladder into an Individual tab and a Teams & Enterprise tab and carries a monthly/yearly toggle labelled "Save 20% with yearly billing". The yearly per-plan amounts are not printed on either page, and none of the discounted figures circulating for them are published by Cursor, so confirm the amount in checkout rather than trusting any summary, including this one.

Why "ten percent markup versus no markup" is the wrong comparison

Zed publishes a per-model rate table, and its plans page states that hosted-model usage converts to dollars at the provider's published rate plus ten percent. You can therefore compute a Zed bill exactly. Cursor charges no blanket markup on individual on-demand usage — its pricing documentation says you continue at the same API rates, and its pricing page says on-demand usage is billed in arrears — but Cursor publishes no dollar value for the usage each plan includes. One documented exception sits on the same docs page: opting in to regional data residency adds a ten percent uplift on model pricing for eligible models. Its usage-limits page splits the allowance into a Cursor Models pool and an Other Models pool and gives no amount for either; the pricing page expresses the difference only as multipliers of the Pro tier. Any figure you have seen for "Pro includes $X of usage" is not published by Cursor, and this guide will not repeat one.

So the two structures cannot be reduced to a single number. Zed's markup applies to every hosted token with a $5 credit in front of it; Cursor's zero markup applies only after an allowance whose size you cannot look up. There is one place Cursor does charge for tokens it did not buy: on Teams and Enterprise plans, third-party model requests carry a Cursor Token Rate of $0.25 per million tokens, and Cursor's documentation says that rate applies to included usage, on-demand usage and bring-your-own-key usage alike. First-party Cursor models are exempt. Max Mode's 20 percent surcharge still exists but is described as legacy request-based plans only, so do not budget a new account around it.

A worked token estimate

Assume a hypothetical month of 2 million uncached input tokens and 200,000 output tokens on one model, with no caching and no tool charges. This is token arithmetic over a stated assumption, not a measured task cost, not a bill ceiling, and not a claim about which editor consumes fewer tokens for the same work — neither vendor publishes that, and nothing here measured it.

Zed's rates are read from its hosted-models page; Kunavo's come from the live catalog. Note what the Opus row does to a Zed Pro plan: Zed's documentation states the default additional spend limit for Pro is $10, "for a total monthly spend with Zed of $20", which means roughly $15 of hosted tokens before the cap stops you. A $16.50 month exceeds that by default and requires raising the limit. Zed also restricts Claude Opus models to Pro and Business, so the Free plan cannot reach that row through Zed's own hosting at all.

A caveat on the Sonnet row deliberately left out of the table: Zed's hosted-models page carries a note calling its Claude Sonnet 5 rate introductory through August 31, 2026 while still listing that rate on September 18, 2026. Nothing on the page resolves which is out of date, so no example here depends on it. If you are comparing Sonnet specifically, re-read that row before committing.

A cheaper published rate is not the same as a lower cost to finish the task. A model that needs three attempts at a low rate can cost more than one that succeeds once at a higher one, and no promise of an absolute lowest price is being made here. On a Cursor Teams seat, the same 2.2 million tokens would additionally carry $0.55 of Cursor Token Rate, on top of whatever the model itself costs — including when the key is yours.

Bring-your-own-key: first-class on one side, narrow on the other

Zed documents two custom provider types in settings.json, and the difference between their api_url values is the detail that wastes the most time. The OpenAI-compatible form takes the URL with the /v1 suffix; the Anthropic-compatible form takes the bare origin, because the client appends the path itself. Every documented example of both forms declares an available_models array, and Zed's own Add Provider flow asks for a model ID and context window, so you name the models yourself rather than having them discovered — and Zed's documentation says plainly not to put the API key in settings.json; it goes in the provider settings UI or the environment variable Zed generates.

{
  "language_models": {
    "openai_compatible": {
      "kunavo": {
        "api_url": "https://api.kunavo.com/v1",
        "available_models": []
      }
    },
    "anthropic_compatible": {
      "Kunavo": {
        "api_url": "https://api.kunavo.com",
        "available_models": []
      }
    }
  }
}

Kunavo serves both shapes: https://api.kunavo.com/v1 for OpenAI-compatible clients and https://api.kunavo.com for Anthropic-style ones, the same convention described in the base-URL reference. A complete, filled-in block is published in the Zed integration guide, which was written against Zed's documentation on September 6, 2026. That is a documentation check, not a runtime test: nobody has recorded a request from a running Zed to Kunavo, so treat the first task as a trial and watch the usage record. In particular, the Anthropic-compatible form exposes a prompt_caching capability flag that has not been exercised against Kunavo — leave it off until you have verified the cache accounting on a small request of your own.

Cursor's documented path is narrower. Its API-keys page lists OpenAI, Anthropic, Google, Azure OpenAI and AWS Bedrock, and none of the Cursor help pages read for this guide describe a custom base URL. The editor itself does carry one: an Override OpenAI Base URL toggle in Cursor Settings then Models, which a Cursor staff reply on Cursor's own forum confirmed on February 26, 2026, while naming a known limitation — the override "applies to all models, not just the ones using your custom key", so a Cursor-hosted model and a custom endpoint cannot be used side by side without toggling it. Absence from the help centre is not absence from the product, but it does mean Cursor has published nothing this guide can hold it to about the setting's behaviour or its future. Kunavo publishes no instructions for pointing Cursor at it and has not tested it, so nothing here should be read as a supported Cursor route. Also weigh the stated trade: Cursor says its Zero Data Retention policy does not apply when you use your own API keys.

If you choose the token route

Configure Zed's Agent Panel against a custom provider, or run Claude Agent or Codex as an external agent inside Zed on your own key — both paths work on Zed's free plan, so the editor adds nothing to the bill. Kunavo starts at a $10 prepaid top-up with no subscription; the catalog estimate above is a floor rather than a cap, because a charge can land higher when upstream-cost billing exceeds it, as the billing guide explains. Create a Kunavo account when you are ready to run that first trial task.

Still weighing editors? Cursor usage limits covers what happens when an allowance runs out, and Cursor alternatives widens the shortlist beyond these two.

FAQ

Every Zed and Cursor figure on this page is read from the vendor page linked beside it, checked September 18, 2026 and re-checked September 21, 2026, along with the public Zed repository. The one claim not taken from vendor documentation is the Override OpenAI Base URL setting, which is sourced to the linked Cursor forum thread and its Cursor staff reply. Kunavo rates are read from the live catalog, and all totals shown are illustrative token arithmetic over the stated assumption, not measured costs.