Zed AI pricing: hosted rates, BYOK costs and model choice

For Zed's own AI surfaces the best API is one you configure yourself, because Zed's free tier can already reach an OpenAI-compatible or Anthropic-compatible endpoint and you do not even need a Zed account to do it. The Personal plan's own feature list reads "Unlimited use with your API keys or external agents like Claude Agent, Codex CLI, and more", and the same page's FAQ answers "I don't want to pay for Zed, can I use my own keys?" by naming Anthropic, Google AI, OpenAI, OpenRouter, Ollama and others (zed.dev/pricing, checked September 19, 2026); Zed's authentication documentation says "Signing in to Zed is not required" and that "To use AI without signing in, you can bring and configure your own API keys". What a paid plan actually buys is unlimited Tab completion and Zed-hosted inference — two things a gateway cannot substitute for in the same way, and the difference is the whole decision.

Two naming traps first, because search results mix them. ZED is also a line of stereo cameras from Stereolabs with its own hardware prices; none of that is this product. And "Zed AI" is not one product in today's documentation, which splits the surface into Zed Agent (native, driven by your configured providers), External Agents over the Agent Client Protocol, and Terminal Threads. A write-up that treats "Zed AI" as a single hosted service is describing a shape the current docs do not use.

Three separate things Zed can charge you for

The most common budgeting error here is treating "Zed" as one line item. It is three, and they are priced by different mechanisms.

Zed's external-agent documentation states that "Zed does not charge for External Agents" and that "Billing, legal terms, retention, and data handling are between you and the agent provider". Its curated list names Claude, Codex, OpenCode, Copilot, Cursor, Gemini CLI, Pi Coding Agent and Poolside. If you already run one of those on a key, most of the Zed pricing question is already answered for you — see the OpenCode provider comparison and the Codex one for the routes those own.

Zed AI pricing: the plans, and the caps that actually stop the bill

Three mechanics matter more than the headline numbers. First, Zed's plans and usage page says "The default value for Pro users is $10, for a total monthly spend with Zed of $20" and that it "can be set to $0 to limit your spend with Zed to exactly $10/month"; once reached, "we'll stop any further usage until your token spend limit resets". So an untouched Pro account is a $20 ceiling, not an open meter. Second, Zed's billing page says the Business org-wide AI spend limit "starts at $0, so it must be increased before members can use any hosted models" — a Business seat on its own buys zero inference. Third, the same page describes threshold billing: charges for incremental token spend start at $10 pre-tax and Zed "may automatically raise your pre-tax invoicing threshold in $10 increments, up to $100", and "Once raised, the invoicing threshold is not automatically lowered during the same subscription".

One conflict to resolve in your own account rather than from any summary. Zed's education page says "The Zed Student plan does not include Claude Fable, Claude Opus, or GPT Pro models", while the plans and usage page describes the same plan as "all hosted AI models except Claude Opus". Those are different exclusion lists. Confirm which one your account enforces before planning around a specific model. Eligibility on the education page is current enrolment at an accredited university and being at least 18, and after 12 months the plan "will automatically downgrade to a Free plan". Zed also publishes no annual price anywhere on the pricing page — which is not the same as proving no annual option exists behind checkout. Business has no free trial, no minimum seat count, and order-form contracts at 25 or more seats; SSO, SAML and SCIM are described as "planned but not currently available".

The hosted rate card, and the one column that does not line up

Zed publishes a per-model table headed "Provider Price per 1M tokens" and "Zed Price per 1M tokens", and says "Any usage of a Zed-hosted model will be billed at the Zed Price (rightmost column above)". Zed's pricing page states the Pro rule as the API list rate plus ten percent, and on every row below that arithmetic is exact. The interesting question is what the provider column is compared against.

Sources, all read September 19, 2026: Zed's columns from zed.dev/docs/ai/models.md, the Anthropic rates from claude.com/pricing, the OpenAI rates from developers.openai.com, the Google rate from ai.google.dev, and Kunavo's from the live catalog. The Anthropic and Google rows agree with their vendors. The three GPT-5.6 rows do not, on the day both pages were read. This page publishes both numbers and stops there: it does not assert why they differ, and it is not claiming Zed is charging beyond its stated policy. One published fact worth knowing before you guess: OpenAI's own page notes that GPT-5.6 Sol's promotional pricing is available at least through November 21, 2026, so at least one of these rows sits on a promotion with an end date. Re-read both pages before budgeting a GPT-5.6 workload on the hosted path.

Two structural notes on the hosted table. Zed caps hosted Gemini 3.1 Pro requests at 200k tokens "because pricing changes above that context size" — Google's published rate above 200k is $4.00 input and $18.00 output. Its Gemini rows also carry no cache-write or cached-input price at all: Zed's models page states that hosted Gemini requests do not use Google context caching, so that usage bills purely as input and output tokens. Zed's Sonnet 5 row still carries a note calling the rate introductory "through August 31, 2026" while listing that same rate in September; Anthropic's published rate has not moved, so read the footnote as stale rather than the price. Zed also gates its top models: "Claude Fable 5, Claude Opus models, GPT-5.5 pro, and GPT-5.4 pro are only available on Zed Pro and Zed Business", and it keeps a retirement list worth checking before you trust any older write-up: the Grok entries went on May 15, 2026, Gemini 3 Pro on March 26, 2026 and Claude Sonnet 3.7 on February 19, 2026.

Three boundaries no single Zed page states together

This is the part a rate table cannot tell you, and each one costs an afternoon or a surprise charge.

1. The two provider types take different api_url shapes. Zed's API-access documentation shows openai_compatible with an api_url ending in /v1 (https://example.com/v1) and anthropic_compatible with a bare origin (https://api.someprovider.com). That documentation does not explain the difference and carries no warning about getting it wrong, so the reason is the client convention rather than anything Zed states: the Anthropic SDKs append /v1/messages to the base URL themselves, which is why the base-URL reference says including /v1 there produces requests to /v1/v1/messages. Copy the two shapes Zed shows. Kunavo serves both — https://api.kunavo.com/v1 and https://api.kunavo.com.

{
  "language_models": {
    "openai_compatible": {
      "kunavo": {
        "api_url": "https://api.kunavo.com/v1",
        "available_models": [
          {
            "name": "claude-sonnet-5",
            "display_name": "Claude Sonnet 5 (Kunavo)",
            "max_tokens": 200000,
            "max_output_tokens": 64000,
            "capabilities": { "tools": true, "images": true }
          }
        ]
      }
    },
    "anthropic_compatible": {
      "Kunavo": {
        "api_url": "https://api.kunavo.com",
        "available_models": [
          {
            "name": "claude-sonnet-5",
            "display_name": "Claude Sonnet 5 (Kunavo, Messages)",
            "max_tokens": 200000,
            "max_output_tokens": 64000,
            "capabilities": { "tools": true, "images": true, "prompt_caching": false }
          }
        ]
      }
    }
  }
}

The key does not go in that file. Zed's documentation says "Do not put API keys in settings.json" for both provider types, and adds that keys saved through Zed are "stored in the system keychain, not in settings.json" and that "Non-empty environment variables take precedence over keychain values". The naming rule it publishes is scoped to the OpenAI-compatible type — the variable is the configured provider id in upper snake case plus _API_KEY, so the kunavo block above would read KUNAVO_API_KEY. No equivalent rule is stated for the Anthropic-compatible type, so read Zed's own environment-variable table for that one rather than assuming the same pattern. There is also no model discovery on either custom path: available_models is the catalog, and a model you did not declare is not in the picker. If a model only works through the Responses API, Zed's documentation says to set capabilities.chat_completions to false and it will use the Responses endpoint instead.

2. prompt_caching defaults to false on the Anthropic-compatible path. Zed's documented defaults for that provider type are tools: true, images: false, prompt_caching: false. Zed's wording is "Enable prompt_caching to send explicit cache_control breakpoints for prompt caching; leave it disabled if the provider rejects requests containing them". With the flag off, Zed sends no breakpoints on that path, and Anthropic's protocol bills an uncached prefix as ordinary input — so a long agent thread re-bills its full input on every turn. This is a statement about the Anthropic-compatible path only; the OpenAI-compatible path has its own prompt_cache_key flag and its own caching behaviour, which this page does not characterise. No cache-savings figure is published here, because that combination has never been exercised against Kunavo and a number nobody measured would be worse than no number. Read how caching is billed and what prompt caching does to a bill, then verify on one small request of your own before you assume the discount.

3. Bring-your-own-key cannot reach Tab completion. Edit prediction is configured under edit_predictions, not language_models, and its non-Zeta providers are copilot, mercury, codestral, ollama and open_ai_compatible_api. That last one is not the chat API: Zed's edit-prediction documentation says your server "must implement the OpenAI /v1/completions endpoint" and posts a fill-in-the-middle body of model, prompt, max_tokens, temperature and stop. Kunavo exposes no /v1/completions route at all, so Kunavo cannot serve Zed edit prediction. A local model through Ollama or an FIM server can, which is the one job on this page where local genuinely wins.

Two smaller boundaries, stated because guessing them costs money. A provider entry configured for Zed Agent does nothing for an External Agent: Zed's documentation states that "An Anthropic API key configured for Zed Agent does not automatically configure Claude Agent", and the same for an OpenAI key and Codex. It does not say whether Terminal Threads inherit that separation, so treat that one as unestablished rather than settled. And Zed's Business admin controls are described over hosted models, edit predictions and data sharing; nothing in the documentation describes admin visibility into member-configured custom providers, so whether a Business administrator can see or cap that spend is not established here rather than answered.

A worked cost estimate under three stated workloads

These are illustrative token arithmetic, not measured task costs and not a bill ceiling. Three shapes, all assumptions: A an inline assist at 20,000 uncached input and 1,500 output tokens; B one agent thread on a feature at 400,000 input and 25,000 output; C an agent-heavy month at 8,000,000 input and 500,000 output. No caching, no tool charges, one model throughout, and no long-context surcharge — the GPT-5.6 models and Gemini 3.1 Pro each bill a large enough single prompt at a higher tier, which this arithmetic does not model. Rates come from the live Kunavo catalog.

Now put workload C against Zed's hosted rates for three of those models, using the published figures above. Claude Opus 5: $57.75 hosted against $21.00 at catalog rates. GPT-5.6 Terra: $30.25 against $7.70. Claude Sonnet 5: $23.10 against $21.00 — close, because Kunavo prices that model at parity with Anthropic's published rate while Zed adds its markup. The gap is not uniform across the shelf, and a page that told you it was would be selling you something — note that the catalog currently lists Claude Opus 5 and Claude Sonnet 5 at the same rate, so those two rows differ only in what Zed charges for each.

The sharper reading is about the cap, not the rate. A $57.75 hosted month is more than double a default Pro account's $20 total ceiling, so that workload does not merely cost more on the hosted path — it stops partway through until you raise the spend limit, and Zed's billing page says the invoicing threshold it raises in response is not lowered again during the same subscription. Scale all of this by your own sessions before treating it as a budget. Kunavo's catalog amount is a billing floor rather than a cap: when the upstream reports its charge, the bill is the greater of catalog cost and upstream cost times the applicable markup, as the billing guide explains. The minimum top-up is $10 in prepaid credit — a funding minimum, not a task fee or a subscription.

Which route wins, and when

One troubleshooting note that belongs with the gateway route. Zed 1.20.2, released September 17, 2026, records a fix for "payment errors from non-Zed model providers incorrectly prompting users to upgrade to Zed Pro instead of displaying the original provider error message" (zed.dev/releases). If you are on an older build and a custom provider returns a payment or balance error, Zed may have shown you a Zed Pro upsell instead of the real message. Update before you conclude your key is fine. Kunavo's error reference covers what the endpoint actually returns.

Choosing a model for the Zed agent

Zed reads a specific set of capability flags for custom providers, and getting those wrong looks like a model problem when it is a configuration problem. On the Anthropic-compatible path the defaults are tools: true, images: false, prompt_caching: false, plus optional extra_beta_headers, a mode for thinking with a token budget, and default_temperature. On the OpenAI-compatible path they are tools: true, images: false, parallel_tool_calls: false, prompt_cache_key: false, chat_completions: true, interleaved_reasoning: false, max_tokens_parameter: false, with reasoning_effort accepting none, minimal, low, medium, high, xhigh or max. An agent model needs tools on; anything you expect to read a screenshot needs images turned on explicitly.

On max_tokens, which Zed uses as the context window and requires: prefer a conservative value such as 200000 over the model's advertised maximum. Kunavo's published context windows are vendor figures rather than per-route measurements, so a settings file declaring a million-token window is asserting something nobody has tested. Raise it after you have sent a large request successfully, not before.

And the honest limit on "best model": no completion-rate measurement was run for this page, and neither Zed nor Kunavo publishes a Zed-specific benchmark. Unit price ranks budgets, not outcomes — a cheaper model that needs three attempts on a refactor can cost more than one that lands it once. Use the table above to shortlist by cost per workload, then judge on your own repository. The coding-model comparison and cost optimization cover that method in more depth.

Setting it up

Kunavo publishes a filled-in configuration reference for this client. That is a documentation check, not a compatibility test: no request from a running Zed to Kunavo has been recorded, and every behaviour described above was read from Zed's own documentation on September 19, 2026. Keep a working route available, run one bounded task, then read the charge your account recorded for it — especially if you turn prompt_caching on. Start at the Zed integration guide, and create a Kunavo account when you are ready to fund a key.

Still comparing editors rather than providers? Zed vs Cursor puts the two purchase models side by side, and the agent API directory covers which other clients accept a custom endpoint.

FAQ

Every Zed, Anthropic, OpenAI and Google figure on this page was read from the page linked beside it on September 19, 2026; Zed behaviour claims come from Zed's documentation and release notes, not from a runtime test of Kunavo inside Zed, which has not been performed. Kunavo rates are read from the live catalog, and all dollar totals shown are illustrative token arithmetic over the stated assumptions.