Obsidian Copilot pricing and API cost: which backend pays for the tokens

Copilot for Obsidian costs $0 to install and run with your own API key. The paid tiers are Lite at $7.99 a month, Plus at $14.99 a month and Supporter at $349.99 once — and none of those is the number most people are looking for, which is the model bill the plugin runs up on your behalf. In the 4.x releases that bill changed shape: the plugin no longer builds its own semantic index over your vault, so vault size is no longer a cost driver at all. What decides the number is which agent backend you run.

Three name collisions first, because they contaminate this search. GitHub Copilot is a different coding product from a different company, and its per-seat prices describe that product. Obsidian itself is free for personal use and sells Sync at $4 per user per month billed annually and Publish at $8 per site per month billed annually (obsidian.md/pricing, checked September 19, 2026) — those are the app maker's prices, not the plugin's. And an unrelated open-source project at eugeneyan/obsidian-copilot shares the name; its main branch has been quiet since June 30, 2024, so its instructions are not the ones below.

Obsidian Copilot pricing: what the plugin charges

The plugin is open source under AGPL-3.0, authored by Logan Yang and shipped commercially by Brevilabs. The current release is 4.0.9, requiring Obsidian 1.11.4 or newer (manifest.json), published September 16, 2026, with 4.0.8 on September 12 and 4.0.7 on September 9 (releases feed). Manifest, licence and release feed checked September 19, 2026. Obsidian's own community plugin stats recorded 2,011,167 downloads when re-checked on September 21, 2026 — that counter moves daily, so read it rather than quoting this figure.

From the official pricing page, checked September 19, 2026. Two gaps are worth naming rather than filling. The included allowance on Lite and Plus is published only as a 5-hour and weekly quota and a "3x" ratio — neither the pricing page nor the plan documentation puts a token count or a message count on it, so neither can be converted into a number of sessions for you. And the top-up credits you buy when you pass the cap carry no price on the pricing page either; it only says they are purchased on the usage dashboard. Note also that the pricing page lists no Believer tier, so a Believer price quoted in an older blog post is not a current anchor.

What 4.x changed, and why the embedding bill is gone

This is the part that invalidates most of what is written about Copilot's costs. In V3 the plugin kept its own embedding index over your vault, and vault size drove an embedding bill. In 4.x that is removed. The vault search documentation states that semantic search from Copilot V3 has moved to Miyo and that Copilot no longer exposes controls for its old in-plugin index. Budgeting "embedding tokens per re-index" today prices a component that was deleted.

Miyo is a separate local-first desktop app for macOS, Windows and Linux from the same maker. Its own page states that the desktop app and CLI stay free; the paid tiers are Relay at $4.99 a month, needed only for remote access from outside your machine, and a $149.99 Lifetime licence (checked September 19, 2026). So the retrieval layer that used to scale with vault size now costs $0 locally. Kunavo does not serve an embedding model and should not be your provider for that step — but in 4.x nothing is asking for one.

Two consequences for the budget. First, Agent Chat's cheapest search path needs no index at all: the documentation says file search finds exact words, phrases, filenames and paths with the active agent's normal tools. Second, the token cost driver is now the agent reading notes into the prompt, not an upload of your vault — the plan documentation says an agent can read a local note and include relevant text in the prompt, and that opening Agent Chat alone does not upload the whole vault. A 20,000-note vault and a 200-note vault cost the same if the agent reads the same five notes.

One more boundary that surprises people: the documentation states that Free Chat never searches the vault automatically, and that automatic keyword retrieval in Quick Chat is a Plus behaviour. Agent Chat, including the opencode, Claude and Codex backends, is desktop-only; on mobile you get Quick Chat.

Which backend actually pays for the tokens

The manifest describes 4.x plainly: run AI agents such as Claude Code, Codex and OpenCode inside your vault. That sentence is also the billing model. Copilot is the host; the backend you pick decides who charges you and through which credential.

Read from the provider documentation and the settings reference on September 19, 2026. The practical read: two of the three agent backends take no key from Copilot at all — the documentation says Claude and Codex inherit authentication and models from their own CLI, and that BYOK models are never added to those two lists — so a generic "best API for Obsidian Copilot" ranking that assumes one settings tab decides everything is answering a question the software does not ask.

A worked Obsidian Copilot API cost estimate

These figures are token arithmetic under stated assumptions, not measured task costs and not a bill ceiling. Assume a Quick Chat ask sends 8,000 uncached input tokens — a note or two plus the conversation — and returns 800 output tokens. Assume an Agent Chat session sends 150,000 uncached input tokens across a multi-turn tool loop that reads several notes and re-sends context each turn, and returns 10,000 output tokens. Those ratios are illustrative choices, not measurements. Rates are live Kunavo catalog prices per million tokens.

The spread between the two columns is the whole point. A Quick Chat ask on Claude Haiku 4.5 lands near $0.0048 under these assumptions, which is rounding error; the same model in an agent loop lands near $0.080, and Claude Sonnet 4.6 near $0.240. Vault size does not appear in either figure — how many notes the agent chooses to read does.

A comparison worth making carefully: $7.99, Lite's monthly price, buys roughly 33 sessions of the shape above at Claude Sonnet 4.6 catalog rates, or roughly 99 at Claude Haiku 4.5. That is not a like-for-like comparison, and treating it as one would be dishonest: Lite's allowance is published only as a 5-hour and weekly quota with no token figure, so what it includes cannot be counted. It is only useful as a floor — if your real usage is a handful of Quick Chat asks a week, metered tokens are very hard to beat on price; if it is hours of agent work a day, a subscription route deserves a serious look.

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. Cache charges and external tools sit outside this example. The minimum top-up is $10 in prepaid credit, a funding minimum rather than a task fee or a subscription — see billing details.

Best model for Obsidian Copilot: choose per surface

Copilot 4.x does not expose temperature, top-p or similar tuning in its model lists. In Agent Chat the per-chat controls are the model, the effort level when the model reports one, and the permission mode; Quick Chat offers model selection only. That makes model choice the main quality lever — and makes it worth splitting by what you are doing rather than picking one winner.

Rates are live catalog prices and the pairings are a selection method, not a benchmark result — this page has run no benchmark on Obsidian vault work and cites none. Two constraints from the software itself. Which models you can pick at all is set by your backend: the Claude and Codex backends show their own CLI's list, and the documentation says the available lineups change so the lists in Copilot are the source of truth. And the documentation does not state a tool-calling capability requirement for Agent Chat, while BYOK model rows show capabilities only "when known" — an absence of documentation, not a guarantee either way, so if you plan to run an agent loop, confirm the model you pick actually handles tools before budgeting around it.

Cheapest API for Obsidian Copilot: two different questions

The cheapest listed rate and the cheapest way to finish the work separate as soon as an agent loop is involved. A Quick Chat ask that summarises one note is genuinely near-free on a small model, and there is no reason to pay more for it. An agent session that reads five notes, edits three and gets one edit wrong pays for the same context twice, so a model that needs two attempts at your work can cost more than one that needs a single attempt at three times the rate. That is a property of your vault and your prompts, not of a price list, which is why this page does not name a cheapest model.

If Claude is the family you have settled on, cheapest Claude API compares that lane specifically, and OpenAI-compatible API covers the wire format the BYOK tab expects.

Pointing Copilot at a gateway, and where it stops

There are two doors, and knowing which one your backend uses saves an evening. Door one is the BYOK tab, which works on the free tier: Settings → Copilot → BYOK → Add a provider → Add a custom provider, then a display name, an API key, a base URL, and at least one model either discovered or typed by hand. Discovery reads the endpoint's /v1/models response, and saving needs at least one model plus a routable endpoint. For Kunavo that base URL is https://api.kunavo.com/v1. Keys entered here are stored in the device keychain rather than the vault's data.json, so syncing a vault does not carry the credential to another machine.

Two caveats on that door. If Test passes but Quick Chat will not send, the documented fix is Enable CORS — and the same documentation says that with it on, responses arrive only after completion instead of streaming token by token. That trade is current in 4.x, not a legacy footnote. And a BYOK provider is guaranteed only in Quick Chat: inside Agent Chat it can appear under the opencode backend only, and the docs state that as a condition — "when the provider is shown there" — without giving a rule for which providers qualify. The Claude and Codex backends never list BYOK models at all, so there is nothing to test there. That path has not been runtime-tested here, so try one task before you move your workflow.

Door two is per-agent environment variables, which is the only documented hook for the Claude and Codex backends, because both explicitly ignore BYOK keys. Settings → Copilot → Basic → Agents → the agent → Environment variables → Add variable passes a name and value to that agent's process only; values are used literally, and the documentation warns not to treat the editor as a secret manager — so do not assume keychain storage for anything typed there. For the Claude backend the variables are Claude Code's own, and the more reliable place for them is Claude Code's settings file, because an Obsidian launched from the dock does not necessarily inherit the environment your shell profile sets:

{
  "env": {
    "ANTHROPIC_BASE_URL": "https://api.kunavo.com",
    "ANTHROPIC_AUTH_TOKEN": "sk-kn-..."
  }
}

ANTHROPIC_BASE_URL is an origin with no /v1 — Claude Code appends the route itself — and ANTHROPIC_AUTH_TOKEN sends the credential as a bearer token; the base URL page covers the failure modes. Anthropic's environment-variable reference documents two side effects of pointing that variable at a non-first-party host: MCP tool search is disabled by default unless you set ENABLE_TOOL_SEARCH=true, and Remote Control is disabled as of Claude Code v2.1.196. Whether the Claude backend inside Copilot picks these up was not tested here; the documentation only says it follows the CLI environment.

Kunavo publishes setup references for all three backends — opencode, Claude Code and Codex CLI. Those are configuration references, not compatibility tests: none of them has been runtime-tested inside Obsidian's plugin host. Keep a working route available, run one bounded task, then read the charge your account recorded for it. Create a Kunavo account when you are ready to fund a key.

Choosing the backend first? OpenCode pricing and Claude Code pricing cover what each one costs on its own, and best API for opencode goes deeper on the provider block that Copilot hands to opencode at startup.

FAQ

Plugin manifest, licence, release feed, vault-search, provider, settings, model and plan documentation, the official pricing page, miyo.md and obsidian.md/pricing all fetched September 19, 2026 and re-checked September 21, 2026; Anthropic's environment-variable reference was read September 21, 2026. Nothing on this page was runtime-tested inside Obsidian, and every route that depends on a live test says so. Kunavo token rates come from the live catalog; all dollar examples are illustrative token arithmetic rather than measured task costs.