The best API for Cherry Studio is whichever provider answers on the protocols your features need, because Cherry Studio itself is free and charges nothing: it is AGPL-3.0 open source, and every cent you spend is billed by the provider behind the endpoint box. Cherry's own API-gateway documentation puts it plainly — "usage and charges are still generated by the actual model provider" — so "Cherry Studio pricing" is really a question about token rates, multiplied by a handful of defaults that quietly fire more requests than you asked for.
Four things share the name and none of them is the free desktop client. Cherry Studio Enterprise is a separate, privately deployable product the README sells by quote. CherryAI is the built-in free provider. CherryIN is the paid gateway Cherry Studio ships as a built-in provider, at open.cherryin.ai. And the "Cherry Studio" listing on the Apple App Store (id 6476445971) is a Health & Fitness app from an unrelated developer, so do not treat it as the mobile version. This page is only about the desktop client from CherryHQ/cherry-studio.
One version note before any tutorial you follow, including older ones on this site. The repository is live and not archived, carries the AGPL-3.0 licence and more than 52,000 stars, and was last pushed on September 21, 2026 (GitHub API, checked September 21, 2026). v2.0.0 shipped August 5, 2026 — two days after the final v1 release — and v2.1.0 on September 18, 2026 (releases API, same check; the last v1 release, v1.9.13, is dated August 3, 2026). The custom-provider dialog changed in v2 — a single "Type: OpenAI" dropdown became independent per-protocol base URLs. Screenshots and field names from a v1-era walkthrough no longer match.
Cherry Studio pricing: the software line and the model line
| Line item | What it costs | Source |
|---|---|---|
| Cherry Studio Community Edition (Windows, macOS, Linux) | $0, AGPL-3.0 | Repository licence field and the README |
| Commercial use of the community edition | $0 if you comply with AGPL-3.0; otherwise a commercial licence by contact | README licence section |
| CherryAI, the built-in free provider | $0; quotas and rate limits are not published | CherryAI provider doc |
| Every model call, including through Cherry's local API gateway | Your provider's per-token or per-image rate | API gateway doc |
| Cherry Studio Enterprise | A different product; no price is published — see below | README comparison table and the enterprise documentation |
The enterprise line is where the search results mislead, because the highest-ranking official page for "cherry studio pricing" in the results sampled for this page is the enterprise documentation — and it publishes no plan price at all. Its body renders through JavaScript; a plain fetch on September 21, 2026 returned only the page's own description line, which says subscription-based plans plus token-based usage, supporting Alipay and bank transfer. The README's edition comparison is no more specific, giving the enterprise cost as "Buyout / Subscription Fee" against the community edition's AGPL-3.0 licence, and directing anyone interested to request a quote. The enterprise marketing site is a JavaScript application whose figures this page could not read, so no amount from it is quoted here. If enterprise is what you want, get the quote in writing.
For everyone else, the client is free and the question is what the calls cost.
The defaults that bill more than you asked for
This is the part a price table cannot show you. Cherry Studio bills several model roles independently, and its history-compression default quietly puts your most expensive model on work that does not need it.
| What fires a request | Default | Effect on the bill |
|---|---|---|
| Comparing models in one message | You choose how many | "Selecting multiple models at once will send separate requests" — three models is three bills for one question |
| Quick model | Set in Settings → Default Model | Conversation naming and search-keyword extraction; Cherry's doc recommends a "cheap and fast" lightweight model, not a reasoning one |
| Translation model | Set separately | Message and page translation bills on its own, not against your chat model |
| Painting model | Set separately | The default for image generation, billed by whichever provider serves it |
| Automatic history compression | Enabled, compression model "Follow the current model" | Summarising earlier history bills at your main model's rate unless you point it elsewhere |
| Model call retry | Off; when on, up to 3 attempts with exponential backoff plus ordered fallback models | Retry and fallback only take effect before output begins, so each extra attempt is another request sent to your provider |
| Tool output truncation | 50,000 characters | Caps how much tool output can be pushed into the prompt in one go |
Read from Cherry Studio's default models, context and compression and model comparison pages, checked September 21, 2026. Cherry's own advice on the comparison feature is worth repeating: when fees, speed or sensitive data are involved, confirm with a short question first, then send the long material.
The in-app cost figure is an estimate, and it reads high on a discounted gateway
Cherry Studio's Usage Statistics panel shows a total cost, a cache hit rate, and a breakdown by provider, model, API key and assistant over the last 30 days, 90 days or year. That total is not a reading of your account. The documentation states that cost "is an estimate: it is converted according to the model's public pricing", with actual billing still coming from each model provider, and that free and local models incur no charge.
The consequence is specific and easy to miss: if you buy the same model below its public list price, Cherry Studio will systematically over-report what you spent. The fix is in the model list — the custom-provider doc says you can open a model's settings and edit its name, type and price, so entering your actual rates makes the panel usable again. The cache hit rate is the other number worth watching, which the Usage Statistics page above defines as cache reads hit divided by observable input; on a route that supports prompt caching, that ratio is where repeated long contexts get cheap.
A worked estimate for two Cherry Studio turn shapes
These are illustrative token arithmetic, not measured task costs and not a bill ceiling. Assume a short question at 1,500 input and 600 output tokens, and a document turn at 120,000 uncached input and 1,500 output tokens — roughly one long paper pasted into the context. Rates are live Kunavo catalog prices per million tokens.
| Model | Input / output per 1M | Short question | Document turn |
|---|---|---|---|
| GPT-5.6 Luna | $0.07 / $0.42 | $0.0004 | $0.009 |
| Gemini 2.5 Flash | $0.09 / $0.75 | $0.0006 | $0.012 |
| Claude Haiku 4.5 | $0.40 / $2.00 | $0.0018 | $0.051 |
| GPT-5.6 Terra | $0.70 / $4.20 | $0.0036 | $0.090 |
| Gemini 3.1 Pro | $0.70 / $4.20 | $0.0036 | $0.090 |
| Claude Sonnet 5 | $2.00 / $10.00 | $0.0090 | $0.255 |
| Claude Opus 5 | $2.00 / $10.00 | $0.0090 | $0.255 |
Two readings that matter more than the ranking. First, the fan-out: asking that same document question of Claude Sonnet 5, Gemini 3.1 Pro and GPT-5.6 Terra together is not one bill but three, $0.436 under these assumptions — which is the whole point of Cherry's advice to test with a short question first. Second, compression: an assumed summarisation pass over 60,000 input and 1,200 output tokens costs $0.1320 when it follows Claude Sonnet 5, against $0.0063 pointed at Gemini 2.5 Flash. Neither number is large alone; both repeat all day.
Scale by your own conversations before treating this 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. Cache charges and image generation sit outside this example. The minimum top-up is $10 in prepaid credit, which is a funding minimum rather than a task fee or a subscription — see billing details.
Best API for Cherry Studio: which route wins when
| Route | Wins when | What you give up |
|---|---|---|
| Direct vendor API | You live in one vendor's flagship and want its own caching and batch terms | A second vendor means a second provider entry and a second balance |
| CherryIN, the gateway Cherry Studio ships with | You want the built-in paid provider, whose published model list carries both OpenAI and Anthropic endpoint types, with no third-party setup | On the Claude models compared below, its published rates match the vendor list rates |
| A third-party OpenAI- and Anthropic-compatible gateway | You switch models per task, want one key and one balance, and want rates below list | You configure it by hand, and the in-app cost estimate reads high until you edit each model's price |
| CherryAI, the built-in free provider | Casual chat with nothing to configure and nothing to pay | Undisclosed lineup, quotas and rate limits, all server-controlled and changed at Cherry's discretion |
| A local model | Small, private or offline work with no per-request charge | Capability gap against hosted frontier models, and your own hardware |
The Claude comparison is the one that is double-sourced, so it is the one this page prints. CherryIN's figures are computed from its own public pricing data at the single default group, whose ratio is 1 and applies no discount, and for all four models they match the vendor rate exactly.
| Model | Kunavo catalog, in / out per 1M | CherryIN published rate | Anthropic published rate |
|---|---|---|---|
| Claude Opus 5 | $2.00 / $10.00 | $5.00 / $25.00 | $5.00 / $25.00 |
| Claude Sonnet 5 | $2.00 / $10.00 | $2.00 / $10.00 | $2.00 / $10.00 |
| Claude Sonnet 4.6 | $1.20 / $6.00 | $3.00 / $15.00 | $3.00 / $15.00 |
| Claude Haiku 4.5 | $0.40 / $2.00 | $1.00 / $5.00 | $1.00 / $5.00 |
Note the honest row: Claude Sonnet 5 is at parity across all three columns, so none of these three routes undercuts the others on that model — the gap opens on Claude Opus 5, Claude Sonnet 4.6 and Claude Haiku 4.5. Sources: CherryIN pricing (computed from its public pricing data, default group, ratio 1) and claude.com/pricing, both read September 21, 2026 — all four figures in the Anthropic column taken directly from that page. CherryIN's rates for OpenAI and Google models are not reproduced here because they could not be reconciled against those vendors' own published pages. For the wider gateway comparison see best LLM gateway and OpenRouter alternatives.
Best model for Cherry Studio, role by role
Because Cherry Studio bills roles separately, the useful answer is an assignment rather than a single name. Put the frontier model where judgment happens and nothing else: Claude Opus 5 or Claude Sonnet 5 for the main assistant, Gemini 3.1 Pro or GPT-5.6 Terra as a mid-priced alternative. Put a cheap, fast, non-reasoning model in the Quick slot — GPT-5.6 Luna, Gemini 2.5 Flash or Claude Haiku 4.5 — since it only names conversations and extracts search keywords. Give translation the same treatment, and point automatic compression at the cheap lane instead of letting it follow your main model. Opus vs Sonnet vs Haiku covers where that judgment line usually falls.
Two traps when you pick. First, the model ID: Cherry Studio no longer merges similarly-named models and shows the provider-returned API ID beneath the display name, precisely so you do not pick a date-stamped or vendor-prefixed twin — copy the exact slug, with dashes, as your provider writes it. Second, the knowledge base, where the usual warning is stated too strongly: Cherry's create page starts the embedding field at "Do Not Use" and its retrieval settings page states that the knowledge base still works without one, on BM25 keyword retrieval — an embedding model is what adds vector retrieval alongside it. Kunavo does not serve an embedding model, so that vector half comes from a local model downloaded in Cherry Studio or from a second provider. LibreChat vs Open WebUI compares two clients that separate retrieval models from chat models the same way.
What to paste in each endpoint box
In v2 the custom-provider dialog takes a provider name, an API key, and independent base URLs per protocol: OpenAI and Anthropic by default, with OpenAI Responses, Gemini, image generation and image editing behind an extra-endpoint toggle. Two official sources name that toggle differently — the documentation page calls it More Settings, while the v2.1.0 client itself labels it 更多端点 (More Endpoints) — so read the label in your own build rather than either page. Cherry Studio builds the final URL itself — its v2.1.0 source normalises the host, appends /v1 unless the address already carries a version segment or ends in #, then appends the protocol path. So https://api.kunavo.com and https://api.kunavo.com/v1 land in the same place; a trailing # is what breaks it, by switching off the version append entirely.
| Endpoint box | Path Cherry Studio appends | Kunavo |
|---|---|---|
| OpenAI | /chat/completions | Served |
| Anthropic | /messages | Served |
| OpenAI Responses (更多端点) | /responses | Served |
| Image generation (更多端点) | /images/generations | Served |
| Image editing (更多端点) | /images/edits | Served |
| Gemini (更多端点) | /models/{model}:generateContent | Not exposed — leave empty and reach Gemini models through the OpenAI box |
Paths read from the v2.1.0 source on September 21, 2026. Fill both the OpenAI and the Anthropic box if you can. Cherry's Coding Partner page — which launches Claude Code, OpenAI Codex, Gemini CLI, OpenCode and others against a Cherry Studio provider — states that it filters the list by the interface type each CLI needs, and that incompatible providers are not listed (checked September 21, 2026). The Agent runtime goes through the same local gateway, which Cherry's own documentation calls an internal service required for Agent operation, and its agent-creation page says the model selector filters out models incompatible with the run mode you chose. Which of your models survive that filter was not tested here, so configure both boxes rather than betting on one protocol.
Three operational notes from the provider documentation, checked September 21, 2026. The Enable switch is the single most common "my key does not work": a configured provider stays disabled and its models never appear in the picker until you turn it on. Get Model List fetches the provider's own model list with the key you just typed, so an empty list usually points at the key or the address rather than at Cherry Studio. And one provider can hold several keys, used in rotation front to back, which is a practical way to spread rate limits.
Set it up, then check the first charge
Kunavo publishes a configuration reference for this client. That is a published setup guide, not a compatibility test — Cherry Studio has not been runtime-tested here against Kunavo's endpoint, and everything above was read from Cherry's own documentation, its v2.1.0 source and the public pricing data named inline. Note that the setup doc was written against the v1 dialog and describes a single Type field; the v2 flow above is the current one. Keep a working route available while you try it, run one bounded task, then compare Cherry Studio's estimate against what your account actually recorded in usage. Start at the Cherry Studio integration guide, and create a Kunavo account when you are ready to fund a key.
FAQ
How much does Cherry Studio cost?
The Cherry Studio desktop client costs nothing. It is open source under AGPL-3.0, the CherryHQ/cherry-studio repository carries that licence, and the README states that commercial use is permitted subject to full compliance with AGPL-3.0, with a separate commercial licence available by contacting the project. What you pay is the model bill from whichever provider you configure inside it. Cherry Studio Enterprise is a different, privately deployable product sold by quote, and its terms say nothing about what an individual running the free desktop client pays.
What is the best API for Cherry Studio?
The best API for Cherry Studio is the one that covers the protocols the features you use actually need, at a rate you accept. Cherry Studio v2 configures a custom provider as independent per-protocol base URLs — OpenAI and Anthropic by default, plus OpenAI Responses, Gemini, image generation and image editing behind the dialog's extra-endpoint toggle — so a provider that only answers on chat completions will light up the chat window and can leave Agent and coding-CLI features filtered out of the picker. A provider that answers on both OpenAI chat completions and Anthropic messages covers the widest surface. Rate matters second: Cherry Studio fires one billed request per selected model when you compare models, so a per-token difference is multiplied by how many models you fan out to.
What is the cheapest API for Cherry Studio?
Cheapest listed rate and cheapest way to finish the work are different questions, and in Cherry Studio the gap is unusually wide because one question can generate several billed requests. Before comparing per-token rates, set the Quick model to a cheap lightweight model, point automatic history compression at a cheap model instead of letting it follow your main one, and remember that selecting three models in the compare view bills three requests. Cherry Studio's own built-in CherryAI provider is free, but its documentation states that the model lineup, quotas and rate limits are delivered in real time by Cherry's server and adjusted from time to time, so it cannot be budgeted against. No provider should be described as absolutely the lowest price; compare the routes on your own workload.
What is the best model for Cherry Studio?
There is no single best model, because Cherry Studio bills at least four roles separately. The main chat or assistant model is where a frontier model earns its rate. The Quick model handles conversation naming and search-keyword extraction, and Cherry Studio's own documentation recommends a cheap, fast, non-reasoning model there. The translation model and the painting model are set independently and should be chosen on the same cost logic. One practical trap: models with similar names are no longer merged, and the real API model ID is shown beneath the display name, so copy the exact slug rather than the label.
Why does Cherry Studio's cost display not match what my provider charged?
Because it is an estimate computed from public list prices, not a reading of your account. Cherry Studio's Usage Statistics documentation says plainly that cost "is an estimate: it is converted according to the model's public pricing", with actual billing still coming from each model provider. If you buy through a discounted gateway, the in-app figure will therefore read high. The fix is in the model list: Cherry Studio lets you open a model's settings and edit its name, type and price, so entering your real per-token rates brings the estimate back in line. Reconcile against your provider's own usage record either way.
Can Cherry Studio's knowledge base run on a chat-only provider?
Yes, with a caveat about retrieval quality. Cherry Studio's own create-a-knowledge-base page tells first-time users to set the embedding model to [Do Not Use], and its model-and-retrieval page states that the knowledge base still works without one, on BM25 keyword retrieval; what an embedding model adds is vector retrieval running alongside BM25. If you do want that, the documentation gives two routes — a cloud embedding model from a service that serves one, or a local embedding model downloaded under Settings then Local Models. Kunavo serves no embedding model, so the vector half of that setup has to come from a local model or from a second provider.
Does a Claude Pro or Max subscription work inside Cherry Studio?
Not as a provider. Claude Pro and Max are chat subscriptions and issue no API key, so there is nothing to paste into a Cherry Studio provider entry; Anthropic publishes Pro at $17 per month on an annual subscription billed at $200 up front, or $20 billed monthly, and Max from $100 per month (claude.com/pricing, checked September 21, 2026). The route Cherry's documentation does describe is the Coding Partner page, which launches an external CLI such as Claude Code that signs in on its own account — and then only inside that CLI, not in Cherry Studio's chat window.
Checked September 21, 2026: the CherryHQ/cherry-studio repository and releases API, the README, the Cherry Studio documentation pages linked above, CherryIN's public pricing data, claude.com's published subscription and API rates, and Cherry Studio's v2.1.0 endpoint-path source. No price is quoted for Cherry Studio Enterprise, because neither official page publishes one that a plain fetch can read. Kunavo token rates come from the live catalog; every dollar example on this page is illustrative token arithmetic rather than a measured task cost, and nothing here was runtime-tested against Kunavo's endpoint.