Grandet: find the cheapest API for any model

简体中文 · Website · Installation

Grandet compares listed API prices for each model across official APIs, inference providers and relay sites. Your AI agent can look them up and connect the cheapest one to your app.

Install

Paste this into the agent you already use (Codex, Claude Code, Cursor and others):

Install Grandet Skill for me: https://grandet.ai/download.md

The agent does the rest. If it needs you to sign in or approve something, it tells you exactly what to click. Then open a new conversation and ask:

I code in Claude Code. Find me the cheapest Claude Sonnet 4.6 API.

No need to mention Grandet or attach files again. Price lookups need no account or API key. To install by hand, or for apps such as ChatGPT or Claude on the web that take an uploaded skill file, see Manual installation.

Related MCP server: TokenLab MCP Server

Install with git

The website and the install sentence above install build.20261018b (our internal name: j2), marked Latest on GitHub. Cloning its tag gives you the same Skill: the installed folder has the same content hash as the release ZIP.

git clone --depth 1 --branch skill-v0.3.0-build.20261018b https://github.com/hanzohanzhe/Grandet.git
cd Grandet
node scripts/install-price-finder.mjs --app claude-code --receipt-source checkout

• Replace claude-code with your app: codex, cursor, opencode, gemini-cli, cline, roo-code, kilo-code, windsurf, continue, hermes, openclaw, trae-cn or trae-intl. --list shows every app and the folder it installs to.

• The installer needs Node.js 20 or later. Without Node, copy the whole skills/grandet-api-price-finder folder into your app's skill folder: ~/.claude/skills/ for Claude Code, ~/.agents/skills/ for Codex, other apps in the per-app guide.

• Then open a new conversation and ask a price question. The last line of the installed SKILL.md names the build.

• The installer sends one anonymous install receipt (no path, name or key) so installs can be counted; --receipt-source checkout marks it as a git install. Add --no-telemetry to send nothing.

• Another version: put its tag from Versions after --branch and run the same install command. It replaces the installed copy and keeps the old one as a backup outside the skill folder. In a clone you already have: git fetch --depth 1 origin tag <tag>, then git checkout <tag>.

• Clone a tag, not main. main usually matches the latest release, but between releases it can carry changes that have not been tested yet.

MCP server

Add it to any MCP client that supports remote servers:

{ "mcpServers": { "grandet": { "url": "https://grandet.ai/mcp" } } }

In Claude Code: claude mcp add --transport http grandet https://grandet.ai/mcp

Tools: grandet_find_cheapest_api, grandet_find_free_credits, grandet_resolve_model, grandet_quote_evidence, grandet_glossary, grandet_list_apps.

What it does

1. Finds the lowest listed price. It compares the relay and official prices on grandet.ai for your model and kind of use (chat, coding, long output) and lists the cheapest few: the site, the group to pick when you create the token, the price, its unit and when it was observed.

• It ranks by listed price only. It never says which site is "good" and never repeats a site's own claims such as "full strength" or "stable".

• USD credit, CNY credit and cash prices are listed separately and never compared with each other.

• Conditions such as a minimum top-up or members-only access are shown as a one-line hint. Prices that contradict the site's own text are listed as a reference. Groups a site limits to one client are folded away when your app is a different one.

2. Finds free credit. Ask "is there anything free?" and it lists official free tiers, trial credit and relay sign-up, check-in and free-group offers, with whether a card, ID check or application is needed. If something you already said makes you eligible for a free offer, it mentions it once.

3. Connects it for you. When you say "connect it":

1. You sign up and top up on the site yourself, and create the token in the group it names.

2. You run one command in your own terminal to store the key. Typing is hidden and the key never enters the chat.

3. It shows the exact configuration change, backs up the file, and writes only after you confirm.

4. It checks the key, group and model name without spending credit.

5. Say "undo" to restore the previous setup.

For Codex, Claude Code, OpenCode, Kilo Code, Continue, Gemini CLI, Hermes, OpenClaw and WorkBuddy, the Skill's own connect program (scripts/connect.mjs, also usable as a local MCP server) makes the change the same way every time. Other apps get a field card and their documentation link. In a cloud chat, it can only hand you the command and configuration to run yourself.

Try asking

20 million tokens a month, a quarter input and the rest output, no cache. Which is cheapest, and what does it cost a month?

Is there a free Gemini API? I don't have a credit card.

Use the second one and connect it to Codex on Windows.

What do "官转" and "逆向" mean?

Answers include a link to the price sources. Listed prices and suppliers' claims say nothing about quality. Asked about quality, speed or downgraded models, it reports only what Grandet has on record and says so when there is no measurement.

What Grandet includes

Versions

Every build keeps product version 0.3.0; builds differ by build ID. The tag of a build is skill-v0.3.0-build.<build> (for example skill-v0.3.0-build.20261018b), and every tag and its release files stay on GitHub Releases, so any of them can be installed with the git commands above. The website installs 20261018b (j2).

From 20261017 on, each build was tested against the build before it: the same 48 questions, asked in the Hermes agent with deepseek-v4.1-flash and glm-5.3. PASS means every target was met. HOLD means some targets were missed; those builds were published after the misses were reviewed. "Used the Skill" counts sessions that answered through the Skill.

Which one to pick

• Most people: 20261018b (j2). It is what the website installs and has the best test results so far. When grandet.ai cannot be reached, it still answers from a price snapshot taken on 18 October 2026 and says the date first. The cost: the largest package (ZIP 522 KB, .skill 218 KB).

• No snapshot, smaller package: 20261018 (i8). ZIP 436 KB. It cannot answer when grandet.ai is unreachable, and in testing 1 of 16 group-condition sessions asked for a page the website has retired.

• Chinese conversations only, fewer tokens on relay-jargon questions: 20261017 (i5). The English glossary column added in 20261017b doubled the glossary file (33.9 KB → 64.2 KB). 20261017 has no English column, no WorkBuddy connection, no named-site comparison and no 30-second switch to SkillHub.

• Longer answers: 20261016h or earlier. These have no 8-line limit (median answer 841 characters in h, 484 in 20261017), but call addresses that don't exist far more often (18 bad-parameter errors in 48 questions, against 0). A small check of j2 against 20261016c (32 questions, deepseek-v4.1-flash) found: Skill used 90% → 100%, API errors per question 0.47 → 0.03, median answer 890 → 530 characters.

• Avoid on native Windows: anything before 20261016d. Its Node installer can fail with EPERM there; copy the folder by hand instead. Builds before 20261016e have no SkillHub download for mainland China, and builds before 20261016h have a description that weaker models pick up less often (47% vs 85% in an A/B test).

• Each build was written for the website's API at the time. The site has changed since, so older builds are more likely to ask for pages or fields that no longer exist.

All builds

skill-v0.3.1 was a candidate that was withdrawn before release; don't use it.

Feedback

Please open an issue if something goes wrong, with the app, version, system and the step where it stopped. Don't paste API keys or private conversations.

Data and developer entry points

Local development needs Node.js 20+ and no third-party packages:

node scripts/catalogue.mjs validate
node scripts/catalogue.mjs generate --check
node scripts/validate-public.mjs
node --test tests/*.test.mjs
node scripts/serve.mjs

serve.mjs opens a local browser view at http://127.0.0.1:8787/. These commands check local files only and never update the website's prices.

Sourced corrections and improvements are welcome; see the contributing guide and methodology.

Friends

• LINUX DO: Chinese-language tech community

Licence

• Skill files: Apache-2.0.

• Catalogue code and documents written by Grandet: AGPL-3.0-only; scope in LICENSE.

• Supplier data, original text and trademarks belong to their owners; the factual compilation carries no blanket open-data licence. See DATA-LICENSE.md, TRADEMARKS.md and each release manifest.