Frequently Asked Questions
Common questions about FastMetal — pricing, models, the API, and how we handle your data.
Getting started
What is FastMetal?
FastMetal is an AI gateway that puts models from many different providers behind a single API key and a single endpoint. Claude, Gemini, Grok, DeepSeek, Qwen, Kimi, GLM, MiniMax and others are all reachable the same way, alongside image and video generation models.
The API is OpenAI-compatible, so existing code works after changing the base URL and the key. Credits are prepaid — there is no subscription, no monthly minimum, and no seat count.
How do I get started?
Two steps. Create an account — an API key is provisioned for you automatically — then add credit from the dashboard. You can send requests the moment the payment clears.
Card payments post to your balance immediately.
Is there a free trial?
There is no blanket signup credit, but a handful of models are free to call, so you can evaluate the platform before spending anything.
Why use FastMetal instead of going direct to each provider?
Three reasons. First, one contract and one invoice instead of a separate account, payment method, and billing cycle per provider.
Second, switching models is a string change. You are not reimplementing a new SDK and a new auth scheme every time you want to try something else.
Third, billing is denominated in Japanese yen rather than USD, which matters if you are budgeting or expensing in yen and would otherwise absorb the FX conversion on every provider invoice.
How do I get support?
Use the form on our contact page, or email [email protected]. Support hours are 9:00 to 17:00 Japan time.
For API questions, the guides in these docs usually get you there faster — most integration issues come down to the base URL or the key.
Pricing and billing
How does pricing work?
Text models are billed per token, with separate input and output rates. The pricing page lists the rate per million tokens for every model we carry.
Image generation is a flat rate per image and video generation is a flat rate per clip. Every request records its actual cost, which you can review on the usage page.
Are there platform fees or monthly costs?
No. There is no monthly base fee, no minimum spend, and no surcharge on topping up your balance. You pay the listed per-token rate for what you actually use.
If you pay by bank transfer, your bank's transfer fee is yours to cover.
What payment methods do you accept?
Credit card, purchased from the dashboard. Payments are processed by Stripe, so card details never reach FastMetal's servers.
You choose the top-up amount. When the balance runs down, top up again the same way.
Do credits expire?
Yes — prepaid credits carry an expiry date. If you want to keep using the service past that point, buy additional credits.
The full terms are set out in our commercial transaction disclosure.
Can I get a refund?
Because credits are consumable and delivery begins immediately, cancellations and refunds after service has started are generally not available.
Where we do issue a refund, the bank transfer fee is deducted from the refunded amount. See the commercial transaction disclosure for the full terms.
Can you issue a qualified invoice for Japanese consumption tax?
Yes. FastMetal is operated by a registered qualified invoice issuer under Japan's invoice system. The registration number is T5010001015990.
The registered entity name and address are on the company information page.
How do I check my balance and usage?
Your current balance is on the dashboard. The usage page breaks spend down by model and by day.
To read it programmatically, call the balance endpoint with your API key. It returns the budget ceiling and the amount spent — the difference is what you have left.
What happens when I run out of credit?
Further requests are rejected with a budget_exceeded error, returned as HTTP 429. Requests simply stop; nothing is billed beyond your balance and there is no overage invoice.
Top up and you can resume immediately. Because budgets are tracked per key, exhausting one key does not affect the others.
Models
Which models are available?
A broad catalog covering Anthropic Claude, Google Gemini, xAI Grok, DeepSeek, Alibaba Qwen, Moonshot Kimi, Z.ai GLM, MiniMax and GPT-OSS, plus image and video generation models.
The models page lists everything currently available with live per-token rates. Programmatically, call the /models endpoint.
Which models are best for Japanese?
We carry llm-jp-3.1-8x13b-instruct4, a domestically developed model from the LLM-jp project trained primarily on Japanese data.
That said, the large multilingual models — Claude, Gemini, GLM — also perform strongly in Japanese, and often better on reasoning-heavy work. Which one wins depends on the task, so the model comparison page is the reliable way to decide: it shows real output from each model on the same prompt.
How often do you add new models?
We add notable releases on an ongoing basis, once we have verified that the route works and the pricing is correct. New additions are announced on the blog.
If a model you need is missing, send us a request from the contact page.
Do you support image and video generation?
Both. Image generation is billed at a flat rate per image and uses the same API key as text.
Video generation is an asynchronous job that turns a text prompt into a short clip, billed at a flat rate per clip. Polling for completion and downloading the result cost nothing extra. See the video generation guide for details.
Can I send images as input?
Yes, on models that support vision. Use the standard OpenAI-compatible format and pass either an image URL or base64-encoded data.
Support varies by model — and a model being advertised as vision-capable upstream does not always mean the route serves image input. Check the models page before you build against it.
How do I choose a model?
Start cheap and fast, then move up only where quality actually falls short. Switching is a one-line change, so there is little cost to being wrong initially.
The model comparison page shows how each model answers the same prompt, which is usually more informative than benchmark scores for deciding what fits your workload.
API and technical details
Will my existing OpenAI code work?
Yes. Point the base URL at https://api.fastmetal.ai/v1 and swap in your FastMetal key. Nothing else needs to change.
Every official OpenAI client library — Python, Node.js, Go, Ruby and the rest — supports overriding the base URL, so this is a configuration change rather than a code change.
Do you support the Anthropic Messages API?
Yes. Set the base URL to https://api.fastmetal.ai/anthropic/v1 and you get /messages and /messages/count_tokens.
If you are on the official Anthropic SDK, changing the base URL and the key is the whole migration.
Is streaming supported?
Yes. Set stream: true on the request and tokens arrive incrementally as server-sent events.
Use it for long responses and for anything rendering into a chat interface.
What are the rate limits?
We do not impose a fixed requests-per-minute cap. The practical limit is your balance: while there is credit on the key, requests go through.
Upstream providers do apply their own limits. When one of those trips you will see an HTTP 429, and backing off briefly before retrying resolves it.
Does FastMetal work with coding agents?
Yes — Claude Code, Hermes, Crush, opencode, Octofriend, GitHub Copilot and Xcode all work, along with any other tool that lets you set an OpenAI-compatible or Anthropic-compatible endpoint.
The Agents section of these docs has per-tool setup instructions.
Do you have an MCP server?
Yes, at https://mcp.fastmetal.ai/mcp. Any agent that speaks the Model Context Protocol can call FastMetal models as tools, authenticating with your normal API key as a bearer token.
It is useful for things a base URL cannot do: getting a second opinion from a different model without leaving your session, or generating an image mid-task.
Privacy and data handling
Do you log my prompts and completions?
No. Message-body logging is switched off on the production gateway, so prompt and completion content is not recorded on our side.
What we do record is the metadata billing and operations require: model name, input and output token counts, latency, cost, and timestamp.
Is my data used to train models?
FastMetal does not use your data to train models.
Inference itself runs on each model provider's infrastructure, so their own policies govern what happens there. If you are working with sensitive data, get in touch from the contact page first and we can talk through which routes suit your requirements.
Where is FastMetal hosted?
The gateway and databases run in AWS Tokyo (ap-northeast-1). Account data and usage history stay in Japan.
Inference runs on each model provider's own infrastructure, which is located wherever that provider operates.
Accounts and API keys
Can I create more than one API key?
Yes, as many as you need, each with its own independent balance and budget ceiling.
Separate keys per environment or per project make it straightforward to see where spend is actually going, and to cap any one of them.
What should I do if a key leaks?
Delete the key from the dashboard. Requests using it stop being accepted the moment it is removed. Then issue a replacement and roll it out.
This is the practical argument for per-key budgets: a leaked key can only ever cost you what was loaded onto it.
Is there a referral program?
Yes. When someone signs up through the referral link on your dashboard and verifies their email, you receive 1,000 yen in credit. They receive 500 yen when they generate their first API key.
Referral rewards are capped at 20 successful referrals per account.