Help center
Support
Most questions are answered faster by the docs than by us. When they are not, here is how to reach a person.
Try these first#
- FAQ — the questions that come up most often.
- Troubleshooting — an ordered checklist for a failing request.
- Error codes — what each status means and whether retrying helps.
Getting in touch#
Contact channels are listed in the contact page, and account-specific issues are fastest through the console, where your account is already identified.
What to include#
The difference between a one-reply resolution and a four-message exchange is almost entirely what is in the first message.
- What you expected and what happened. One sentence each.
- The exact model ID, including the vendor prefix.
- Timestamp with timezone, and how often it occurs — once, or every request.
- The full error body:
type,codeand message. - Whether curl reproduces it. This single fact rules out half the possible causes.
- Any proxy or gateway between you and the API.
Never send an API key
Not in a message, not in a screenshot, not in a request dump. If a key has appeared anywhere it should not, revoke it in the console first — that takes effect within about a minute — and investigate afterwards.
Billing questions#
Usage is broken down per key, model and day in the console, which answers most billing questions directly. If a charge still looks wrong, include the date range and the key name, and read Pricing & billing first — the two things that most often look like errors are reasoning tokens counting as output, and conversation history being resent on every turn.
A payment did not go through
Payments are handled by an external provider that runs its own fraud screening. A declined attempt means no money was taken — the charge you may see on a card statement is usually a temporary authorisation that releases on its own. Trying a different card or payment method is the fastest resolution.
Reporting a security issue#
Report vulnerabilities privately rather than in a public channel, and give us a reasonable window to fix the issue before disclosing it. Please do not test against other people's accounts or data — use your own.
Checking for an outage#
The status page is the first place to look. Beyond that,GET /v1/modelsis public and unmetered, so it is a quick way to tell whether the API is reachable at all. A failure there points at network or infrastructure; a failure only on one model points at that model's upstream, which routing usually handles by failing over before you ever see it.