FAQ

Accounts, security, billing, compatibility, troubleshooting. Check here first — 90% of issues are solved in under a minute.

Updated 2026-04-17

Accounts & keys

Can I recover a lost API key?

No. A key is shown in plaintext only once at creation; afterwards only a hash is stored. If lost, just revoke the old key → create a new one in the console — two steps.

How many keys can one account create?

20 by default; you can request more. We suggest splitting by purpose: cursor-mac, claude-code-work, ci-prod, etc. for easier auditing and revocation.

How do I cap a key’s spending?

In the console, API Keys → Edit lets you set per key:

• Daily cap (USD)
• Monthly cap (USD)
• A model allowlist

On overage that key is paused automatically, without affecting others.

Can I delete my account?

Yes. Settings → Account security → Delete account in the console; a soft delete is reversible for 30 days, after which it is physically deleted per the GDPR process. See the Privacy Policy.

Billing

How is cost calculated?

Pay-as-you-go = (input tokens × input price + output tokens × output price) × model multiplier. Multipliers are between 0.3x and 0.4x, below official. See the pricing page.

How much should my first top-up be?

From $1. We suggest $2–10 first to test the integration, then top up in bulk once your toolchain works.

Can you issue invoices?

We do not issue Chinese VAT invoices (neither general nor special). If genuinely needed, we can provide an overseas proforma invoice on request; a proforma invoice is not a Chinese tax invoice and cannot be used for domestic VAT deduction or reimbursement.

What are the refund rules?

• Credits: non-refundable on a no-reason basis (including unused balance)
• Used portion: non-refundable
• Passive refunds: processed only where required by law or after a major platform incident (see the Refund Policy)
• Losses from platform issues: compensated to your balance via a reviewed claim (public rules)

See the Refund Policy and Refund rules as product.

Compatibility

Do you support the OpenAI SDK?

Yes. All OpenAI endpoints — /v1/chat/completions, /v1/embeddings, etc. — are fully compatible; just set base_url = "https://api.staging.cosmos-router.com/v1".

Do you support the official Anthropic SDK?

Yes. The anthropic Python / Node SDK works after changing two environment variables:

export ANTHROPIC_BASE_URL="https://api.staging.cosmos-router.com"
export ANTHROPIC_API_KEY="sk-xxxxx"

Do you support function calling / tool use?

Yes. OpenAI’s tools parameter and Anthropic’s tools are both passed through to the upstream unchanged.

Do you support streaming / SSE?

Yes. stream: true behaves exactly as in the official APIs.

Do you support prompt cache?

Yes. OpenAI’s automatic caching and Anthropic’s explicit cache_control are both billed at official rates (cache hit = 10%–50% of the input price).

Troubleshooting

The call returns 401 Unauthorized

• Check the key starts with sk- (not an official sk-ant- key)
• Check the key isn’t revoked (see its status on the console API Keys page)
• Check the header format: Authorization: Bearer sk-xxxxx (note the space after Bearer )

The call returns 402 Insufficient Quota

Insufficient balance, or the key hit its daily/monthly cap. Top up in the console or raise the cap.

The call returns 429 Rate Limit

• Default limit is 600 RPM / 300k TPM
• The client should retry with exponential backoff (the Retry-After header gives the seconds)
• For high-frequency use, request a higher limit in the console

Calls are slow or time out

• Check the status page (/en/status) for an ongoing failover
• 1–3 seconds to first token on large Claude / GPT models is normal
• If it persists > 10 seconds, check the upstream response time for that request under Recent calls in the console — the upstream may be queuing

The response differs from official

ZetaAPI passes requests and responses through verbatim — no prompt rewriting, no result filtering. If you see a difference, it’s almost always:

1. A model-version mismatch (gpt-4o has multiple snapshots)
2. Different temperature / top_p, etc.
3. Randomness (fix the seed to reproduce)

If you still suspect a proxy issue, open a ticket with the request ID (the x-request-id response header).

Security

Are my prompts and responses logged?

• Metadata (model name, token counts, latency, status code) is kept for billing and monitoring for 90 days
• Prompt and response content is not logged by default; only when you explicitly enable “debug logs” is it kept for 7 days
• Staff cannot access user content, and the audit log is public

Is traffic encrypted?

End-to-end TLS 1.3 — HTTPS enforced both to the client and to the upstream.

What if my key leaks?

Immediately revoke the key on the console API Keys page (effective instantly, globally invalid within 10 seconds). The platform also monitors for keys exposed on GitHub and proactively notifies you.

Still no answer?

• Browse the API Reference for technical details
• Check the Changelog for a recent change
• Open a ticket from the bottom-right of the console; we respond within 2 business hours