Troubleshooting

“DeepSeek / MiniMax не отвечают”

  1. Open Status. If error rates are high, the network is busy — not your key.
  2. Confirm the id is still on GET /v1/models.
  3. Try zai-org/GLM-5.3-Flash.
  4. Retry with backoff. Do not 404-storm an unlisted model.

Healthy shards on the capacity table can still 429 when load is 100%.

401

Wrong or revoked key. Mint a new one. Do not send the dashboard JWT to /v1.

402

Free grant gone or USD balance empty. Top up. Over-limit images also 402.

404 model_not_found

Typo, deprecated id, or the model was unlisted. Refresh /v1/models.

429

Host concurrency full. Back off or enable persist through rate limits and raise the client timeout to 120s+.

502 / timeout / empty body

Upstream miss, fallback miss, or the client gave up. Raise timeouts for fat prompts. Check Status. See Errors.

Tool calls look like raw JSON

Enable reasoning compatibility for DeepSeek. Confirm the model supports tools.

MCP does nothing on the API

Settings → MCP tools is off. Chat MCP can still work. Local npx servers never run on the proxy. See MCP.

Files ignored

Settings → Convert files and images is off.

Dashboard shows no usage

You are on the wrong account, privacy hid text (counts should still exist), or you looked before the request finished. See Usage.

SDK strips web_search / MCP tools

Use extra_body in the official OpenAI SDK. See OpenAI SDK.