Questions and next steps
XiuRouter questions about APIs, keys, models, pricing, usage, and failed requests
Check current models and prices on public pages, then continue in the console or docs.
Protocols and API keys
Choose the client protocol first, then confirm the key’s service tier and model scope.
Should I choose Responses, Chat Completions, Messages, or Gemini?
Choose the protocol the client actually sends, not one guessed from the model brand. Codex currently uses Responses, Claude Code uses Anthropic Messages, tools limited to OpenAI-compatible chat use Chat Completions, and native Gemini SDKs use GenerateContent. Confirm against the protocol page and the target client guide.
Can one API key be used across protocols?
A key can be used for the models and service tier it is allowed to access, but each protocol also depends on the endpoint types currently declared by the target model. Confirm the model scope when creating the key, then verify with a small request on the matching protocol.
Should API requests use router.xiu.ai or router-api.xiu.ai?
New SDK and client configurations should use the router-api.xiu.ai address shown in Docs. router.xiu.ai hosts the public site and console and retains selected legacy-compatible routes, but it is not the default API origin for new integrations.
Where do I create an API key and verify the first request?
Create and save the key on the console’s API keys page, then confirm the model ID and service tier on the live Models and pricing page. The quickstart maintains the exact request parameters, examples, and verification steps.
Models and service tiers
Model availability changes over time. Check Models and pricing for the current list.
Which models does XiuRouter currently support?
Use the current catalog on Models and pricing. It shows current model IDs, service tiers, supported API formats, and prices.
Does a service tier represent speed, quality, or reliability?
A service tier determines the current model and price source. Its name is not a speed, quality, reliability, failover, or SLA promise. Evaluate the target model with a real request and its resulting usage record.
Where do I get the model ID?
Copy the exact model ID from the current Models and pricing page. Clients that support discovery can also read /v1/models. Do not construct an ID from a marketing name.
Why can a request fail even when the model is visible in the catalog?
Catalog visibility means the current public catalog can describe the model. A request still depends on the protocol, key status, model scope, service tier, balance, and the upstream response at that moment. Check the request record rather than treating visibility as permanent availability.
Pricing and usage
Check current unit prices separately from costs already recorded in Usage.
Where do I see current prices?
Models and pricing reads the current public pricing response and presents it by model and service tier. Use /api/pricing for machine-readable data; content pages do not copy a fixed price list.
How do input, output, cache-read, and cache-write prices differ?
They are separate per-1M-token rates for each usage type. Only usage produced by the request is billed, and models billed per request are marked separately.
What is the difference between a reference price and a XiuRouter tier price?
A reference price is a comparison point for the same model, not the XiuRouter rate. Your recorded cost uses the model and service tier that handled the request plus the applicable input, output, and cache rates.
Why can the same model cost differently across service tiers?
Each service tier has its own model availability and rates. If a key uses automatic routing or cross-tier retry, the tier that handles the request determines the price. Check the actual tier and recorded cost in Usage.
How do I reconcile tokens and cost for one request?
Find the request in Usage by API key, time, or request ID. Check the model, service tier, input, output, cache usage, status, and recorded cost. Models and pricing shows the unit rates; Usage shows what the request used.
Common request failures
Keep the error, time, and request ID before deciding whether the issue is the protocol, key, model, or upstream.
Why check the Base URL and path first for a 404?
Different SDKs append different paths. A common failure is putting /v1 in the Base URL when the client adds another /v1, or mixing Responses, Messages, Chat, and Gemini routes. Check the final request URL against the protocol page.
What should I do for model not found or no available model?
Copy the model ID again from the current pricing page, confirm the target tier still lists it, and check the key’s model scope. If the catalog or upstream state changed, choose a currently available combination and send one small request.
Should I immediately retry after a timeout, interrupted stream, or incomplete response?
Check Usage first to see whether the previous request was recorded and charged before retrying. Keep the request ID and partial output; an incomplete client response does not prove the request never happened.
What does it mean when text works but tool calls or file edits fail?
The text path may be working while the client-specific tool protocol, streaming events, or the selected model’s tool ability does not meet the task. Check the client guide’s boundary and verify with a smaller tool task.
Agents and clients
The same client name can still have different desktop, CLI, and IDE setup surfaces.
Which integration path should Codex use?
The current Codex path uses OpenAI Responses and selects XiuRouter through a local provider configuration. It affects new local Codex tasks, not cloud ChatGPT chats.
Why should Claude Code not reuse an OpenAI Base URL?
Claude Code speaks Anthropic Messages and appends /v1/messages itself, so its gateway root differs from the /v1 Base URL used by OpenAI SDKs. Follow the current Claude Code guide.
Will every Cursor feature use XiuRouter after setup?
No. The current Base URL override mainly affects the relevant OpenAI-compatible model path. Cursor-specific models, Tab completion, and other proprietary features can remain on Cursor’s service. Ask working does not automatically prove Agent file edits work.
How do I evaluate another agent or client?
First confirm it accepts a custom Base URL and API key, then identify whether it sends Responses, Messages, Chat Completions, or Gemini requests. Even then, verify one real request with the target model and reconcile its usage record.
Production checks and support
Before production use, check the catalog, send a real request, and review its Usage record.
What should be verified before production use?
Confirm the final request URL, authentication header, model ID, service tier, and current rate. Send one small request, reconcile status, tokens, and cost in usage records, then add application handling for timeouts, limits, and upstream failures.
Does catalog visibility equal a reliability or SLA promise?
No. The pricing page shows the current catalog and rates; it does not guarantee uptime, latency, output consistency, or failover. Plan retries, fallback behavior, monitoring, and spend limits in your application.
When should I use Router pages versus Docs?
Router pages help you choose a use case and API format and point to current models and prices. The docs provide parameters, setup, verification, recovery, and troubleshooting steps. Use the current pricing page and API responses for live data.
What information should I provide if self-service troubleshooting does not resolve the issue?
Contact support with the time, request ID, client, protocol, model ID, service tier, and error text. Do not send a full API key; use the name or masked digest shown in the console if the key must be identified.