Skip to content
Console
On this page

OpenRouter compatibility and differences

What carries over from OpenRouter, and what to change

Sylphx AI and OpenRouter both expose a Responses API. That does not make their routing controls, model catalogues or hosted tools interchangeable. This page describes the Sylphx contract, not a claim that every OpenRouter feature is available or that a production request has passed.

Compatibility at a glance

AreaSylphx contractMigration action
ResponsesCreate with POST /responses relative to https://api.sylphx.com/v1; streaming uses Responses events.Change the base URL and bearer key. Check the model in the current catalogue.
Chat Completions and MessagesAlternative encodings of the same Responses document.Use the encoding your client supports; do not mix fields between encodings.
Stored conversationsStore, retrieve, delete, cancel and compact routes are documented.OpenRouter Responses is stateless. Treat stored-response continuation as a Sylphx capability, not a portable OpenRouter request.
Function toolsYour application executes its own function tools.Preserve call IDs when returning tool outputs.
Hosted toolsweb_search, web_fetch, tool_search and datetime have documented platform execution semantics.Use the Sylphx tool declarations, not OpenRouter plugins or provider-specific tool names.
Provider selectionA public model ID selects admitted lanes; availability and features are catalogue facts.Do not assume OpenRouter provider, plugins, routing suffixes or fallback lists are accepted.
Model breadthThe Sylphx catalogue is curated, not a mirror of OpenRouter's catalogue.Resolve the exact model and supported capabilities before sending traffic.
PrivacyData-policy selection is a Sylphx contract.Review the policy separately; a familiar model name does not establish equivalent retention or routing.

See the Responses reference, model catalogue and hosted tools for the supported fields. Unsupported input is not a request to silently weaken a policy or choose a different tool engine.

Hosted-tool execution and usage

Declaring a hosted tool is not executing it. With tool_choice: auto, the model decides whether to call it. A search that returns no results is still a search; a function merely named web_search is not proof of platform execution.

Inspect the response's output items for tool activity. Streaming requests report progress and then continue to one terminal response. For platform web-search execution, tool_usage.web_search.num_requests reports a count aligned with executed searches. It is not a count of tool declarations and is not a currency amount. Do not infer usage for every hosted-tool type from this web-search field, or treat an absent field as a verified zero bill.

To verify a migration, retain the request ID, HTTP status, final response, tool-call items and reported usage from a real authenticated request. Check the corresponding usage or billing evidence separately before claiming end-to-end metering. A documentation example or a successful ordinary text response alone does not establish server-tool billing correctness.

Minimal Responses request

Use a model currently listed in the catalogue and a key with inference access:

Shell
curl --fail-with-body https://api.sylphx.com/v1/responses \
  -H "Authorization: Bearer $SYLPHX_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "openai/gpt-5.5",
    "input": "Reply with OK."
  }'

A 401 means authentication failed; it is not a successful inference. Likewise, a published comparison page does not prove inference availability. Handle typed failures using the errors reference.

OpenRouter references

Compare against OpenRouter's current Responses overview, model catalogue and provider routing. Provider support and model availability change independently on the two platforms. Re-check both contracts before relying on a feature in production.