Ox Alpha streaming: API Setup Guide for Live Responses

Ox Alpha streaming is an API workflow for receiving a response as it is generated instead of waiting for the entire completion. OpenRouter lists Ox Alpha as a reasoning model intended for coding, sustained agentic work, production workloads, and tasks that combine text with visual context. The model is identified by the slug stealth/ox-alpha.

The provider is described as an anonymous third-party operator during the preview period. OpenRouter routes requests to that provider but does not identify itself as the model’s developer, owner, or provider. This distinction matters when you review retention terms, operational responsibility, and production suitability.

The OpenRouter model page lists a 1M context window, text, image, and video input support, and text output. The page also shows a listed release date of August 20, 2026. Because this is a preview-style stealth model, verify current behavior before depending on a specific modality, limit, or provider policy in a critical application.

The fastest setup path is to create an OpenRouter API key, store it as an environment variable, and send a request using stealth/ox-alpha. OpenRouter describes its API as OpenAI-compatible, so many existing SDK patterns can be adapted by changing the base URL, authentication configuration, and model slug.

Never place a production API key directly in browser code, public repositories, screenshots, or client-side applications. A server-side route should receive the user request, attach the secret key, call the provider, and return only the data your application needs.

The basic environment setup is:

export OPENROUTER_API_KEY=sk-or-v1-...

The request endpoint is a POST operation. The essential headers are the content type and bearer authorization. OpenRouter also documents optional HTTP-Referer and X-Title headers for identifying an application on its rankings and related platform displays.

OpenRouter documents streaming as a server-sent event workflow. In the request body, set "stream": true; the server then returns incremental event data as the model produces output. Your client should append each non-empty text delta to a buffer and render it progressively.

A minimal cURL request looks like this:

curl -N -H "Content-Type: application/json" -H "Authorization: Bearer $OPENROUTER_API_KEY" -d '{"model":"stealth/ox-alpha","stream":true,"messages":[{"role":"user","content":"Explain recursion in simple terms."}]}'

The -N option helps cURL display the stream without unnecessary buffering. In an application, the equivalent behavior requires an SSE-aware parser or an SDK that exposes an asynchronous iterator.

The OpenRouter TypeScript SDK example uses openrouter.chat.send, passes the model and messages, and iterates over returned chunks. Text is read from chunk.choices[0]?.delta?.content. The final chunk may also contain usage information, including reasoning-token details when available.

For user-facing applications, separate the accumulated answer from the transport state. This lets you display “connecting,” “generating,” “completed,” or “failed” without corrupting the text already received.

A robust stream handler should also:

• Stop reading when the server signals completion.
• Handle a connection ending before a final usage object arrives.
• Avoid rendering raw provider errors to end users.
• Preserve partial text when a retry is offered.
• Cancel the request when the user navigates away or presses Stop.
• Record request duration without logging private prompt content unnecessarily.

OpenRouter’s published snapshot provides several operational signals for Ox Alpha. The listed provider showed a P50 latency of 5.30 seconds and throughput of 23 tokens per second on the provider summary. The wider performance panel displayed additional percentile averages, including average P50 throughput of 36 tokens per second and average P50 latency of 3.38 seconds across the shown measurement view.

These figures should not be treated as a promise for every region, prompt, SDK, or time period. Latency can change with prompt size, output length, traffic, provider load, tool use, and the complexity of the task. Streaming improves perceived responsiveness because users can see output before the full completion arrives, but it does not necessarily reduce total generation time.

The page showed 99.99% uptime and 99.51% availability over three days in the captured period. It also showed an average tool-call error rate of 2.27% and an average cache hit rate of 81.72% for the listed provider. Monitor your own workload because these metrics may differ from the public aggregate.

For production agent workflows, use bounded retries with exponential backoff. A retry should be safe for the operation being performed; do not automatically repeat an external action merely because the response stream disconnected. Tool calls deserve additional validation because a partial answer and a completed action are different states.

Use a staged rollout for Ox Alpha streaming. First validate a simple text-only request. Then test longer prompts, structured output, image or video inputs where supported, and tool-enabled agent tasks. This progression isolates integration errors before they become difficult to diagnose.

A practical troubleshooting sequence is:

1. Confirm the environment variable exists on the server.
2. Check that the bearer token is attached exactly once.
3. Verify the model slug is stealth/ox-alpha.
4. Confirm the request body is valid JSON.
5. Check that "stream": true is a Boolean rather than a quoted string.
6. Inspect the first response status before attempting to parse events.
7. Log transport metadata while redacting prompts, keys, and sensitive output.
8. Retry only when the failure is safe to repeat.

For a maintained integration, link your implementation to the Ox Alpha model and provider page on OpenRouter. It is the primary reference used for the published model slug, capabilities, provider information, performance snapshot, pricing display, and quick-start examples.