- Ox Alpha multimodal accepts text, images, and video while returning text responses.
- Primary use cases include coding, long-horizon engineering, reasoning, and agentic workflows.
- API access uses the
stealth/ox-alphamodel slug through OpenRouter’s compatible endpoint. - Pricing status is listed as free, but preview terms and provider retention policies still apply.
- Performance profile shows 23 tokens per second, 5.30-second P50 latency, and 1M context.
Ox Alpha multimodal Overview
Ox Alpha multimodal is a reasoning-focused AI model designed for software engineering, sustained agentic work, production workloads, and tasks that combine text with visual context. The model is listed by OpenRouter under the slug stealth/ox-alpha, with a reported 1M-token context window and support for text, image, and video inputs.
The model is operated by an anonymous third-party provider during its preview period. OpenRouter routes requests to that provider but does not identify itself as the model’s developer, owner, or upstream provider. This distinction matters when evaluating support, availability, data handling, and long-term model stability.
As of August 22, 2026, the public listing identifies Ox Alpha as free to use through the displayed provider. Pricing and availability can change, so teams should check the live Ox Alpha API pricing and provider page before deploying a workload.
| Model Detail | Ox Alpha Status |
|---|---|
| Model slug | stealth/ox-alpha |
| Model type | Reasoning and coding model |
| Input modalities | Text, images, video |
| Output modality | Text |
| Context window | 1M tokens |
| Listed release date | August 20, 2026 |
| Displayed price | Free |
| Provider count | One provider |
| Provider identity | Anonymous third-party operator |
Coding Focus
Ox Alpha is positioned for software engineering, codebase work, production tasks, and sustained development sessions.
Visual Context
The documented input format supports text alongside image and video URLs for multimodal analysis.
Agentic Work
Its target workflows include long-running agents, tool use, iterative tasks, and complex reasoning across multiple steps.
Treat Ox Alpha as a preview model with strong workflow potential, not as a fully documented general-purpose product. Test it against your own prompts before committing production traffic.
Multimodal Capabilities and Best Uses
The defining feature of Ox Alpha multimodal is its ability to process multiple input types in one request. The public quick-start example combines a text instruction with an image URL and a video URL, asking the model to describe the visual content. The documented response format returns text.
This makes the model suitable for workflows where code, screenshots, diagrams, recorded sessions, or visual references need to be considered together. The available material does not establish specialized benchmarks for optical character recognition, video understanding, or image generation, so those capabilities should be validated through practical testing rather than assumed.
The strongest documented use cases are:
- Reviewing code while inspecting screenshots of an application.
- Analyzing visual context during an agentic software workflow.
- Working with long project instructions or large codebase excerpts.
- Combining written requirements with image or video references.
- Supporting iterative production tasks through an OpenAI-compatible API.
| Input Type | Supported Role | Example Workflow |
|---|---|---|
| Text | Instructions, requirements, code, questions | Ask for a refactor plan or debugging explanation |
| Image | Visual context supplied by URL | Review a UI screenshot alongside implementation notes |
| Video | Temporal visual context supplied by URL | Ask what appears in a recorded screen session |
| Mixed input | Text combined with image or video | Compare a design reference with a written specification |
| Output | Text response | Return analysis, recommendations, code, or explanations |
Code Review
Provide implementation details with a visual reference to help connect code behavior and interface results.
Debugging
Use screenshots, logs, and written symptoms together when a problem cannot be explained by text alone.
Planning
Give the model large requirements or repository context for long-horizon engineering plans.
Agent Support
Integrate Ox Alpha into tools that need reasoning, visual context, and iterative task execution.
The listing confirms that image and video inputs are accepted, but it does not publish detailed limits for file size, duration, resolution, or supported URL sources. Validate those constraints in a controlled test before building around them.
Ox Alpha multimodal API Setup
OpenRouter presents Ox Alpha through an OpenAI-compatible API style. The key configuration change is the model identifier: applications should send requests using stealth/ox-alpha. Most compatible SDKs can be adapted by changing the base URL and model slug, although exact integration behavior depends on the SDK and request format.
The basic setup begins with an API key stored as an environment variable. Streaming can then be enabled with "stream": true when the application needs server-sent events and incremental output.
Create an API Key
Generate an API key from the OpenRouter dashboard and store it in an environment variable such as OPENROUTER_API_KEY. Avoid placing the key directly in browser code or public repositories.
Select the Model
Set the model value to stealth/ox-alpha. This slug identifies Ox Alpha on the OpenRouter route and should replace the model name used by your existing compatible client.
Build the Request
Send a chat request with a text message, or provide structured content containing text, an image URL, and a video URL. Keep the initial test narrow so errors are easier to diagnose.
Enable Streaming When Needed
Add "stream": true when your interface should display incremental output. Streaming is useful for long answers, agent updates, and interactive development tools.
Verify Usage and Errors
Inspect the returned message, usage information, latency, and tool-call behavior. Record results across several prompts before moving from testing to production.
| Parameter | Type | Default | Practical Purpose |
|---|---|---|---|
max_tokens | Integer | Not specified | Sets an upper limit for generated response tokens |
temperature | Float | 1 | Controls response variety |
top_p | Float | 0.95 | Limits selection to a probability mass of likely tokens |
tools | Array | Not specified | Supplies tool definitions using the compatible request shape |
tool_choice | String or object | Not specified | Controls whether and how a tool is selected |
top_k | Integer | 0 | Restricts token choices at each generation step |
response_format | Map | Not specified | Requests a specific output format |
API Readiness Checklist:
- Create and securely store an OpenRouter API key
- Confirm the model slug is stealth/ox-alpha
- Test text-only requests before adding visual inputs
- Test image and video URL handling separately
- Log latency, output quality, and tool-call errors
Start with a text-only request, then add one image, then add video or tools. This staged approach isolates input-format and integration problems more effectively than testing every feature at once.
Performance, Pricing, and Reliability
The public provider listing reports a single upstream provider for Ox Alpha. That provider is shown with free input and output pricing, a P50 latency of 5.30 seconds, throughput of 23 tokens per second, and 100.00% uptime in the displayed provider row.
The broader performance panel reports different aggregate percentile values across the available monitoring view. For example, it lists average P50 throughput at 36 tokens per second and average P50 latency at 3.38 seconds, while end-to-end P50 latency is listed at 16.65 seconds. These figures should not be treated as interchangeable: provider throughput, round-trip latency, and end-to-end latency measure different parts of the request lifecycle.
| Metric | Displayed Value | How to Read It |
|---|---|---|
| Provider throughput | 23 tok/s | Reported P50 throughput for the listed provider |
| Provider latency | 5.30 s | Reported P50 round-trip latency |
| Provider uptime | 100.00% | Displayed provider uptime over the monitored period |
| Overall uptime | 99.99% | OpenRouter monitoring view for the recent three-day period |
| Overall availability | 99.51% | Percentage of requests successfully served in the same view |
| Tool-call error rate | 2.27% | Average error rate shown for the provider |
| Cache hit rate | 81.72% | Average cache hit rate shown in the performance panel |
Pricing is currently displayed as zero dollars per million input and output tokens, with weighted average input and output prices also shown as $0. However, “free” should be interpreted as the current listing status rather than a permanent guarantee. Preview access, provider changes, usage restrictions, and terms can affect future availability.
Data handling also deserves attention. The listing states that prompts and completions are retained by the third-party provider and are not used for training. Other use is governed by the applicable Stealth Model Terms. Review those terms before sending confidential source code, customer data, credentials, or regulated information.
A strong recent uptime display does not remove the need for retries, timeouts, logging, and fallback planning. A single-provider route creates a different operational profile from a model with multiple independent providers.
Production Guidance and FAQ
Ox Alpha can be a useful candidate for multimodal coding and agent experiments, especially when a large context window and visual inputs are valuable. Production adoption should be based on measured task quality, not only the public price or headline modality support.
Use a small evaluation set that reflects your real workload:
- Code changes across multiple files.
- Screenshot-based UI diagnosis.
- Video or screen-session interpretation.
- Tool-use tasks with recoverable failures.
- Long prompts that test context retention.
- Structured outputs that can be validated automatically.
For sensitive systems, place an application layer between users and the model. Redact credentials, restrict tool permissions, validate generated code, and preserve request metadata for debugging. Because the upstream provider is anonymous in the preview listing, document the model’s data-handling assumptions and review them whenever the provider terms change.
Q: What is Ox Alpha multimodal?
Ox Alpha multimodal is a reasoning model designed for coding, sustained agentic work, production workloads, and workflows that combine text with visual context. It accepts text, images, and video as inputs and returns text.
Q: How do I call Ox Alpha through an API?
Use an OpenRouter API key and set the model to stealth/ox-alpha. The listed integration follows an OpenAI-compatible request style, and streaming can be enabled with stream set to true.
Q: Is Ox Alpha free to use?
As of August 22, 2026, the OpenRouter listing displays zero-dollar input and output pricing. This is a current preview listing and should not be assumed to be permanent.
Q: Does Ox Alpha support images and video?
Yes. The documented quick-start format includes text, an image URL, and a video URL in the same request. The model returns a text response describing or analyzing the supplied content.
Use Ox Alpha for controlled multimodal evaluations first. Keep production safeguards in place, monitor provider behavior, and confirm current pricing and terms at the official listing before scaling usage.