Union Alpha FAQ

Union Alpha is a stealth AI model that appeared in OpenRouter's catalog on September 16, 2026 under the id stealth/union-alpha. It is multimodal (text and image input), has a 262,144-token context window, supports tool calling, and is free during its preview. No company has publicly claimed it as of September 17, 2026.

Unknown. OpenRouter lists the provider as 'Stealth' and describes it as developed and operated by a third-party provider. Community speculation has focused on the GLM family because the previous stealth model, Ox Alpha, turned out to be Z.ai's GLM-5.3-Flash — but the published specifications of the two models do not match.

Yes, during the preview period it is listed at $0 on OpenRouter. Reported anticipated pricing is $0.50 per million input tokens and $1.50 per million output tokens, which would put a typical coding task at roughly $0.65.

Very good for the price. It scores 74.2% on SWE-bench Verified and 74.0% on DeepSWE, which is level with GPT and Opus-class models that charge roughly 50 times more per task. On Terminal-Bench v4.0 it lands around 52%.

262,144 tokens, usually described as 256K. That is large enough for repository-scale context in a single request.

According to OpenCode's published note, no — your data is not used for training. Note that this claim comes from the routing platform, not from the unnamed developer.

No. Independent measurement puts it in the 6th percentile for speed, meaning it is noticeably slower than most catalogued models. The quality-per-dollar is excellent; the quality-per-second is not.

Yes. It accepts image input alongside text, which makes it usable for screenshot debugging, UI review and diagram-driven tasks. There is no documented video or audio support.

No. As of September 17, 2026 the provider remains undisclosed. OpenRouter's catalog entry states only that the model is developed and operated by a third-party provider.

Unconfirmed, and the public specifications argue against it. GLM-5.3-Flash is documented with a 1.3M-token context window, video input, 320B total / 18B active parameters and MIT-licensed weights. this listing shows 262K context, no video, no parameter count and no weights.

The previous stealth model on OpenRouter. It was later identified as Z.ai's GLM-5.3-Flash. That precedent is why people assume Union Alpha will also be attributed eventually — but it does not tell us to whom.

Blind testing produces unbiased feedback, a free preview generates real-world load data at scale, and staying anonymous means a weak debut leaves no permanent mark on the brand. See our stealth models page for the full picture.

The DeepSWE figure of 74.0% is described as audited, and SWE-bench Verified at 74.2% is widely reported. Terminal-Bench v4 positioning comes from an Artificial Analysis chart that plots score against cost without printing exact values. The SMF result is a community run and should be read as indicative.

A software engineering benchmark that measures how reliably a model resolves real coding tasks. its 74.0% resolution rate is what first put the model on people's radar, because it matched models costing far more.

Because benchmarks measure quality, not throughput. A 6th-percentile speed ranking means that for interactive use — chat, inline completion, anything a human waits on — the model will feel sluggish even though its output quality is high.

For the OpenRouter path, yes — you need an account and an API key even though the model itself is free during the preview. The OpenCode path routes through OpenCode instead.

stealth/union-alpha on OpenRouter, or opencode-go/union-alpha on OpenCode. Everything else in an OpenAI-compatible request stays the same.

Cursor, Cline and Roo Code all work by adding OpenRouter as a custom OpenAI-compatible provider. OpenCode supports it natively. Any terminal agent that speaks the chat completions format will also work.

Yes. OpenRouter exposes an OpenAI-compatible chat completions endpoint, so the official openai SDK, LangChain, LlamaIndex and similar tooling all work with a base URL change.

Use the standard OpenAI multimodal message format — a content array mixing a text object and an image_url object. the model accepts image input alongside text.