Mysterious model Union Alpha airborne: performance close to GPT-6 Astra, limited to one week of speed pedaling

Another "mysterious model" has appeared in the AI circle. On September 16, a model called Union Alpha quietly appeared on OpenCode and OpenRouter. There was no press conference, no signed laboratory, and no technical report—only a codename: stealth/union-alpha.

But the report card it throws out is very low-key: multiple benchmarks are close to GPT-6 Astra and Claude Opus 5, but the cost is said to be as low as only 1/18 of its opponents. And it’s free for a limited time only for one week.

Let’s look at the hard parameters first

• Model ID:stealth/union-alpha (OpenRouter/OpenCode)
• context window:262,144 tokens (256K)
• maximum output:131,072 tokens
• multimodal:Support text + image input, text output
• Ability: Tool calling, structured output, streaming response
• Positioning: For agentic coding, research and long context tasks
• Online time:September 16, 2026

What does 256K context mean? You can swallow an entire code base at once, handle documentation-intensive debugging tasks, or keep large amounts of project material within a single agent session. For people who work as coding agents, this capacity is really useful.

Performance: The data is available, but you need to be able to read it

The current public reviews mainly come from the community and independent tests, and are not a single standardized list. When reading, you must distinguish the source:

• OpenCode encoding snapshot: The four coding projects totaled 23.14/40, with an average cost of about $0.03 per prompt.
• Terminal-Bench v4: Some developers report about 51%, and each task costs about $1.50 (the methodology is not fully disclosed and cannot be used as a standard score for direct comparison)
• Cline's statement: Performance close to GPT-6 Astra and Opus 5, but expected to cost about 1/18th the

Note the word "close" - not equal, not even beyond. There is also negative feedback from the community: long context recall is unstable in some scenarios, has failed in OpenCode, and the response is slow under certain routes. So don’t just look at the running score, try it on your own real tasks.

Free for a week, but don’t get too happy too early

During the preview period, the input and output prices are $0, and OpenCode clearly states that the preview will take about a week. But long-term pricing has already shown signs: the expected paid price given by the official website is $0.50 input / $1.50 output (per million tokens). Compared with cutting-edge models, this price is really tough.

In terms of privacy, OpenCode's privacy table labels Union Alpha as "not used for training user data, and data retention is 0 days." This is kinder than many free models, but it is still recommended not to feed really sensitive things.

Whose family does it belong to?

This is the biggest suspense. OpenCode's documentation only calls it a "stealth model" and does not identify the laboratory behind it. Speculations are flying all over the community, and GLM, Kimi, etc. have all been named, butNone have been officially confirmed.

Anonymously release the model, let everyone run scores, create a topic, and then reveal it after a while - this set of gameplay has been repeated in the past two years. Sometimes it’s a team holding back big moves, and sometimes it’s a well-planned marketing campaign. It's too early to draw conclusions now.

How to get started

Calling through OpenRouter is very simple, just one line of model name:

POST https://openrouter.ai/api/v1/chat/completions
Header: Authorization: Bearer $OPENROUTER_API_KEY
Body: { “model”: “stealth/union-alpha”, “messages”: […] }

Supports common parameters such as stream, tools, tool_choice, response_format, temperature, top_p, etc.

It is recommended to use this

• Run the real task first: Don’t just look at the running score, try it on the actual job you have at hand. Does the good running score mean it is suitable for your scenario?
• Do comparison tests: The same prompt is thrown to Union Alpha and GPT-6 Astra respectively, and the difference can be intuitively felt.
• Control risks: High-risk changes undergo automated testing + second model review. Don’t rely solely on an anonymous model.
• hurry up: The exemption is limited to about one week, after which the price will no longer be available.

Should I be excited?

Stay calm. Anonymous models have natural information asymmetry: you don’t know who trained them, what data they used, whether there are any security issues, whether they will suddenly disappear, or what has been done with your data.

But it’s definitely something worth trying – with 256K contexts, image inputs, tool calls, and a free preview period, the cost of trial and error is almost zero. It may be a prelude to a dark horse, or it may just be a topical marketing. The answer will not be clear until the identity is revealed and long-term pricing is implemented.

Since it’s free, run a few tasks of your own and save the comparison results. When it is announced, you will have first-hand information in your hands.

Mystery is mystery, and ease of use is the last word. Take advantage of the limited exemption and run.

Reference: OpenRouter stealth/union-alpha model page, OpenCode release notes, community reviews (buildfastwithai, etc.)