Union Alpha is a multimodal model built for research, coding, agentic workflows

A New Stealth Model Appears on OpenRouter With a Free Preview

On September 16, 2026, a model called Union Alpha quietly showed up on OpenRouter with a price tag of zero, no developer attribution, and a claim of "frontier-level" performance across general-purpose tasks. The listing is short on details. There is no model card, no weights, and no lab publicly taking credit for it. That pattern will sound familiar to anyone who has followed OpenRouter's stealth model releases over the past year.

What the Listing Actually Says

Union Alpha is described as a multimodal model built for research, coding, and agentic workflows. It accepts text and image inputs and returns text output. The context window is 262,144 tokens, large enough to fit an entire medium-sized codebase or a stack of documentation in a single prompt. It supports tool calling and structured JSON output, both requirements for any model that wants to work inside automated agent pipelines rather than just chat.

Performance numbers from OpenRouter show 21 tokens per second throughput and a P50 latency of 11.33 seconds. Uptime sits at 100% during the preview. The model accepts standard Chat Completions API parameters including temperature, top_p, max_tokens, and streaming.

Under the hood, Cloudflare's documentation describes Union Alpha as a "blended AI model" that runs multiple language models in parallel for each request and synthesizes a single answer. That is a different architecture from a single monolithic model, and it raises questions about which models are actually being combined and how the synthesis works. Nobody is answering those questions yet.

The Free Preview Window and What Happens After

During the initial preview period, expected to last roughly one week, Union Alpha is completely free. No credits are charged for prompts or completions. After the preview ends, pricing has not been announced. Developers who build on top of the model during the free period should be prepared for that to change.

The data policy during preview is notable: prompts and completions may be retained by the provider but are not used for training. All other use is governed by OpenRouter's Stealth Model Terms. That means your code snippets and research prompts could be stored, but at least during the preview, they are not feeding into the next version of the model.

Part of a Growing Pattern

Union Alpha is the latest in a series of anonymous models appearing on OpenRouter. In February 2026, a model called Pony Alpha launched stealth before Zhipu AI confirmed it was part of its GLM-5 system five days later. In March, Hunter Alpha appeared and sparked speculation about DeepSeek's involvement, given its self-identification as "a Chinese AI model primarily trained in Chinese" and a May 2025 knowledge cutoff matching DeepSeek's chatbot.

Then in August, Ox Alpha arrived with a one-million-token context window and scored 80% on DeepSWE coding benchmarks, beating Anthropic's Claude Fable 5 at 65%. Technical analysis pointed to Zhipu AI's GLM-5.3 family based on tokenizer fingerprinting. None of these connections have been officially confirmed for Union Alpha.

Why Developers Should Pay Attention

The free access is the obvious draw, but the real signal is the agentic focus. Function calling, structured JSON, large context windows, and multimodal input are the building blocks for AI agents that can debug repositories, generate tests, and run multi-step workflows autonomously. If Union Alpha delivers on those capabilities, it fills a gap between expensive frontier models and smaller open-source alternatives that often struggle with long-horizon tasks.

The blended architecture adds another wrinkle. If the model is genuinely routing requests to multiple underlying systems and combining the results, it could offer better reliability and fault tolerance than a single model. OpenRouter already supports provider-level failover, and a model that builds that philosophy into its own design could be interesting for production deployments where uptime matters.

The anonymous nature cuts both ways. Without a known provider, there is no public training data, no published benchmarks to verify, and no accountability if the model degrades or disappears after the preview. Developers who build critical pipelines on top of Union Alpha should treat it as experimental until someone claims it and publishes real evals.

How to Try It

Access is through OpenRouter's standard Chat Completions API at https://openrouter.ai/api/v1/chat/completions. Set the model field to "stealth/union-alpha" and include your OpenRouter API key in the Authorization header. Streaming is supported, as are all the standard optional parameters.

A one-week free window with no strings attached is rare for a model claiming frontier performance. Whether Union Alpha turns out to be the work of a major lab testing a new release strategy, a startup generating buzz, or something else entirely, it is worth a few minutes of experimentation before the pricing changes.