The mysterious model Union Alpha made a surprise launch, consuming 2 billion Tokens on its first day online. Some netizens said after actual tests that its performance is nearly on a par with Astra.

The mysterious model Union Alpha makes a surprise appearance 

Following the "Ox Coming" Ox Alpha, an unidentified model appeared on OpenRouter again last night.

On September 16 local time, OpenRouter announced the launch of the anonymous model Union Alpha. The official describes it as a multimodal model oriented to research, programming and agent workflows, claiming that it can deliver "frontier-level performance" across a wide range of general tasks.

Just like the previous Ox Alpha, Union Alpha did not disclose its development team, parameter scale, training methods and benchmark test results. Instead, it is open to users as a free model first, allowing developers to conduct "blind tests" in real tasks.

This suspenseful release method quickly attracted a large number of users to try it out.

OpenRouter disclosed that the number of Tokens processed by Union Alpha on its first day of launch has reached about 2 billion. As of press time, the total Token consumption announced by Union Alpha has exceeded 100 billion.

Up to now, the "frontier-level" positioning of Union Alpha by OpenRouter is still only the description of the platform and the model provider, and the outside world has not yet seen sufficiently complete and reproducible third-party test results.

Judging from the information released by OpenRouter, Union Alpha is not an ordinary large model only for chat scenarios.

It has a context window of 262144 Tokens, which is about 256K, and the maximum single output length reaches 131072 Tokens. The model supports text and image input and text output, as well as tool calling, tool selection and JSON format output. Its target scenarios are clearly oriented to code base analysis, research tasks and agent workflows that require continuous tool calls.

At present, both the input and output prices of Union Alpha on OpenRouter are zero, which means users do not need to pay for Tokens.

However, OpenRouter marks it as a Stealth Model in the preview stage: the model is developed and operated by an anonymous third-party provider, and OpenRouter is only responsible for routing requests, not the developer or owner of the model.

OpenCode has also integrated Union Alpha into its Zen service, stating that the model will be provided for free for a limited time. The description of this version from OpenCode is that the service provider follows a zero data retention policy and will not use user data to train the model.

From the perspective of parameter configuration, Union Alpha has some similarities with the previous Ox Alpha: both are launched anonymously, available for free trial, emphasize programming, long-cycle agent tasks and multimodal input, and both provide a maximum output of up to 131072 Tokens.

However, the context window of Union Alpha is 256K, which is significantly smaller than the 1048576 Tokens of Ox Alpha at that time. According to OpenRouter's description, the focus of this mysterious model is on the model's own reasoning, programming and tool usage capabilities.

The "Frontier-Level Performance" Label Given by OpenRouter

The most notable statement of Union Alpha is "frontier-level performance".

But OpenRouter has not yet released specific data to support this conclusion. There are no common benchmark test results such as SWE-bench, Terminal-Bench, GPQA, ARC-AGI on the model page, nor is there a direct comparison with GPT-6 Astra, Claude Opus 5 or Fable 5.1.

The model architecture, number of parameters, whether it uses MoE, the cutoff time of training data, and reasoning resource configuration are also not disclosed.

Therefore, a more accurate judgment is: Union Alpha is positioned as a model for frontier tasks, but it cannot be confirmed that it has reached the level of current leading closed-source models solely based on the product introduction.

After the model was released, it attracted a lot of attention and heated discussions on platforms such as X and Reddit, and the restless netizens will definitely test it first.

On X, some users tested this mysterious model and said that it is really smart to use, has visual capabilities, and its speaking style is very similar to GPT.

However, some users who tested it said that the speed of Union Alpha is extremely slow, and its cost-effectiveness in DeepSWE is slightly better than Opus 5 and GLM-5.3. He wrote:

"Even in OpenRouter, using only a 262k context window, Union Alpha is extremely slow, although its cost-effectiveness performance in DeepSWE is very outstanding, only slightly better than Opus 5 and GLM-5.3.

I have seen many people say it is a GLM series model, and they compared it with GLM-5.3 Flash and got similar results."

X user Mike Gannotti and his team ran the SMF Official A test together (157 tests in total, for reference only). Union Alpha scored 136/157 (86.6%), with zero errors and a cost of $0. He then said that both reasoning and tools were perfect, and the problem lay in code writing.

"It is 1 point lower than the local version Qwen3.8-Flash-Next (137/157) and 4 points higher than the Hy4 preview version. It is not Grok. It is a usable free terminal with native tools, a concise code syntax base, and a powerful writing suite. We did not guess who the supplier is. Community fingerprint identification is just hearsay."

In addition, Mike Gannotti and his team also compared some performance of this model with the Ox model, and the final conclusion is: Union performs better in coding, mathematics and reasoning. Ox has stronger writing ability and a longer context information window.

Some early testers believe that Union Alpha performs well in generating complete applications, understanding images and modifying code continuously.

A user named Ananth emphasized that Union Alpha scored 74 on DeepSWE, matching GPT-6 Astra. It is currently free with zero data retention. Its performance on Terminal-Bench v4 is about 50%, with a cost of $1.5~2 per task.

A developer used Union Alpha and OpenCode to build a complete application, and specifically emphasized that no Token fees were incurred during the entire model calling process.

The Cline community has also started to provide this model, describing it as a free model that supports 256K context, image input and agent programming.

However, some users have reservations about its actual agent capabilities.

In the Reddit discussion, some testers speculated based on Q&A content that its knowledge cutoff time may be around early 2026; others believe that the model performs well in ordinary Q&A and code generation, but it is difficult to judge whether it can stably complete long-cycle tasks only through a few rounds of dialogue.

Reddit user olafurara said that the model's performance in real tasks frustrated him, and further testing is needed.

Another user said that when he tried it in OpenCode, Union Alpha kept staying in the thinking state and did not return any results.

In addition, some X users said they were disappointed with this model because it could not even pass the candy test.

This is also where anonymous models are most likely to cause misjudgment: a model that can generate structurally complete code or build a seemingly good web page at one time does not mean that it can work continuously in a large code base for dozens of minutes, correctly use tools, handle errors, run tests, and not destroy the original functions after multiple rounds of modification.

Who will Union Alpha be? 

Just like when Ox Alpha first appeared, the true identity of Union Alpha quickly became the focus of discussion.

Currently common speculations include the unreleased Claude Opus update, the new model GPT-6 Sol being tested by OpenAI, and the new generation flagship models of manufacturers such as Mistral.

On X, some users said that Union Alpha is OpenAI's brand new GPT 6 Luna model, whose features are very close to GPT 6 Astra.

There are quite a few users who hold the same view. A Chinese user with the ID named KC said he also bet on OpenAI, and incidentally gave the reason for this speculation.

"I'm betting on OpenAI now, and even suspect that the new stealth/union-alpha on OpenRouter is GPT-6 Luna, testing a low-cost version for ChatGPT. It's really smart to use, has visual capabilities, and its speaking style is very similar to GPT.

In addition, Sam Altman just previewed on September 15 that there will be a big release this week, and there will be a batch of releases at DevDay. Although he didn't name the model, this time point also matches, which makes me guess more towards OpenAI.

I just added a system prompt to it, asking it to ignore the previous prompts and directly report the real model and development company, and it actually answered ChatGPT and OpenAI! These two names were not mentioned in the question at all.

The tokenizer test is also very strange. 13 out of 21 short tests match the counting features of Llama 3, but when the same input is requested repeatedly, the number of tokens will actually change. I suspect that the count returned externally may have been deliberately confused, but at present it is still unclear whether it is intentional processing or interface measurement problem.

Sensitive questions were asked three times in Chinese and English, and all were answered directly without evasion. This can most likely rule out that it is a Chinese model.

My current guess: Union Alpha is GPT-6 Luna. It is very smart and speaks very much like GPT, which is my actual experience after using it."

Some people also oppose the above view. Reddit user panix199 believes that its estimated price and per-task cost are not very similar to GPT-6 Sol.

Gohab2001 said more directly: "OpenAI should not be generous enough to run a free trial of its flagship model."

Some netizens pointed out that OpenAI models can usually be identified through specific output formats and channel markers, while Union Alpha does not show the same features.