Union Alpha on OpenRouter: Tests and Reviews

A user generated a result with Union Alpha in one shot, found it slightly worse than Astra alone, and suspects a large-model orchestrator with a smaller model behind a single facade, similar to Devin Fusion.

A developer praised Union Alpha's visual taste, said it would suit iterative game development workflows, and offered to make demos and showcases.

A user lists muse spark 1.3, Union Alpha, and Devin SWE-2 pro as currently free inference options and says they have started migrating their workflow gradually.

A long-running Union Alpha task took 7 hours 34 minutes, used 52.89M recorded tokens (83.8% cached) across 552 tool calls, and the user says it looks like Astra.

Cloudflare's documentation says Union Alpha calls several language models in parallel on each request and merges them into one answer, which explains the inconsistent reply style the author had noticed.

eliebakouch says the Union Alpha model served on OpenRouter a few hours earlier was not the same one up now, and with no visible chain of thought it is probably a closed-source model rather than GLM.

In a Three.js Lord of the Rings opening comparison, Union Alpha performed well, with less detail than Fable/Astra but clearly better than GLM 5.3 Flash and other recent open models.

A user complains that the stealth model Union Alpha on OpenRouter is now hard to test and calls on the provider to add more compute.

Asked to introduce itself in a first conversation, the stealth model Union Alpha listed web search, web and PDF reading, image generation, a sandboxed Linux environment and a "council" of independent AI models.

The author finds Union Alpha both smart and cheap and says its tone resembles OpenAI's GPT, while GPU shortages and tokenizer tests leave guesses like GLM, Llama and Mistral uncertain.

A user ran the same prompt at the highest reasoning setting on Union Alpha and Kimi K3; Union Alpha took 30 minutes and K3 took 10, and the author calls the two roughly on par.

Run on OpenRouter's free stealth/union-alpha endpoint with thinking off, Union Alpha scored 136/157 (86.6%) on SMF Official A with zero errors and no cost, acing reasoning and tools but weak on writing.