Union Alpha on OpenRouter: Tests and Reviews

Someone matches GLM-5.5 rumors of 3T+ parameters and 1M token context against the anonymous Union Alpha model in OpenCode and asks whether they are the same.

A long post walks through the reveal: Pareto has several open and frontier models work the same task and check each other, calling a stronger model when needed; the team says there was no new training run, and the free preview ended after traffic hit billions of tokens per minute.

A post lists a guessed breakdown for Pareto 26.9: Llama 3.3 70B and Qwen 3.6 for easy tasks, GLM 5.3 Flash for coding, DeepSeek V4 Flash as judge, and GPT 6 Astra for hard tasks.

The author, initially skeptical of Hermes' multi-model theory, says Cloudflare documentation confirms Union Alpha was a routing system called Pareto 26.9, and laments that the stealth phase ended so quickly.

A day after launch, Union Alpha is said to be Pareto 26.9 from TheUnbiasedCo, a router that runs several open and frontier models on the same task and calls a stronger one when needed, priced at $2.50/$7.50 per million tokens, under a quarter of Astra's price.

After testing Union Alpha, a user concludes it is neither Zhipu's nor Kimi's model and falls far short of the earlier blind test, so there is little reason to expect much.

A user's inference from the call path suggests Union Alpha could be a gateway that fans out to multiple models in parallel and merges the results, mixing self-hosted open-source models with commercial API relays.

The poster says Union Alpha is not a routing layer but is currently thought to be GLM 4.5, Cloudflare's first stealth model, and links to the official documentation.

A Polish user notes that the free stealth model Union Alpha showed up on OpenRouter just as Ox Alpha did a month earlier, and wonders which model is behind it this time.

A roundup of Union Alpha's 256K context, multimodal and agentic coding positioning and low-cost rumors draws a parallel to Ox Alpha turning out to be GLM-5.3-Flash and guesses it is an anonymous GLM-5.4 test.

The author reports Union Alpha runs at 300-400 tokens per second in Codex with 262k context and image input, free for a week, and compares it to Ox Alpha, which was revealed as GLM-5.3-Flash after a week.

The author says Cloudflare once described Union Alpha as a blended system, archived costs match Pareto's published rates, and testing found multiple token-count patterns, though nothing is confirmed.