Union Alpha on OpenRouter: Tests and Reviews

Cline plugged Union Alpha into its own harness and says its agentic coding performance is close to GPT-6 Astra and Opus 5 at roughly 18x lower expected cost.

One user says Union Alpha performs around Fable 5.1 level at a much lower cost and wants to see how it holds up on agentic coding.

A user's Terminal-Bench v4.0 comparison puts Union Alpha at 50%, against Astra max at 59%, Sol at 40% and Grok 4.6 at 21%, with the model free this week.

A model called Union Alpha, described as Fable and Sol class, is free for a week on OpenCode with image support, no data training and an agentic coding focus.

A roundup of Union Alpha's reported scores (74% on DeepSWE, ~50% on Terminal-Bench v4, 256K context, multimodal and tool calling) and doubts about who could give away 5 trillion tokens a day.

alexgetmancom says Union Alpha showed up in stealth on OpenCode, scoring around 51% on Terminal Bench v4 at roughly $1.5 per task, close to GPT 6 Astra high for far less, and guesses it may be Kimi.

An anonymous model, Union Alpha, is free on OpenCode for a week, built for agentic coding with image support and no training on user data, and sits near GPT-6 Astra (high) on Terminal-Bench.

The new stealth model Union Alpha, similar to Ox Alpha, offers 7 days of free access, zero data retention, multimodal support and up to 5T tokens/day capacity, with strong Terminal-Bench 4.0 scores.

Union Alpha is free on OpenRouter until about September 22, with zero data retention and a focus on agentic coding; the author suggests testing it on real tasks rather than trusting leaderboards.