OpenRouter and OpenCode released a multimodal AI called Union Alpha for agentic coding and research that scores 74% on the DeepSWE benchmark. The stealth model is free for one week starting September 16 and offers a 256K context window along with image support and tool calling. OpenRouter reported early traffic of approximately 1.96 billion tokens, and OpenCode advertised a daily capacity of 5T tokens for the service.
Early users allege the system is a model router rather than a new AI, claiming it switches between Llama-3.3-70b, Qwen 3.6 35b-a3b, and GLM 5.3 Flash. These reports suggest DeepSeek v4 Flash serves as a judge to route requests when the system is under load. OpenRouter announced additional capacity arrived on September 18 to resolve initial stability issues encountered by developers.
Key sources
- SOURCE@openrouter“Significantly more capacity for Union Alpha is coming online tomorrow morning”x.com
- SUPPORT@ycombinator“Union Alpha (stealth model) is free for the next week”x.com
- SUPPORT@teortaxestex“union alpha is just an extremely sloppy router that uses llama-3.3-70b, qwen 3.6 35b-a3b and glm 5.3 flash”x.com
- SUPPORT@theo“released 'union alpha' without telling us it was a crappy model router”x.com
- SOURCE@cline“near GPT-6 Astra and Opus 5 performance for ~18x lower expected cost”x.com
- SUPPORT@testingcatalog“handles 262K tokens of context and up to 131K tokens of output”x.com
- SUPPORT@kimmonismus“scores 74% on DeepSWE, beating GPT-5.6 Sol and most other models”x.com
- SUPPORT@rohanpaul_ai“OpenRouter app traffic showing ~1.96B tokens. And OpenCode advertised ~5T tokens/day capacity”x.com