• Home
  • Technology
  • Gaming
  • Entertainment
  • World & Business
  • Science
  • Sports
  • AI
HomeTechnologyGamingEntertainmentWorld & BusinessScienceSportsAI
    • Home
    • Technology
    • Gaming
    • Entertainment
    • World & Business
    • Science
    • Sports
    • AI
    AI

    Union Alpha's “stealth” AI blends multiple models into one answer

    OpenCode describes Union Alpha as built for coding agents, with image support and no data training.

    Y CombinatorYC
    Yuchen JinYJ
    Pliny the Liberator 🐉󠅫󠄼󠄿󠅆󠄵󠄐󠅀󠄼󠄹󠄾󠅉󠅭PT
    24 Sources, ,

    TLDR

    Cloudflare's documentation describes Union Alpha as a blended AI model that runs multiple language models in parallel for each request and synthesizes one answer. It lists support for text and vision inputs.

    On September 16, 2026, OpenCode advertised Union Alpha as a stealth model built for coding agents, with no data training and free access for the following week.

    One user who initially guessed it was GLM 5.4 later withdrew that identification, saying the wrong model had been served. The same user subsequently described Union Alpha as a router, linking to Cloudflare's documentation.

    Combined views

    4.2M

    24 Sources, first seen 21d ago

    Combined views

    4.2M

    24 Sources, first seen 21d ago

    21K likes
    21d ago
    first seen 21d ago
    21K likes
    1.4K comments
    5.2K saves
    1.5K reposts
    1.4K comments
    5.2K saves
    1.5K reposts

    Sentiment

    Positive20%80%Negative

    Summary

    Some accounts welcomed the frontier-level performance and low prices from new open models like Union Alpha on OpenRouter, while many others called out unreliable tool calls, secret leaks, and quick failures during agent use.

    Based on 30 sentiment-bearing replies from 30 accounts across 4 conversations.

    Sentiment

    Positive20%80%Negative

    Summary

    Some accounts welcomed the frontier-level performance and low prices from new open models like Union Alpha on OpenRouter, while many others called out unreliable tool calls, secret leaks, and quick failures during agent use.

    Based on 30 sentiment-bearing replies from 30 accounts across 4 conversations.

    Today's Rank

    —

    Not ranked yet

    Today's Rank

    —

    Not ranked yet

    24 Sources

    Santiago@svpinoA new, mysterious model just showed up. We don't know who created it. It's free for about a week on OpenRouter, with zero data retention for prompts and code. It's built for coding and multi-step agent tasks. Anyone could be behind this model. I can't wait to see the benchmarks.21d
    OpenCode@opencodeUnion Alpha (stealth model) is free for the next week - no data training - built for agentic coding - supports images let's see what you can do21d
    dax@thdxrRT @opencode: Union Alpha (stealth model) is free for the next week - no data training - built for agentic coding - supports images let's…21d
    Alex Atallah@alexatallahHere's the DeepSWE result for Union Alpha, a stealth model we just launched. Try it now! Works in every harness. $ ori [code | your-fav-harness] --model=stealth/union-alpha21d
    Yuchen Jin@Yuchenj_UWAny guess what this Union Alpha beast on OpenRouter is? Claude Fable 5.1-level, over 10x cheaper. Waiting for it to be open-sourced.21d
    Rohan Paul@rohanpaul_aiUnion Alpha is the brand-new anonymous “stealth” multimodal model that dropped today on OpenRouter and OpenCode. And in this 3D scenes rendering test by AI/ML API (@aimlapi) it used 3x fewer tokens vs Ox Alpha. i.e. high task efficiency per token. Union Alpha is built for research, coding, and agentic workflows with near frontier level coding strength. Usage is already huge for a day-zero model. OpenRouter app traffic showing ~1.96B tokens. And OpenCode advertised ~5T tokens/day capacity.21d
    elie@eliebakouchhi GLM 5.4? seems to be the same/similar tokenizer as the GLM 5 series + the model answers quite often that it's GLM from ZAI (screenshot in thread)21d
    Chubby♨️@kimmonismusHere we go again: a new stealth model scores 74% on DeepSWE, beating GPT-5.6 Sol and most other models. It also reportedly outperforms GPT-5.6 Sol on Terminal-Bench 2.1 and SWE-Bench Verified, and being much cheaper at the same time. My bet is on another chinese model.21d
    Alex J. Champandard 🍃@alexjc@eliebakouch Guessing bigger model, mid-generation jump. So GLM 5.5?21d
    Ben Tossell@bentossellgood first experience with this model 😳 also...sounds like another model21d

    24 Sources

    Santiago@svpinoA new, mysterious model just showed up. We don't know who created it. It's free for about a week on OpenRouter, with zero data retention for prompts and code. It's built for coding and multi-step agent tasks. Anyone could be behind this model. I can't wait to see the benchmarks.21d
    OpenCode@opencodeUnion Alpha (stealth model) is free for the next week - no data training - built for agentic coding - supports images let's see what you can do21d
    dax@thdxrRT @opencode: Union Alpha (stealth model) is free for the next week - no data training - built for agentic coding - supports images let's…21d
    Alex Atallah@alexatallahHere's the DeepSWE result for Union Alpha, a stealth model we just launched. Try it now! Works in every harness. $ ori [code | your-fav-harness] --model=stealth/union-alpha21d
    Yuchen Jin@Yuchenj_UWAny guess what this Union Alpha beast on OpenRouter is? Claude Fable 5.1-level, over 10x cheaper. Waiting for it to be open-sourced.21d
    Rohan Paul@rohanpaul_aiUnion Alpha is the brand-new anonymous “stealth” multimodal model that dropped today on OpenRouter and OpenCode. And in this 3D scenes rendering test by AI/ML API (@aimlapi) it used 3x fewer tokens vs Ox Alpha. i.e. high task efficiency per token. Union Alpha is built for research, coding, and agentic workflows with near frontier level coding strength. Usage is already huge for a day-zero model. OpenRouter app traffic showing ~1.96B tokens. And OpenCode advertised ~5T tokens/day capacity.21d
    elie@eliebakouchhi GLM 5.4? seems to be the same/similar tokenizer as the GLM 5 series + the model answers quite often that it's GLM from ZAI (screenshot in thread)21d
    Chubby♨️@kimmonismusHere we go again: a new stealth model scores 74% on DeepSWE, beating GPT-5.6 Sol and most other models. It also reportedly outperforms GPT-5.6 Sol on Terminal-Bench 2.1 and SWE-Bench Verified, and being much cheaper at the same time. My bet is on another chinese model.21d
    Alex J. Champandard 🍃@alexjc@eliebakouch Guessing bigger model, mid-generation jump. So GLM 5.5?21d
    Ben Tossell@bentossellgood first experience with this model 😳 also...sounds like another model21d