Union Alpha: Anonymous Model Unmasked in a Day

Episode 1 · Mystery stealth model Union Alpha launches free, sparking identity speculation (2026-09-16, 16 posts)

On September 16, an anonymous stealth model called Union Alpha quietly appeared on OpenRouter and OpenCode, with Cline later offering it free for about a week. It claims multimodality, a 256K/262K context window, image input and tool calling, targeting research, coding and agentic workflows, with an explicit promise not to train on user data. OpenRouter co-founder Alex Atallah announced the launch and shared its DeepSWE benchmark results; users can call it via --model=stealth/union-alpha. rohanpaulai shared free trial access and AI/ML API MCP integration instructions for Claude Desktop, Cursor and Claude Code. User doodlestein reported burning roughly 5.5 billion tokens in a day for free, calling the model quite good.

Confirmed

• Union Alpha launched on OpenRouter and OpenCode on Sept 16, then free on Cline for about a week: 256K/262K context, image input, tool calling, no training on user data
• Researcher YuchenjUW tested it as reaching Claude Fable 5.1-level capability at over 10x lower price; Cline says performance approaches GPT-6 Astra and Opus 5 at roughly 1/18 the expected cost
• Per rohanpaulai, an AI/ML API 3D scene-generation comparison showed Union Alpha using about one-third the tokens of competitors
• TNG Technology's reverse engineering suggests it is not a single model but an AutoRouter-style mixture, consistent with its name "union"

Unconfirmed

• The provider's identity is unknown. Community guesses include an unreleased Alibaba Qwen4 27B (relayed by cephaloform), unverified
• eliebakouch retracted his earlier GLM-tokenizer analysis, clarifying the backend had served a wrong model; Union Alpha is likely not GLM. peterjliu notes it uses a Llama-family tokenizer, pointing elsewhere—no conclusions
• TNG's router conclusion is third-party analysis with no official response; its true internals remain unconfirmed

Why it matters

Stealth models are a common channel for labs to anonymously test unreleased models. Union Alpha's large context, free access, agent focus and benchmark results triggered widespread identity speculation; the GLM lead has been withdrawn, leaving Qwen and Llama-tokenizer theories, plus TNG's alternative that it may be a multi-model routing system rather than a single new model.

• Mystery stealth model Union Alpha hits OpenRouter: free, 256K context, agentic focus — gaganghotra_ · 2026-09-16
• Mystery stealth model hits OpenRouter, built for coding and agentic workflows — kevinkern · 2026-09-16
• Stealth model Union Alpha appears on OpenRouter: multimodal, 256K context, frontier-level claims — FlunkyGraphics · 2026-09-16
• OpenRouter quietly launches stealth model Union Alpha with DeepSWE results — airesearch12 · 2026-09-16
• Mystery OpenRouter stealth model Union Alpha rumored to be Qwen4 27B — cephaloform · 2026-09-17
• Stealth Model Union Alpha Free for a Week, Built for Agentic Coding — AiBreakfast · 2026-09-17
• Mystery stealth model Union-Alpha goes free on OpenCode, built for agentic coding — gaganghotra_ · 2026-09-17
• Mystery model Union Alpha on OpenRouter claims Claude Fable 5.1-level at 10x cheaper — Yuchenj_UW · 2026-09-17
• Stealth model Union Alpha hits OpenRouter, uses 3x fewer tokens than Ox Alpha in 3D tests — rohanpaul_ai · 2026-09-17
• Union Alpha free trial link and MCP setup guide for Claude, Cursor, Claude Code — rohanpaul_ai · 2026-09-17
• Stealth model Union Alpha on OpenRouter rumored to be Zhipu's GLM 5.4 — zephyr_z9 · 2026-09-17
• Cline's free stealth model Union Alpha: near-flagship agentic coding at 1/18 cost — WolframRvnwlf · 2026-09-17
• OpenRouter's Mysterious Union Alpha Looks Like A 'Brain Of Brains' AutoRouter — airesearch12 · 2026-09-17
• Mystery OpenRouter model 'Union Alpha' was a wrong serving, not GLM — eliebakouch · 2026-09-17
• Mystery OpenRouter model "Union Alpha" rumored near-frontier, uses Llama 3 tokenizer — peterjliu · 2026-09-17
• Mystery 'Union Alpha' model on OpenRouter let one user burn 5.5B tokens free — doodlestein · 2026-09-18

Episode 2 · Mystery Stealth Model 'Union Alpha' Tops DeepSWE Leaderboard (2026-09-17, 4 posts)

A stealth model dubbed Union Alpha appeared on OpenRouter, scoring 74% on DeepSWE to rival GPT-6-level models while being free with no data retention, though some argue the benchmark is saturated.

• Mystery model 'Union Alpha' tops DeepSWE, claiming GPT-6-class capability at DeepSeek-level prices — daniel_mac8 · 2026-09-17
• Union Alpha scores 74 on DeepSWE matching GPT-6 Astra, but benchmark may be saturated — brandon_galang · 2026-09-17
• Stealth model Union Alpha scores 74% on DeepSWE, beating GPT-5.6 Sol at lower cost — ZeroStateReflex · 2026-09-17
• Stealth Model Union Alpha Benchmarked: Matches Frontier Models on Agentic Coding Tasks — Prompt Engineering · 2026-09-17

Episode 3 · The Mystery of Cloudflare's Stealth Model Union Alpha (2026-09-17, 5 posts)

Cloudflare quietly introduced a stealth model, Union Alpha, in its Workers AI docs and AI Gateway, describing it as a multimodal model for research, coding, and agent workflows with frontier-level general performance, accessible via API. The community quickly began hunting for its identity, with no official answer yet.

Confirmed

• Union Alpha appeared as Cloudflare's first stealth model with an official frontier multimodal positioning and API access (koltregaskes).
• eliebakouch found the OpenRouter-launched version differed from the current one, initially suspecting GLM or a closed model without chain-of-thought (relayed by teortaxesTex).
• Plinius mentioned GLM-5.4 in an interaction with opencode, fueling speculation (relayed by koltregaskes).

Unconfirmed

• One account says Union Alpha is not a single model but a blended model that runs multiple LLMs in parallel per request and synthesizes one answer (relayed by ritakozlov and teortaxesTex).
• koltregaskes later posted it is not a router and the prevailing guess is Zhipu's GLM 4.5 (mistyped as GLM 5.4 in the post); the rebrand theory and the blending theory conflict and neither is officially confirmed.

Why it matters

If Union Alpha is a multi-model synthesizing router, it signals a new orchestration capability at Cloudflare's AI Gateway layer; if it is a rebranded GLM 4.5, it reflects the practice of anonymously testing frontier models to gather unbiased eval data. Until Cloudflare reveals the answer, the model's identity remains an open mystery.

• Cloudflare Launches First Stealth Model Union Alpha, Blending Multiple LLMs Per Request — ritakozlov · 2026-09-17
• Cloudflare's mysterious Union Alpha revealed: a router that queries multiple models in parallel — teortaxesTex · 2026-09-17
• Cloudflare's first stealth model Union Alpha appears; community guesses it's GLM 4.5 — koltregaskes · 2026-09-17
• Cloudflare's First Stealth Model Union Alpha Suspected to Be a Rebranded GLM — koltregaskes · 2026-09-17
• Union Alpha, Cloudflare's first stealth model, rumored to be GLM 4.5, not a router — koltregaskes · 2026-09-17

Episode 4 · Hands-On Test Suggests Anonymous Model union-alpha Is No Flagship (2026-09-17, 2 posts)

Blogger karminski3 tested the anonymous model union-alpha on OpenRouter and concluded it is neither a Chinese flagship nor from OpenAI/Anthropic, noting its three.js knowledge is stuck at r127—well below 2026 flagship level.

• OpenRouter's mystery model union-alpha flagged as non-flagship by benchmark-less test — karminski3 · 2026-09-17
• Quick test of OpenRouter's anonymous union-alpha model suggests it's no frontier flagship — karminski3 · 2026-09-17

Episode 5 · Mystery Model Union Alpha Unmasked as Pareto 26.9 Within a Day (2026-09-18, 2 posts)

The anonymous model Union Alpha launched on OpenRouter was identified by the community within a day as Pareto 26.9 from Unbiased, while critics argue such model routing effectively means silent downgrading.

• Stealth model Union Alpha unmasked in a day as eval outfit's Pareto 26.9 — karminski3 · 2026-09-18
• Community unmasks stealth model Union Alpha as Unbiased's Pareto 26.9 in one day — gaganghotra_ · 2026-09-18