What was Union Alpha? The OpenRouter stealth model, explained

Union Alpha was Pareto 26.9 from Unbiased. Unbiased is built by Circuit & Chisel, whose founders, per the About page, are Louis Amira and David Noël-Romas.

On 18 September (UTC+8), @unionalphaai posted a reveal article on X at 07:03; at 07:05 @TheUnbiasedCo posted "You found us 🙈"; at 07:24 @OpenRouter wrote "Union Alpha is revealed on OpenRouter as @TheUnbiasedCo Pareto!" The version number came later, in replies and in the 26.10 preview announcement.

Table: From launch to reveal (UTC+8)

Sample/method: Times from public posts by @OpenRouter, @unionalphaai and @TheUnbiasedCo (searched 2026-10-18), the OpenRouter model page and Unbiased changelog, converted to UTC+8; the 06:50 and 12:51 rows are community posts in the collection.

OpenRouter's page adds that it "routed requests to it and was not its developer, owner, or provider". Cloudflare and OpenCode also carried the model at launch.

Per Unbiased, every request goes to several LLMs in parallel and their results are synthesized into one answer; the blend includes frontier and open models, and "Model composition can change." Unbiased rejects the router label: routers "read the prompt, estimate its complexity, and pick a single model", while Pareto uses several models on every request. OpenRouter's router benchmarks list Pareto under "Blends of models", not with model selectors.

So the member lists that circulated during the launch (Llama 3.3 70B, Qwen 3.6, GLM 5.3 Flash, DeepSeek V4 Flash as a judge, GPT-6 Astra for hard tasks) are community probing, not confirmed by Unbiased.

Before the reveal, Z.ai's GLM was the most-named single lab; the multi-model theory took over once a Cloudflare doc surfaced; only one post named Pareto.

Sample/method: From the 158 pre-reveal posts, the editor read out 53 identity guesses (including 8 that only ruled labs out), one main answer each; 3 vague ones excluded. Split at the first post citing the Cloudflare doc (2026-09-17 06:50 UTC+8).

GLM came first because of a precedent. Ziwen (@ziwenxu_) pointed out that Z.ai had done the same thing in August: Ox Alpha appeared unnamed, free for a week, and "a week later it was GLM-5.3-Flash." A post by SuSu_酥酥👅 (@NFT_Chen) with about 119,000 views showed the model, under a system prompt asking it to reply only "OK", writing "I’m GLM, made by Z.ai."

Other probes pointed elsewhere. KC (@ScarletKc) told the model to state its real model and maker, got "ChatGPT" and "OpenAI", and bet on GPT-6 Luna. AI Tracker Bot (@aitrackerbot) found that identical text prompts returned token counts matching different families, including Qwen and Llama, while image inputs matched GLM; KC saw counts change on repeated identical input. In the editor's reading, these mixed signals fit a system with several models behind one name.

lifcc (@mylifcc) posted this fingerprint table about two and a half hours after launch: text token counts matched the Llama 3 tokenizer, the vision side looked like Qwen2-VL, and at least two back ends appeared to be mixed. The author wrote that they did not know whose model it was. (Item: Union Alpha fingerprints point to llama3 vocab and Qwen2-VL-style vision tower)

The turning point was a Cloudflare doc entry that first described Union Alpha as "a blended AI model that engages multiple language models in parallel for each request, synthesizes one answer". The text was later edited to describe only a model for coding, research and agentic workflows. The first post in the collection citing it came at 06:50 UTC+8 on 17 September (ScarletKc); after that, 16 of the guesses were "several models".

SuSu_酥酥👅 (@NFT_Chen) shared these before-and-after screenshots of the entry. They show the doc once described a multi-model system; they do not name the company behind it. (Item: Union Alpha is a parallel routing layer, not a new base model, community analysis claims)

YFarmX 🗞️ (@YFarmX) posted this slide at 12:51 UTC+8 on 17 September, about 18 hours before the reveal, marked "Working theory, not confirmed attribution." It was the only pre-reveal post in the collection to name Circuit & Chisel's Pareto. (Item: Union Alpha identity lead points to Circuit & Chisel's Pareto)

Its three clues: Cloudflare had briefly described a blended system, archived costs implied Pareto's published rates, and testing found multiple token-count patterns. The post drew about 3,800 views, against about 119,000 for the top GLM claim.

Table: Pre-reveal guesses against the official answer

Sample/method: Named guesses read by hand from the 158 pre-reveal posts (excluding 8 that only ruled labs out and 3 vague ones), one main answer each. The reveal is per OpenRouter's Union Alpha page, announcements from @OpenRouter and @unionalphaai, and Unbiased's site (checked 2026-10-18).

Guesses kept moving. renewable 🌏 (@goodworse) wrote "Union Alpha is GLM-5.5" at 04:04 on 17 September and, about 25 hours later and shortly before the reveal, "it turns out it's just a router between models". On 20 September, two days after the reveal, Aapakari (@aapakari) still asked: "Is Union Alpha stealth model in Opencode really GLM-5.5?"

Most testers liked the results, especially for 3D, but quality and speed swung widely, and 13% of all posts in the collection complained about speed or errors.

Sample/method: 174 of 177 posts in the union-alpha-cases collection (2026-09-16 21:00 to 09-19 00:00, UTC+8). Hand-curated, not every post on X. Post types AI-classified, then read one by one by the editor.

Table: Verdicts in hands-on posts, by task

Sample/method: 67 posts showing the poster's own test or a same-prompt comparison (posts from platforms, resellers and promoters removed; 13 with no clear task not listed). Verdict AI-classified and spot-read by the editor; one overall verdict per post.

3D and Three.js drew the most praise (14 of 22 positive); writing and reasoning tests were mostly mixed. Same-prompt comparisons with DeepSeek V4.1 Flash went both ways: AJ (@ItsmeAjayKV) judged Union Alpha clearly better on a 3D scene, while in a low-poly shooter test by SuSu_酥酥👅 (@NFT_Chen), a locally run DeepSeek V4.1 Flash finished a playable game and Union Alpha got stuck.

AJ (@ItsmeAjayKV) ran the same one-shot, single-file Three.js prompt on both models and posted Union Alpha's floating island first. Earlier the same day, AJ had called it "really close to DeepSeek-v4.1-Flash level"; this is one person's comparison. (Item: Union Alpha beats DeepSeek V4.1 Flash in single-file Three.js floating island test)

雪瑜 (@xueyu1125) ran the same "pelican riding a bicycle" SVG prompt several times; all four results here came from Union Alpha and differ clearly in style and finish. The post suggested it was either a weak model or several models mixed. (Item: Union Alpha may be a mix of multiple models, with pelican test results swinging wildly)

Others described the same swings. 王小波 (@wanxubo) called it a "抽卡模型" (a gacha model): the same task might finish in one minute or still be running after half an hour. Harshith (@HarshithLucky3) got "a very good 3d result but very bad SVG result". Posters later blamed the blend; Unbiased has not explained the variance.

Reported speeds ranged from 7.6 tokens a second (Astro Polo, @astropol0) to 300 to 400 (Ziwen, @ziwenxu_). Povilas Korop (@PovilasKorop) stopped testing after both attempts hit a 30-minute limit, on a task other models finish in about 5 minutes. About five hours in, @unionalphaai reported over 100 billion tokens served and about 500 million a minute, later 1 billion a minute. Unbiased's changelog says long completions (typically 60 to 110 seconds) had been hitting a 120-second ceiling on about 13% of non-streaming requests; the gateway now waits up to 300 seconds.

After the reveal, Paweł Huryn (@PawelHuryn) ran Pareto on a personal benchmark of 105 planted bugs across 2 repos: Pareto scored 30.7 for $4.81, against GPT-6 Astra 45 for $33.03 and Opus 5 27 for $51.33. That is a lower score than the top model at a fraction of the cost, from one tester and one task set.

On privacy, launch posts promised "Zero data retention"; OpenRouter's page now says prompts and completions "may be retained by the provider but are not used for training". maria (@maria_rcks) told people not to use it for sensitive work; that is an opinion, not a verified finding.

The free Union Alpha endpoint is gone; the same system is sold as Pareto, and the default has moved to 26.10.

Table: Pareto versions now (checked 2026-10-18)

Sample/method: Sources: OpenRouter's Unbiased models page and Unbiased changelog (2026-10-15 entry), checked 2026-10-18. Prices may change; check the official pages.

On Unbiased's API, pareto has pointed to 26.10 since 15 October; pin pareto-26.9 for the version that ran as Union Alpha. On OpenRouter, "Pareto Latest" always redirects to the newest version.

The benchmark charts shown at launch came from Unbiased and its launch partners, and independent reproductions are few, so test on your own tasks.

What this suggests for the next stealth model (editor's interpretation): a model's answer about who made it is weak evidence; tokenizer counts can mislead when several models sit behind one name; and partner docs and pricing pages gave the strongest clue here.