Union Alpha Is Pareto 26.9: Astra-Level Scores at 1B tok/min

Union Alpha surfaced on OpenRouter on Sep 16 2026 with no developer name and benchmark scores beside GPT-6 Astra and Claude Fable 5.1. One day later it unmasked as unbiased.ai Pareto 26.9 after serving a reported 1B tokens per minute.

• Stealth listing offered multimodal research, coding and agent work with a 262k context window
• A planned week-long free trial ended after one day under reported demand of 1B tokens per minute
• I route mystery models through eval, cost and provenance gates before any production traffic

Monday brought a leaderboard anomaly. A nameless model called Union Alpha sat beside GPT-6 Astra on coding and agent benchmarks with no lab, no paper, no weights page. By Tuesday the account joked about spare GPUs while serving a billion tokens a minute. By Wednesday it had a name: unbiased.ai Pareto 26.9. I have seen this movie before with OX Alpha, and I run the same three-gate drill every time. Here is the event plus the harness.

What happened on OpenRouter in 72 hours

Per reporting, OpenRouter listed Union Alpha as a stealth model on Sep 16 with the developer hidden, a standard practice for labs testing blind preference votes. The model card described multimodal research, coding and agent workflows at state-of-the-art levels across tasks. Reported benchmark scores sat at GPT-6 Astra and Claude Fable 5.1 levels. Context window measured 262,144 tokens. The launch promised a one-week free trial. Demand ended it in a day.

Traffic claims deserve careful reading. The model official account posted about processing 1B tokens per minute during the trial. That figure describes served throughput under free-trial flooding, not sustained capacity or quality. A billion mostly curious tokens differ from a billion paying production tokens. I treat vendor throughput posts as marketing until my own harness confirms latency distributions. Still, the signal stands: frontier-class anonymous supply draws instant floods.

The reveal followed the established stealth pattern. One day after release, Union Alpha identified as unbiased.ai Pareto 26.9. Our OX Alpha stealth-testing coverage documented the same arc with a different model: anonymous outperformance, then attribution. The pattern now repeats often enough to need procurement policy, not just curiosity. Stealth launches are a release strategy. Treat them as such.

For score context, our Muse Spark versus Gemini same-day showdown shows how frontier coding benchmarks separate contenders. Astra-level claims mean DeepSWE-class performance until proven on your suite. Nothing else counts.

graph TD
  A[Stealth model appears] --> B[Gate 1: identity + provenance]
  B --> C[Gate 2: golden-task eval]
  C --> D[Gate 3: cost + rate limits]
  D --> E{All pass?}
  E -->|yes| F[5% shadow traffic]
  E -->|no| G[Lab only, re-check weekly]

Step 1: Provenance before prompts

Mystery models carry mystery policies. Before routing anything sensitive, I establish four facts. Who operates it and under what terms. Where data flows and whether prompts train future models. What the rate limits and abuse policies state. Whether a data-processing agreement exists for customer content. Union Alpha triaged fast because OpenRouter terms applied during stealth, then unbiased.ai terms after reveal. The one-day free window meant most evaluators accepted clickwrap without reading. I read it. Training-data rights were the price of free, the same trade our Contributor SKU analysis flagged on Meta pricing. Free inference is never free. It is priced in data or attention.

First war story. During the OX Alpha episode I routed a client support pilot to the anonymous endpoint for a weekend test. Scores impressed. Monday brought attribution plus a terms change and a 3x price move. The pilot had customer tickets in prompts with no DPA covering the new operator. Legal spent two weeks on remediation. Now gate one blocks any non-public data until operator, terms and DPA status are written down. Stealth endpoints get synthetic benchmarks only. No exceptions. Not even for impressive weekends.

Step 2: Golden-task eval in one evening

Hype scores describe vendor harnesses. I run 60 golden tasks spanning our bug-fix, refactor and agent-tool patterns with fixed prompts and five rounds each, recording pass, tokens and latency. Pareto-class claims must survive contact with my suite before touching roadmaps. Methodology follows our reasoning effort benchmark harness: same prompts, same rounds, same cost-per-solved-task math.

File: requirements.txt

httpx==0.28.1
numpy==2.0.2
pydantic==2.8.0

File: config.py

from pydantic_settings import BaseSettings, SettingsConfigDict
from pydantic import Field

class Settings(BaseSettings):
    model_config = SettingsConfigDict(env_file=".env", extra="allow")
    openrouter_key: str = Field(alias="OPENROUTER_KEY")
    candidate: str = "unbiased/pareto-26.9"
    baseline: str = "gpt-5.6-sol"
    rounds: int = 5
settings = Settings()

File: gate_eval.py

import time, statistics
import httpx
from config import settings

TASKS = open("golden.txt").read().split("
---
")

def run_model(model):
    passes, bills, lats = 0, [], []
    with httpx.Client(timeout=300) as c:
        for t in TASKS:
            for _ in range(settings.rounds):
                t0 = time.time()
                r = c.post("https://openrouter.ai/api/v1/chat/completions", json={"model": model, "messages": [{"role": "user", "content": t}]}, headers={"Authorization": f"Bearer {settings.openrouter_key}"}, timeout=300)
                dt = time.time() - t0
                body = r.json()
                text = body.get("choices", [{}])[0].get("message", {}).get("content", "")
                passes += int("pass" in text)
                usage = body.get("usage", {})
                bills.append(usage.get("prompt_tokens", 9000) / 1e6 * 2.0 + usage.get("completion_tokens", 1800) / 1e6 * 8.0)
                lats.append(dt)
    total = len(TASKS) * settings.rounds
    avg = sum(bills) / len(bills)
    return {"model": model, "pass": round(passes / total, 3), "per_solved": round(avg / max(passes / total, 0.01), 4), "p50": round(statistics.median(lats), 1)}

if __name__ == "__main__":
    print(run_model(settings.baseline))
    print(run_model(settings.candidate))

uv venv --python 3.12 && source .venv/bin/activate
uv pip install -r requirements.txt
python gate_eval.py

Second war story. My first stealth eval compared the mystery model at max sampling budget against our baseline at throttled defaults. The newcomer won by 11 points and I briefed the team to migrate. A teammate reran with matched budgets and the gap shrank to 2 points inside noise. Same harness mistake as mismatched reasoning tiers, different costume. Matched configs or the comparison is theater. I gate the gate: eval diffs must declare sampling params before results count.

Cost projections need the same honesty. Free-trial pricing evaporates. I model candidate bills at three rates: launch promo, plausible steady state from comparable frontier models, and worst-case 3x. Pareto 26.9 at Astra-class quality prices against Astra-class rates until unbiased.ai publishes otherwise. Our provider arbitrage tables show identical weights selling 6x apart across hosts. New models deserve wider bands, not narrower.

Step 3: Shadow traffic, then commit or walk away

Passing gates earns 5% shadow traffic on non-critical paths with full tracing. Our background-thread tracing pipeline attributes every token, so shadow cost and quality report separately from production. Promotion needs two clean weeks: pass within 2 points of baseline and per-solved cost at parity or better. Anything else stays lab-only with weekly re-checks. Most stealth curiosities fail week two when rate limits tighten post-trial. Expect it.

Fable-class context frames the stakes. Our Fable versus Astra throughput study shows frontier margins deciding interactive products. A true Astra peer at sharper pricing reshapes routing tables. An unproven one reshapes incident reports. Gates separate the two.

When NOT to chase stealth models

Let's be clear. Curiosity is not a roadmap.

Skip production trials for regulated or customer-data workloads until operator terms and DPAs are signed. Synthetic benchmarks scratch the curiosity itch without compliance exposure. The model will still exist after paperwork. Hype will not.

Skip migration planning on trial-week pricing. Free tokens and launch rates expire by design. Plan at steady-state bands with 3x headroom. If the math fails there, the model is entertainment, not infrastructure.

Production bottlenecks I hit: stealth endpoints throttle brutally after trials end so capacity tests must rerun post-trial; model IDs churn on reveal breaking pinned references; rate-limit headers differ per host breaking unified backoff; attribution changes invalidate prior terms acceptance. Pin IDs post-reveal. Re-accept terms explicitly. Log everything.

Bottom line: stealth launches deserve a fast eval and a slow commit, in that order, every time.

By Deepak Bagada, Founder & Editor-in-Chief at Daily AI World. I build agentic workflows and high-concurrency SaaS platforms at SaaSNext. Follow my benchmarks on <a href="https://x.com/deeepakbagada">X @deeepakbagada and <a href="https://deepakbagada.in">deepakbagada.in.