Union Alpha is Pareto 26.9: everything you need to know

The mystery lasts a little over a day: Union Alpha, the stealth model that appeared on OpenRouter, OpenCode and Cloudflare, which had already sparked a buzz for its coding level, is in fact Pareto 26.9, the system developed by The Unbiased.

To reveal it, it was the same company, after the account @aitrackerbot managed to trace its identity within twenty-four hours. Before the official confirmation, the community had already started suspecting it: an analysis of the tokenizer had ruled out the initial hypothesis of a Zhipu AI/GLM model, bringing Union Alpha closer to an Llama 3-type architecture.

Union Alpha is Pareto 26.9: it’s not a model, but an entire orchestrated AI system

The focal point of the reveal is precisely this: Pareto does not correspond to a single set of weights, like a traditional model, but to a blended system that runs multiple LLMs in parallel on each request, dynamically synthesizing the results based on the task, with a verification mechanism that calls upon a more powerful model only when the task truly requires it.

The Unbiased does not publicly disclose the specific models that compose the system at any given moment, because the combination would change too frequently to remain an accurate information, while stating that it will announce every change before bringing it into production.

So no training runs are planned: according to The Unbiased, earlier Pareto versions taught the company that benchmarks can be misleading, while the private beta phase helped identify which models were particularly well suited for specific tasks.

Technical specifications and declared benchmarks

Union Alpha, and thus Pareto, handles multimodal inputs of text and images, supports tool calling and works with a very wide context window: 262,144 tokens, with a maximum output of 131,072 tokens.

Based on the most recent public model card available (relating to version 26.8, not yet updated to 26.9 at the time of writing this article), Pareto achieves a composite score of 73.5 across seven public benchmarks at a cost of $0.10 per task, with results of 86.0 on Terminal-Bench 2.1, 86.0 on SWE-Bench Verified, 90.4 on GPQA-Diamond, 47.0 on Humanity’s Last Exam, 69.2 on arXivMath, 58.0 on DRACO and 77.9 on MMMU-Pro (the multimodal benchmark).

The comparison is made with models like Claude Opus 5, Claude Fable 5, GPT-5.6 Sol, Kimi K3 and DeepSeek V4F-0731. In terms of latency, The Unbiased reports no additional delay on agentive tasks, but up to three times slower on particularly complex reasoning tasks.

Why the stealth launch

The Unbiased chose to launch Pareto 26.9 anonymously on OpenRouter, OpenCode and Cloudflare to gather real-user feedback, understand what difficulties would emerge at scale, and use these learnings ahead of the public launch of Pareto 26.10, scheduled for next month (according to a source, the date indicated would be October 10, 2026, not confirmed directly by The Unbiased).

The company specifies that it does not store prompts nor outputs from users, and believes that the absence of data retention should become the standard also for other models.

A week-long free trial that, in fact, lasted only one day

The initial communication spoke of free and unlimited access for a week, but the reality was different: Union Alpha was launched on September 16, 2026, and already the next day, September 17, 2026, at about 32 hours distance, free access was halted to make way for the paid version, precisely due to the outsized demand generated in the first hours.

Same coding parameters as Astra, but at a price much lower

During the short free trial phase, Pareto 26.9 offered coding parameters comparable to Astra’s, but at a clearly lower cost.

Now that the company is coordinating with platforms to enable paid access, Pareto 26.9’s price list stands at $2.50 (about €2.18) per million input tokens, $0.25 (about €0.22) per million cached tokens, and $7.50 (about €6.53) per million output tokens.

Demand has exceeded available capacity

In the early hours of Wednesday, the demand reached one billion tokens per minute, rendering the model practically unusable due to its slowness. The Unbiased describes itself as a small-scale entity, and admits it did not start with sufficient computing capacity to sustain the interest generated. AWS helped the company triple capacity overnight, but even that was not enough to fully satisfy the demand.

How to try Pareto via Cloudflare

For anyone who still wants to try the system through Cloudflare AI Gateway, access requires a Cloudflare account with a domain you own, a credit card and a prepaid minimum deposit of $10 (plus a fee of about $0.50); the integration is via API endpoints compatible with the OpenAI standard, easily usable in existing workflows.

What happens next

The Unbiased considers systems like Pareto, based on the coordinated work of multiple models, the path to achieving more intelligence at a lower cost, and defines them as a new field of research.

From here on, the company states it intends to carry forward this research openly, publishing details on its official model card. The invitation to users is to continue using Pareto 26.9 and report where the system fails, ahead of the public launch of Pareto 26.10 next month.