You found us. Union Alpha — the stealth model on OpenRouter, OpenCode, and Cloudflare — is Pareto 26.9 from @TheUnbiasedCo. The internet needed one day to pull off the mask. We shouldn’t underestimate adversaries like @aitrackerbot.
We partnered with these platforms to do a stealth launch because we wanted to see what real users thought about it, learn what got tricky with scale, and bake that into our planned public launch of 26.10 next month. We don't keep your prompts or outputs, and we think zero data retention should be the default for more models.
Some of you thought you were testing one new model, and were surprised when others said it was multiple. That is on us. It is not on @OpenRouter, @OpenCode, or @Cloudflare. They worked with us to run a stealth listing the way stealth listings are always run.
Pareto is not a single set of weights. It is a system: several models, open and frontier, working the same task, with a harness that checks the work and calls in a stronger model when the task needs one. There was no training run. Earlier versions of Pareto taught us that benchmarks can mislead. Our private beta taught us to measure which models are exceptionally good at which tasks. We think the routers of the future will look and feel like models do today. Max intelligence when the task needs it, cheaper intelligence when you’re asking how to scramble an egg.
Within hours on Wednesday, demand hit billions of tokens per minute and the model was unusably slow. We're a small company, and we didn’t launch with enough compute to meet the moment. If you gave up on it yesterday out of frustration, we understand. AWS helped us triple the capacity overnight! Turns out that still wasn’t enough to support the interest.
We’re coordinating with the gateways to enable paid access for 26.9. The pricing is $2.50 for input, $0.25 cache, $7.50/output per million tokens – less than a quarter of Astra’s list price.
We believe systems like this are the path to more intelligence for less. It’s a new field of research: many models working as one. From here on, we pursue it in the open: https://unbiased.ai/model-card/.
Use 26.9. Tell us where it fails. We're excited to show you what comes next.
@TheUnbiasedCo