Union Alpha revealed as unbiased.ai's Pareto, free tier ends after one day

The stealth launch lasted barely a day

A model that circulated in developer channels as a mysterious free system reportedly stronger than GPT 5.6 was unmasked and put behind a paywall within roughly 24 hours, according to a dev.to post by Nokka (a post that states it was AI-drafted and human-edited). The post reconstructs the timeline: on 16 September 2026 the model appeared on OpenRouter under the id stealth/union-alpha, released by OpenCode, free to use, with a stated plan to keep it free for about a week. Its X account claimed it was serving a billion tokens per minute. On 17 September the operator revealed it was Pareto, built by unbiased.ai, and closed the free window early that same day. The author also found the stealth id has since been removed from OpenRouter entirely.

Why free access ended early

The company's explanation, as quoted in the post, is blunt: demand after launch far exceeded forecasts, hitting a billion tokens per minute within a single day and degrading response speed until the service was unusable. AWS tripled the available compute overnight and it still was not enough. The anonymous launch was deliberate — the company said it wanted to see how real users responded and to learn scaling lessons ahead of an official launch scheduled for 10 October 2026.

Not one model, and not a router

The detail the post flags as most consequential: Pareto is a blended model. Per the official description, several models are invoked in parallel on every request and their outputs are synthesized dynamically into a single answer. Cloudflare describes the system the same way, noting it handles text and images through one API.

unbiased.ai explicitly distinguishes this from a model router, which reads a prompt, judges its complexity and forwards it to a single model. Its argument is that dependable routing does not work in practice, and that switching models mid-conversation breaks prompt caching, with cache misses consuming the savings at scale. The members of the blend are not disclosed: the company says the composition changes frequently as new models are tested and weak ones dropped, and it commits to announcing routing or membership changes before they take effect.

What it costs now

The current release is Pareto 26.9 at $2.50 per million input tokens, $0.25 for cached input and $7.50 for output — figures the post says are consistent across the OpenRouter API, prices. and the company's own pages. The lower $1.25/$0.15/$6.25 numbers still circulating belonged to the earlier 26.8 release, so the price roughly doubled at the version change; the author initially misread the two as conflicting sources before tracing them to two different versions. Even at the higher price, Pareto 26.9 is the cheapest per token in the comparison the post pulled: Claude Fable 5.1 sits at $10/$50, Claude Opus 5 at $5/$25, GPT-5.6 Sol at $4/$20 and Kimi K3 at $3/$15 per million input/output tokens.

The benchmark picture is weaker than the hype

The 26.9 model card undercuts the launch chatter. Pareto ties at 74 on DeepSWE, trails GPT 6 Astra on Terminal-Bench 4.0 (51 vs 58), MMMU-Pro (78 vs 87), HLE without tools (49 vs 54) and ArXivMath (88 vs 91), and trails Claude Fable 5.1 on Terminal-Bench, MMMU-Pro and HLE. It does not top a single benchmark in the table. More importantly, the card states that measured task costs and a composite score have not been published for this release — meaning the widely shared 26.8 figures, such as 86.0 on SWE-Bench Verified at $0.2708 per task and roughly $0.10 per task averaged across seven benchmarks, describe a model that is no longer on sale. The eval harness itself is open on GitHub as circuitandchisel/pareto-evals.

Why it matters

Anyone who wired Union Alpha into a project on the assumption of free access lost that assumption within a day. More broadly, per-token pricing alone no longer answers whether a service is cheap: without cost-per-task numbers for 26.9, nobody can yet say whether the blend actually saves money, even at the lowest token price on the board. And the architecture is worth watching regardless of this vendor — parallel blended inference is a workaround for the routing and caching problems many teams face, and the open eval harness makes at least part of the performance story checkable.