Mystery model Union Alpha hit a billion tokens a minute before vanishing from listings and being revealed as Pareto

The argument worth reading is the one against routing. A router reads a prompt, scores its complexity and picks one model ^[8]. unbiased.ai says it is not doing that, and its explainer, in Nokka's Thai rendering, says it would already be doing so if that could be done reliably ^[8]. The obstacle it names is the cache: swap models mid-conversation and the prompt cache goes with them, and at scale the misses eat the whole saving ^[9]. Cloudflare's description of Union Alpha matches unbiased.ai's description of Pareto, several models called in parallel on every request and synthesized into one answer over a single API for text and images ^[5][4]. The wording here is translated from a Thai-language post drafted with deepseek-v4.1-flash through Hermes Agent and edited by Nokka ^[25].

That design puts a floor under the vendor's cost. Every served request pays for several model calls, and the customer is billed for one answer ^[4]. The published cached-input rate of $0.25 per million tokens only pays off if the same blend members keep seeing the same prefixes ^[13]. Prefix stability is what a router gives up ^[9]. Nokka reports Pareto's per-token price is still the lowest in the competitor table he pulled from prices.json ^[21].

His first draft reported a conflict, $1.25/$6.25 against $2.50/$7.50, and said he could not tell which was live; on recheck he had misread two versions as two sources, and he left the error in the published text on purpose ^[14][24]. Pareto 26.8 launched at $1.25 input, $0.15 cached and $6.25 output per million tokens ^[12]. The version serving now is 26.9 at $2.50, $0.25 and $7.50 ^[13][15]. He confirmed the current figures in four places that agree, among them /how/, /model-card/, prices.json checked on 17 September, and the OpenRouter API ^[19][17][18]. The /pricing/ page changed while he was editing, from the 26.8 numbers with a competitor note dated 1 August to the 26.9 numbers dated 17 September ^[16]. The page that announced 26.8 now returns nothing but its title, so the launch prices have no verifiable source left ^[20].

Input doubled and output rose by a fifth ^[1]. For a request of 10,000 input tokens and 1,000 output tokens, the bill moves from $0.01875 to $0.0325, up 73 percent ^[2]. Retrieval and long-context callers take close to the full doubling; short-prompt chat takes close to the 20 percent ^[1].

The benchmark line is the part I would not carry into a plan. That first draft cited 86.0 on SWE-Bench, 90.4 on GPQA, 77.9 on MMMU-Pro and a 73.5 average ^[22]. Those numbers describe a blend whose membership unbiased.ai declines to publish, on the grounds that a published list would be out of date within a few weeks ^[10]. What the company offers instead is procedural, a commitment to announce routing or membership changes before they go into service ^[11]. For a September score to hold in October, the blend would have to be the same blend. unbiased.ai says it tests new models as they ship and drops the ones it no longer wants in the blend ^[10].