Pareto 26.10 Preview

Pareto is a blended AI model. You send one request to one endpoint, several models work on it at once, and the system combines their results into one answer. Unbiased, the platform run by the lab Circuit and Chisel, released Pareto 26.10 Preview on 1 October 2026, two weeks after Pareto 26.9, at $0.80 per million input tokens and $3.20 per million output.

“Frontier intelligence, zero data retention, and open-source pricing are usually a pick-two. This is all three,” the launch post says. The post calls it a preview that “will continue to improve over the coming week”, and every Pareto 26.10 score on this page is Unbiased’s own preliminary run.

One request, several models

Every Pareto request goes to several models in parallel, and their results are combined according to the task, Unbiased’s documentation says. “You send one request and get one answer. The blend includes frontier and open models.” Unbiased keeps the models’ names private, and says “Model composition can change.”

Routers “read the prompt, estimate its complexity, and pick a single model”, Unbiased’s documentation says; Pareto engages several models on every request, and its home page says it never switches models mid-conversation. Pareto 26.9 first ran on OpenRouter as an anonymous model called Union Alpha, which YFarmX named as Pareto while it was still anonymous, before Unbiased confirmed it.

26.10 costs 57% to 88% less per token than 26.9

Unbiased prices Pareto 26.10 Preview at $0.80 per million input tokens, $0.03 per million cached input and $3.20 per million output, cuts of 68%, 88% and 57% on Pareto 26.9. The same rates appear on its pricing page and on OpenRouter. The launch post sets them against two frontier models:

“What won’t change: the price won’t go up, and the preview slug won’t break under you,” Unbiased says. Subscriptions run alongside the per-token rate: Personal at $10 a month for 25 million tokens a week, Personal Max at $100 a month for 300 million, and a pay-as-you-go Team plan. When 26.10 leaves preview, Unbiased says the weekly allowances double at the same prices.

Unbiased’s first scores for 26.10

On Unbiased’s preliminary runs of 1 October 2026, Pareto 26.10 Preview scores about the same as 26.9 on all four benchmarks, while each task costs between 17% and 83% less. The figures come from Unbiased’s model card, which says the 26.10 results “may change before final publication” and that the other results “are not independently validated”.

Unbiased adds that “Denominators and cost methodology require confirmation”, and that its 26.9 DeepSWE result used a 30-task slice of the benchmark.

Its claim is the cost per task

On raw score, Pareto 26.10 Preview sits below the leading models wherever Unbiased lists them; its claim is that no model scores higher for the same cost per task. “In these preliminary results, it matches or sets the Pareto frontier on all of them,” the launch post says, the frontier being the best score available at each price. On Humanity’s Last Exam, OpenAI’s GPT-6.1 Sol scores 52.9% at $0.078 a task, about ten times Pareto’s $0.008 for three more points.

GPT-6 Astra scores 96.3 on GPQA Diamond, 54.7 on Humanity’s Last Exam, 74.0 on DeepSWE v1.1 and 59.6 on Terminal-Bench 4.0, in Unbiased’s comparison.

The rival scores in that table are other labs’ and evaluators’ published figures, and Unbiased says they “use their own task sets and harnesses, so they are not directly comparable with our runs”.

A full zero-data-retention tier and a bigger window

Pareto 26.10 adds “a full zero-data-retention tier”, and Unbiased says all of its OpenRouter and Cloudflare traffic runs on it. Its terms, effective 24 September 2026, say traffic on that tier will have no prompt or response content retained or accessed after inference. Other traffic can be kept for up to 30 days, the terms say.

OpenRouter’s listing gives the preview a context window of 1,048,576 tokens, four times 26.9’s 262,144, with output capped at 131,072 tokens, and lists text and image input with tool use. Unbiased says 26.10 has “Lower latency across the board”, with “the biggest gains” on long agentic runs.

How to get it

Pareto 26.10 Preview runs on Unbiased’s own API and platform, where the changelog says “The model string stays pareto”, and on OpenRouter as unbiased/pareto-26.10-preview. OpenRouter’s listing advises using Pareto 26.9 for stable behaviour while the preview changes. Unbiased reviews and approves new accounts by hand. For the models Pareto measures itself against, see GPT-6 Sol and Luna, GPT-6 Astra and Claude Sonnet 5.5.