Pareto 26.10 Is Generally Available: What Actually Changed When the Preview Label Came Off

Pareto 26.10 — the release Unbiased has been selling as Pareto 26.10 Preview since October 1, 2026 — went generally available on October 15, 2026. That is the whole of the news, and it is smaller than the word "generally available" usually implies. The model did not change. The price did not change. The benchmark sheet did not change, and, tellingly, the vendor still calls those numbers preliminary on the same page that carries the GA banner. What changed is the routing: the pareto model string now points at 26.10 by default, where two weeks ago it pointed at the previous release and 26.10 was something you had to opt into by name. The documented AI failures and agent incidents behind releases like this one are collected in the OrcaRouter AI Incident Archive where each entry carries its own sources, severity, confidence grade and disputed flag.

That distinction is worth the two minutes it takes to read, because it is the difference between a release you chose and a release that arrives. Anyone calling pareto on October 14 got the older model. Anyone calling it on October 16 got 26.10 — with a different context window, a different price, and a different model string underneath, without changing a line of code.

What moved on October 15

The vendor's changelog entry for that date is short and unambiguous, and it makes three claims worth separating.

• The default alias moved. The pareto string now selects 26.10. Unbiased's own wording is that the alias "now points at 26.10," which is a statement about routing, not about the model. • The dated strings still work. pareto-26.10 pins this release, and pareto-26.9 "remains available" for anyone who wants the previous one. The preview string is retired: preview users were told to switch to pareto or pareto-26.10. That is a migration you have to do by hand if you were on the preview. • Subscription allowances doubled at the same price. Personal went to 50M tokens a week and Personal Max to 600M, both unchanged in monthly cost at $10 and $100. The API rate card itself — $0.80 per million input, $0.03 cached input, $3.20 output — is explicitly "unchanged from the preview."

So the GA event is an alias flip, a retired string, and a bigger weekly allowance. Nothing in that list makes the model better. Two of the three make it cheaper or easier to use, which for most readers is the point.

The spec sheet, as documented

Pareto 26.10 is a text-and-image-in, text-out model with a 1,048,576-token context window and a 131,072-token maximum response. Both of those are current fact as published; neither is independently measured, and Unbiased's own pricing page does not print a context figure at all — you get the number from the model's technical listing record and from the vendor's model card, which is a small documentation gap worth knowing about before you size a workload against it.

On the parameter surface, the documented set is max_tokens, temperature, top_p, tools and tool_choice. That last one is restricted: you can let the model decide whether to call a tool, but you cannot force a call the way you can on most frontier APIs. There is no response_format in the published parameter list either, which matters if your pipeline depends on structured output. There is no Hugging Face identifier anywhere in the record, so the weights are closed, and the vendor describes Pareto as a composite — a blended model that runs several underlying models against the request and keeps the result it judges best. That is a claim about mechanism, made by the company selling it, with no external verification.

The context jump is the part worth pausing on. The previous release carried 262,144 tokens; 26.10 quadruples it to 1,048,576. Roughly speaking, that moves a codebase-sized or transcript-sized workload from "you will need to chunk this" to "you can send it." It is also the single strongest argument for reading 26.10 and 26.9 as genuinely different models rather than a version bump, which is the question both of the comparison pages in this series work through.

The benchmark sheet still says preliminary

Four scores are published for 26.10, all of them from October 1 runs, all of them vendor-run, and all of them carrying the label the vendor wrote for them.

• DeepSWE v1.1 — 69.9%, at a reported mean of $0.24 per task. • Terminal-Bench 4.0 — 50.8%, at $0.48 per task. • Humanity's Last Exam, text-only — 49.9%, at $0.008 per task. • GPQA-Diamond — 92.4%, at $0.004 per task.

The vendor's model card calls this a "preliminary comparison," states that the 26.10 results come from October 1, 2026 runs and "may change before final publication," and notes that denominators and cost methodology "require confirmation." The GPQA-Diamond figure in particular is the eye-catching one — 92.4% would put a blended router model in territory usually reserved for frontier reasoning systems — and it is precisely the number that has no independent counterpart. No evaluator has published a rerun of any of these four as of October 18, 2026.

There is also a comparison on the same card that the vendor itself describes as not independently validated, in which the previous release scores higher on the one benchmark both have in common: 26.9 at 70.0% on DeepSWE v1.1 against 26.10's 69.9%. That gap is a tenth of a point, on a 30-task slice rather than the full evaluation set, and it should be read as "these two are indistinguishable on this measurement," not as a ranking in either direction. If you take one thing from the benchmark sheet, take that: the published numbers do not separate the two releases, so the reason to move to 26.10 is context window and routing, not score.

What GA actually buys you, and what it doesn't

Three practical consequences, in order of how likely they are to bite.

First, if you were pinning the preview string, your calls are now a maintenance item. The preview is superseded; the vendor points you at pareto or pareto-26.10. Anyone who treated the preview string as a stable snapshot — a reasonable instinct for a preview — now has a model string that no longer represents the current release.

Second, if you were calling the bare pareto alias and assuming it was stable, that assumption was always wrong and is now visible. You moved from 262K context to 1M context and from the September rate card to the current one without asking. The vendor's docs note that unknown model identifiers are not rejected yet — a typo is still served and billed as Pareto — so the alias behaviour is the only thing standing between you and a silent version change. If reproducibility matters, pin the dated string; this is the case for it in one sentence.

Third, the doubled allowances change the arithmetic for steady workloads. At 50M tokens a week for $10 a month on Personal, a heavy single-user coding or research load can sit inside a subscription rather than on per-token billing. That is not a claim about the model, and the allowance is a cap rather than unlimited usage, but it is the part of the October 15 entry that changes what you pay.

Where a router fits, honestly

OrcaRouter does not carry any Unbiased model. If you want Pareto 26.10, you call it through the vendor's own API — that is the only route, and we are not going to imply otherwise. What a routing layer is for here is the decision around it: the same key and the same integration for both sides of a comparison, so evaluating 26.10 against what you already run does not mean a second contract and a second SDK, and a model that changes underneath you mid-week is a config change rather than an outage. That last point is what pareto flipping from 26.9 to 26.10 on October 15 looks like from the outside: a vendor-side alias move that landed with no signal in the request. A router's automatic failover and routing DSL exist for the cases where a dependency like that starts erroring or timing out, and where the fallback target should already be named before it does.

OrcaRouter passes the provider's list price straight through with 0% markup, so a vendor rate change is live on our side the same day it lands — which is a real advantage in general and not a reason to route a model we do not serve.

Who should move this week

Move to pareto-26.10 if you need the 1M-token window. That is the concrete, checkable difference between the releases, and it is the only reason on the table that a benchmark sheet cannot argue with. Pin the dated string rather than the alias while you do it, so the next alias flip is something you schedule.

Stay on pareto-26.9 if you are mid-project and your context fits in 262K. The two published scores are a tenth of a point apart on the vendor's own harness; there is no evidence in the public record that moving improves output quality. You also keep a string that the vendor says remains available.

Wait if you are still on the preview string and unsure: the migration is a one-line change, but it is the only part of this release where doing nothing has a defined failure mode.

What would change this read

Three things, none of them visible yet.

An independent benchmark run. Every number in this article is Unbiased's own, on Unbiased's harness, with Unbiased's caveats attached. A rerun by an outside party is the single most useful thing that could appear, and until it does, the 92.4% GPQA-Diamond figure should be quoted as a vendor-reported result or not quoted at all.

A dated price change. The card has moved before without a changelog entry explaining it, which is a better deal and a worse signal. GA carried the rates over explicitly; the next time they move, whether the vendor dates it tells you how much of this model's pricing you can plan against.

A second serving path. The model is served by one provider, so there is no failover target inside the vendor's own stack. For anything where 26.10 is the critical path, that is a design constraint to work around rather than a detail to file.

Questions that keep coming up

Is Pareto 26.10 a new model or the preview with a new label? Functionally it is the preview release: same context window, same pricing, same four published scores, same model underneath. What GA changed is the routing — the pareto alias now selects it — plus the doubled subscription allowances and the retirement of the preview string.

Does the generally-available label mean the benchmarks are final? No, and this is the detail most coverage of the October 15 changelog will skip. The same vendor page that announces general availability continues to describe the four scores as a preliminary comparison and to say the results may change before final publication. General availability describes the model's status as a product; it does not upgrade the evidence about its quality.

What does the alias change actually cost someone who ignores it? Nothing, if they wanted the newest release. If they wanted the version they tested, it costs them a context window they did not choose and a rate card they did not review — both of which are recoverable by pinning, and neither of which can be reconstructed after the fact from a bill.