Pareto 26.9 vs Unbiased's Pareto 26.10 Preview: The Stable Pin and the Preview

Unbiased's Pareto line now offers two callable releases with two different answers to the same question, and as of this week you have to pick one on purpose. Pareto 26.9 is the release the vendor's default alias resolves to: 262,144 tokens of context, $2.50 per million input tokens, no expiration date attached. Pareto 26.10 Preview is the opt-in build: four times the context, roughly a third of the input price, four benchmark scores the vendor itself calls preliminary, and an expiry date written into its catalogue record. Both are live. Both take text and images in and return text. Neither is a variant of the other you can flip with a flag at request time.

What changed is not either model. It is that the stable release finally has its own address. On October 13, 2026, the routed catalogue that lists this family began carrying unbiased/pareto-26.9 as a separately callable entry, alongside an alias entry that always redirects to whichever release is current. Before that, the only way to reach the stable model with certainty was to ask for Pareto and accept whatever the alias handed you. Now the two releases sit in the same list with their own prices, their own limits and their own expiry metadata — and choosing between them is a decision you make rather than one that gets made for you.

The two releases, dimension by dimension

Every published difference between them fits in these seven lines. Left is Pareto 26.9; right is Pareto 26.10 Preview.

• Context window — 262,144 tokens vs 1,048,576 tokens. Four times the room, with no smaller tier offered on either. • Maximum output — 131,072 tokens vs 131,072 tokens. The larger input window bought no larger answer. • Input price — $2.50 per million tokens vs $0.80. The preview is about 32% of the stable rate. • Output price — $7.50 per million tokens vs $3.20. About 43% of it. • Cached input — $0.25 per million tokens vs $0.03. The preview is roughly an eighth, which is the row that most changes a long agent's bill. • Published benchmark sheet — one score, on a partial task slice, vs four scores the vendor labels preliminary. • Stability terms — no expiry on the record, and the vendor calls this the default release vs an opt-in preview whose catalogue record carries an expiration date of October 15, 2026.

Two of those rows are doing most of the work. The cached-input rate is the one that decides whether a long agentic trajectory is affordable on either release, and the expiry field is the one that decides whether the preview is something you can build on at all.

Why the version numbers do not line up

The numbers are month labels, not a sequence you should reason from. Unbiased's own model card dates Pareto 26.9 to a September 20, 2026 run and describes 26.10 Preview as the October 1 release that followed. The catalogue that lists 26.9 shows it entering its record on October 13 — three weeks after the release date the vendor prints. And the vendor's launch post for the preview says the previous version "26.9" had been running for two weeks on third-party gateways under a temporary name before it was named.

Put those three statements together and the practical lesson is that a version number here tells you which month a build came from, and nothing about whether the build is still current, still supported, or still the one behind the alias. It also means the date you see on a catalogue page is a listing date, not a release date. Reading it as the latter is the single most common mistake made about this family, and it is now made in both directions: 26.9 looks three weeks newer than it is, and the preview looks five weeks older.

What the preview buys, and what the benchmark sheet actually says

The case for the preview is arithmetic. Four times the context for a third of the input price, with a cache-read rate low enough that a long trajectory costs a fraction of what it would on 26.9. Unbiased also claims lower latency on the preview with the largest gain on long agentic runs — a claim with no published measurements behind it, from either the vendor or anyone else.

The case against reaching for it reflexively is subtler and lives in the one place the two releases can be compared at all. On DeepSWE v1.1, the agentic coding benchmark both releases appear on, the vendor's model card shows Pareto 26.9 at 70.0% at $0.29 per task and Pareto 26.10 Preview at 69.9% at $0.24 per task. The newer, cheaper release scores marginally lower on the only benchmark where both have a number, and it wins on cost per task by about 17%. That is not a scandal and it is not a rounding error to be waved away — it is exactly what a cheaper blended system should look like, and it is worth knowing before you assume the preview is a strict upgrade.

Two caveats belong on that comparison and the vendor publishes both. The 26.9 figure comes from a 30-task slice, not the full DeepSWE v1.1 set, so it is not the same measurement as the preview's. And every number here is vendor-run: the preview's four scores — 92.4% on GPQA-Diamond, 49.9% on Humanity's Last Exam text-only, 69.9% on DeepSWE v1.1, 50.8% on Terminal-Bench 4.0 — carry the vendor's own "preliminary, may change before final publication" label, and the model card states that denominators and cost methodology still "require confirmation." Nobody outside Unbiased has published a score for either release. Treat the whole sheet as a reason to test, not a reason to switch.

The expiry is the part to plan around

The preview's record carries an expiration date of October 15, 2026 — one day after this was written. What happens on that date is not documented anywhere we could find: not in the changelog, not on the model card, not on the pricing page. Nothing in the vendor's material says the preview stops serving, and the launch post says the opposite of a shutdown in spirit — that "the price won't go up, and the preview slug won't break under you," and that on graduation 26.10 becomes the subscriber default while weekly allowances double, from 25M to 50M tokens on the $10 Personal plan and from 300M to 600M on the $100 tier.

So the honest reading is not "the preview dies tomorrow." It is that the preview is governed by terms that move, and that a build with an expiration field in its record is not the release to pin for a system you do not want to revisit. That is the whole argument for 26.9 as the default: it is slower to improve and it is priced three times higher on input, but it is the one the vendor calls the release rather than the experiment, and its record has no clock on it.

Holding both without rewriting the integration

The reason this is a routing decision rather than a modelling one is that both releases take the same request shape. Same endpoint family, same parameters, same text-and-image-in/text-out surface — the only thing that differs is the string you put in the model field. That makes the migration surface tiny and the pinning discipline the entire problem: exactly one value in exactly one place determines which of two very differently priced systems answers you.

That is the case the routing DSL on OrcaRouter was built for. You compose a call out of named legs and keep the model string as a config value rather than a constant buried in application code, so moving from a preview to a stable pin is an edit to the route rather than a deploy. The same is true of the failover chain: a preview that changes, expires, or starts returning errors becomes a fallback event rather than an incident. We do not carry either Unbiased release — this is a vendor-served family with one provider, the vendor itself — so the way to call it is Unbiased's own API, and if your goal is to compare the two releases against your own workload, both are reachable there today.

Which one to pin

If you are shipping something that has to answer next month without anyone touching it, pin Pareto 26.9. It costs three times more per input token and gives you a quarter of the context, and in exchange it is the release the vendor's alias points at, it has no expiration field, and its only published benchmark — a partial-slice score — happens to be the higher of the two. That is the boring answer and it is the right one for a production path.

If you are evaluating, or if your workload is long-context and cost-sensitive enough that a third of the input price changes what is viable, run Pareto 26.10 Preview deliberately and behind a fallback. A cached-input rate of $0.03 per million against $0.25 is the difference between an agent loop you can afford all day and one you ration, and the 1,048,576-token window is the only one of the two that fits a large repository or a long document set in a single request. Take the price, take the window, and keep the expiry date in a calendar entry rather than in your assumptions.

What would change this advice is narrow and checkable: an announcement that the preview has graduated and the alias has moved, an independent benchmark for either release, or the preview continuing to serve past October 15 on the same rates. Any one of those turns a two-way choice into a one-way one. Until one lands, the two releases in this family are genuinely different products with genuinely different risk profiles, and the only thing you control is which string your requests carry.

A release that can change under you mid-week is the case automatic failover was built for: automatic failover a preview that expires or starts erroring becomes a fallback event rather than an incident.