Aion 3.5 Is Out: The Context Window Doubled, the Price Did Not, and the Small Tier Changed Families

AionLabs published Aion 3.5 and Aion 3.5 Mini on 23 September 2026, and the rate card is the first surprise. Both carry the exact prices their 3.0-generation predecessors carried: Aion 3.5 at $3.00 per million input tokens, $0.75 cached and $6.00 output; Aion 3.5 Mini at $0.70, $0.18 and $1.40. Those are the same numbers the vendor's model list has shown since May. What moved instead is the context window, which doubles from 131,072 to 262,144 tokens on both tiers, and what changed underneath is the model family the smaller tier is built on.

Everything below comes from AionLabs' own published model list and documentation, read directly today, plus the served configuration metadata on the API the models are actually reachable through. There is no independent benchmark of either model yet, which is the single most important thing to know before reading a spec sheet — and it is discussed at length further down rather than buried.

What actually shipped

Two models, both dated 23 September 2026 on the vendor's own list, both text-in and text-out, both flagged as reasoning models.

• Aion 3.5 — $3.00 input / $0.75 cached input / $6.00 output per million tokens, 262,144-token context, 32,768-token maximum output. Described by the vendor as a multi-model system built on the GLM family of models.

• Aion 3.5 Mini — $0.70 / $0.18 / $1.40 per million tokens, the same 262,144-token context and the same 32,768-token output ceiling. Described as a multi-model system built on the GLM family.

Both are reachable through an OpenAI-compatible endpoint at the vendor's own API, under the identifiers aion-labs/aion-3.5 and aion-labs/aion-3.5-mini. The served configuration exposes reasoning at three effort levels — max, high and low — with high as the default, and reasoning is not optional: the model reasons on every call. Tool calling, JSON response format and temperature are all accepted parameters. Image input is not; the architecture declares text to text.

What changed from 3.0, in three lines

The 3.0 generation is still listed and still priced identically, which makes the delta unusually easy to read.

• Context — 131,072 tokens on Aion 3.0 and Aion 3.0 Mini, against 262,144 on both 3.5 tiers. Doubled, and doubled on the cheap tier as well as the expensive one.

• Price — unchanged. Aion 3.5 costs what Aion 3.0 costs, on every one of the three rate lines. This is a refresh that does not ask existing callers for more money per token.

• Family, on the small tier only — Aion 3.0 Mini is described by the vendor as built on the DeepSeek family. Aion 3.5 Mini is described as built on the GLM family. The flagship was GLM-family in both generations; the mini moved.

That third line is the one worth pausing on. AionLabs built its cheap tier on one family and its expensive tier on another through the whole 3.0 generation, and as of today it builds both on GLM. Whatever the ensemble's internal division of labour used to be, the small model is no longer drawing on a different lineage than the large one.

The architecture claim, and what it does not tell you

"Multi-model system" is the phrase the vendor uses, and the description is specific about the mechanism: a collaborative generation process in which multiple specialised models each contribute to a response, producing stronger narrative structure and more compelling tension and conflict. That is a real architectural position, not marketing shorthand for a bigger single model.

It also leaves three things unstated, and they are the three things a buyer would most want.

The vendor does not publish how many models are in the ensemble. It does not publish which models they are, beyond naming the GLM family. And it does not publish how the work is divided — whether every constituent model runs on every call, or whether a router selects among them per prompt.

That last unknown is the one with money attached. A system that runs five models on every request and returns one answer costs its operator five times what a single model costs, and the flat $3.00 / $6.00 rate is a bet by the vendor that the blend averages out. A system that routes to one constituent per request costs roughly what one model costs. From outside, the two look identical, and only one of them is priced generously. The 262,144-token context is the tell that something is genuinely larger here than a repackaged Aion 3.0, but it is not a tell about how many forward passes a call triggers.

Nobody has measured either model

Artificial Analysis has no page for Aion 3.5, Aion 3.5 Mini, Aion 3.0 or Aion 2.0. None of the four appears in the model index. No SWE-bench Verified, Terminal-Bench, GPQA or MMLU-Pro figure is published by the vendor either. The launch materials carry no benchmark table at all.

That is not unusual for a roleplay-and-storytelling system, and it is not a red flag on its own — the evaluations that matter for narrative work are not the ones a coding leaderboard runs. But it does mean the two things every other model page in this price bracket can show you are missing here: a capability number and a cost-per-completed-task number. Aion 3.5's $3.00 / $6.00 rate is a list price with no measured denominator attached. You cannot yet say what a finished job costs on this model, because nobody has run a standardised job on it in public.

What can be said is narrower and still useful: the price is unchanged from a model that has been in production since May, and the context doubled. A caller who was already on Aion 3.0 is being offered more window for the same money with no migration cost beyond the model identifier. A caller who was not is being asked to take an unmeasured model at a rate three times what several measured general-purpose models charge for input.

Calling it without betting on it

The honest routing advice for an unmeasured model is not "wait" — it is "try it on a path you can afford to lose." Aion 3.5 sits behind an OpenAI-compatible interface, so swapping it in is a base-URL and model-name change, and the same is true of the general-purpose models you might use as a fallback. A gateway that fronts several hundred models on one key lets you put Aion 3.5 first for the prompts where its architecture is the reason you are calling it, and declare a measured model as the fallback for everything else — a fallback chain that absorbs a 429 or an outage before the response starts rather than after it fails.

One pricing note belongs here, because it is the kind of thing that changes on a vendor's schedule rather than yours. AionLabs has held this rate card flat across a generation, which is a good sign for anyone modelling spend a year out. Where a vendor does cut prices, a pass-through router bills the provider's list price with nothing added per token, so the cut is live on the routed side the same day rather than at the next billing cycle.

What is not true, and is worth stating plainly: Aion 3.5 is not served on every platform that hosts its neighbours. It is available through AionLabs' own API and a small number of third-party platforms. Check the catalogue before you assume the identifier you used for Aion 3.0 resolves.

What to watch next

Three things would settle the open questions, and all three are observable. An independent evaluation appearing for either tier, which would give the rate card a denominator. A vendor note on how many models the ensemble runs per call, which would explain the pricing. And a 3.5-generation replacement of the DeepSeek-family mini, if AionLabs keeps two families in the lineup after moving one of them to GLM.

Until the first of those lands, the defensible summary is this: Aion 3.5 is a same-price, double-window refresh of a specialist system, released today, with a documented architecture and an undocumented scoreboard. If narrative work is the job, that may be enough to try. If a benchmark is the thing you need before you commit, this model does not have one yet — and the models that do are all general-purpose.