Aion 3.5 Mini Guide: A 262K Storyteller at a Quarter of the Price

The usual trade with a Mini model is capacity for cost: a smaller window, fewer features, a lower price. Aion 3.5 Mini drops only the last part of that. It has the same 262,144-token context as Aion 3.5, the same reasoning controls, the same GLM base and the same multi-model design. What changes is the price, plus whatever quality difference a smaller system brings, which AionLabs has not quantified.

The numbers side by side

• Input: $0.70 per million tokens, against $3.00 for Aion 3.5.
• Output: $1.40 per million, against $6.00.
• Cached reads: $0.18 per million, against $0.75.
• Context and output cap: 262,144 and 32,768 for both.
• Reasoning effort: low, high or max, defaulting to high, for both.

Those are the rates listed on OpenRouter and on AionLabs' model list at the time of writing, and AionLabs is the only provider. Both models arrived on 23 September 2026. The Mini has a dated snapshot, aion-labs/aion-3.5-mini-20260923, for anyone who wants the build pinned rather than following updates.

On every rate the Mini costs about 23 per cent of the full model. Put the other way, the question for any given scene is whether the full model's output is worth more than four times as much to you.

What changed from 3.0 Mini

The previous Mini, Aion-3.0 Mini, was built on the DeepSeek family while the full 3.0 sat on GLM. For 3.5, AionLabs put both on GLM. The price is unchanged at $0.70 / $1.40, and the window doubled from 131K.

Two consequences follow. Moving from 3.0 Mini to 3.5 Mini changes the base family, so expect a different voice, not only an improved one, and compare before migrating a running campaign. And the Mini and the full model in the 3.5 line are now closer relatives than they were in 3.0, which should make routing between them less jarring: a scene handed from one to the other is less likely to change register mid-story. That last point is an inference from the shared family, not a published claim, so check it on your own transcripts.

3.0 Mini keeps one capability the 3.5 Mini lacks: reasoning can be turned off completely. The 3.5 line always thinks, at least at low.

Where a smaller system tends to show

AionLabs does not publish the sizes of the models inside either system, so there is no specification to read. Experience with smaller models in general suggests where to look:

• Late continuity: facts established early in a long session contradicted later.
• Group scenes: several characters' voices blurring into one.
• Subtext: characters stating what they mean too plainly, where a stronger model would leave it under the dialogue.
• Repetition: recurring phrases and sentence shapes across a long run.

None of these show up in a five-turn trial. Test over the session length you actually run, and read blind. A/B testing two models covers how to compare the pair without your expectations deciding the result.

Mini by default, full model for the scenes that matter

Because both models share a context window, a request format and a base family, they suit a simple routing rule: send ordinary turns to the Mini and escalate the few that carry the story.

1. Default every turn to 3.5 Mini.
2. Escalate to 3.5 when the user starts a new chapter or scene, when a planned reveal is due, or when they regenerate a reply twice.
3. Keep the conversation history identical across both, so the switch is invisible apart from the prose.

Escalating on regeneration is a cheap quality signal. A user asking for another attempt is telling you the first one missed. Model routing and fallbacks covers how to wire this up, and how to stop an escalation path quietly becoming the default.

A worked cost example

A 300-turn campaign averaging 60,000 tokens of context and 450-token replies uses 18 million input tokens and 135,000 output tokens. With no caching:

• All on 3.5 Mini: $12.60 input plus $0.19 output.
• All on 3.5: $54.00 input plus $0.81 output.
• Mini with one turn in ten escalated: about $17 in total.

With good cache hit rates on a stable prompt prefix, all of those input figures shrink by roughly three-quarters, since cached reads cost about a quarter of fresh input on both models. Output barely registers in any scenario. For long-form roleplay, input price and cache discipline are the whole bill.

Decision rule

• Start on 3.5 Mini for any new project in this family.
• Move a scene type to 3.5 only after blind reading shows a difference you would pay for.
• Stay on 3.0 Mini if you need reasoning fully off, or if you prefer its DeepSeek-based voice and your sessions fit inside 131K.