Aion-3.0 Guide: AionLabs' Multi-Model Storyteller on GLM

Aion-3.0 is the release where AionLabs stopped shipping a single fine-tuned model and started shipping a system. Its predecessor, Aion-2.0, is one DeepSeek V3.2 variant. Aion-3.0 is described as a multi-model roleplaying and storytelling system built on the GLM family, using "a collaborative generation process in which multiple specialized models each contribute to a response".

That sentence is most of what is public, and it raises more questions than it answers. This guide covers what can be confirmed, what the design implies for anyone building on it, and why the arrival of Aion 3.5 at the same price changes the decision.

The confirmed specification

• Base: the GLM family from Z.ai. Which GLM models, and how many, is not published.
• Context: 131,072 tokens, with output capped at 32,768.
• Capabilities: tool calling, JSON mode, and reasoning with effort levels none, low, medium, high and max, defaulting to medium, per AionLabs' model list.
• Price: $3.00 per million input tokens, $6.00 per million output and $0.75 per million cached input reads, as listed on OpenRouter and on AionLabs' own model list at the time of writing.
• Availability: on OpenRouter from 7 July 2026, with AionLabs as the sole provider. AionLabs' list dates the model 5 May 2026.

OpenRouter also carries a dated snapshot, aion-labs/aion-3.0-20260707, for anyone who needs the behaviour frozen. There are no published weights, parameter counts or benchmark results. AionLabs claims stronger narrative structure, more compelling tension and conflict, and a nuanced treatment of mature and darker themes. Those are claims to test, not facts to quote.

What multi-model means for you as a caller

From the outside, Aion-3.0 is one endpoint. You send a chat completion and get one response. Behind it, several models contribute. AionLabs has not described the pipeline, but any design of this kind has consequences you can observe without knowing the internals.

• Latency is not one model's latency. If contributions run in sequence, a planner then a writer then a reviser for example, time to first token grows with each stage. Measure it with streaming on, because the gap before the first visible word is what a reader feels.
• Behaviour can change without the id changing. A component can be swapped, or the way contributions are combined adjusted, behind the same model name. The dated snapshot is the defence; see pinning model versions.
• The bill is quoted on your tokens. Pricing is per input and output token on your request, like any other model on the listing. Whatever runs internally is priced into that rate. Check the usage block on a few responses to see how reasoning tokens are counted.
• Repeatability is weaker. Several sampled stages amplify ordinary sampling variance. Do not expect the same prompt to reproduce a scene.

The general pattern, splitting a job across specialised models, is covered in multi-model strategies. Aion-3.0 is that idea packaged so the caller does not have to build it, at the price of not being able to see or tune it.

The price step from 2.0

Aion-3.0 costs 3.75 times as much as Aion-2.0 on both input and output. For roleplay, input is where that bites, because every turn resends the whole conversation.

A worked example: a session of 200 turns where the context averages 40,000 tokens and each reply is 400 tokens. That is 8 million input tokens and 80,000 output tokens. On Aion-3.0 the uncached cost is $24.00 for input plus $0.48 for output. On Aion-2.0 the same session is $6.40 plus $0.13. Cached reads at $0.75 per million would bring the Aion-3.0 input line down to about $6 if nearly all of each prompt is a cache hit, and that is the main lever you control.

The point is not that $24 is a lot. It is that the output price, the figure most people compare, is almost irrelevant for this workload. Compare roleplay models on input price and cache behaviour.

Aion-3.0 now that Aion 3.5 exists

Aion 3.5 arrived in September 2026 at the same $3.00 / $6.00 price, on the same GLM family, with the context window doubled to 262,144 tokens. On paper that makes 3.5 the default for new work, and the Aion 3.5 guide covers it in detail.

Two reasons remain to keep 3.0. The first is the effort range. AionLabs lists 3.0 with effort down to none, while 3.5 starts at low and defaults to high. If you want the fastest possible replies with no thinking pass, 3.0 can do that and 3.5 cannot.

The second is continuity. A long-running campaign, or a product with prompts tuned against 3.0's voice, can stay on the snapshot until you have compared the two side by side. If neither reason applies, there is little case for starting a new project on 3.0.

Testing a system you cannot inspect

Because the internals are hidden, judge Aion-3.0 on transcripts alone. A test that suits a multi-model narrator:

1. Pick one long scenario and run it for fifty turns on 3.0, on Aion-2.0 and on 3.5, with the same system prompt and the same user turns replayed from a script.
2. Record time to first token and total time per turn, not just the text.
3. Score structure: did scenes have a set-up, a turn and a consequence, or did the story drift?
4. Score continuity: plant three facts early and check whether they are honoured after turn forty.

Scripted user turns matter. If you improvise, you will steer each model differently and end up comparing your own prompts. The extra cost over 2.0 is justified only if the structure and continuity scores move, and only you can see that in your own genre.