Aion 3.5 Guide: 262K Context and Always-On Reasoning for Fiction

Aion 3.5 is AionLabs' current flagship for roleplay and storytelling, released on 23 September 2026 alongside a cheaper Mini. It keeps the formula of Aion-3.0, a multi-model system built on the GLM family in which several specialised models contribute to each response, and it keeps the price. Two things changed that matter: the context window and the reasoning controls.

What changed from 3.0

• Context: 262,144 tokens, up from 131,072. The output cap is unchanged at 32,768.
• Reasoning effort: AionLabs lists three levels, low, high and max, with high as the default. Aion-3.0 offered none, low, medium, high and max, defaulting to medium.
• Price: unchanged at $3.00 per million input tokens, $6.00 per million output and $0.75 per million cached reads, as listed on OpenRouter and on AionLabs' own model list at the time of writing.
• Everything else: text only, tool calling, JSON mode, AionLabs as the only provider, and no published weights or benchmarks.

OpenRouter also carries a dated snapshot, aion-labs/aion-3.5-20260923. The undated id will follow future updates; the dated one will not, which is what you want under a product whose prompts were tuned against this exact build.

Twice the context, and what to fill it with

For roleplay, context is the campaign. Every turn resends the system prompt, the character sheets, any world notes and the conversation so far. At 131K a long campaign eventually forces you to summarise or drop history. At 262K you can carry roughly twice as much before that happens.

The best use of the extra room is usually not more raw transcript. It is structured material the model would otherwise lose: a running ledger of facts, relationships and open threads, kept near the top of the prompt and updated every few scenes. A model reads a ledger more reliably than it recalls a detail from 150,000 tokens back, because quality degrades across long contexts well before the advertised limit. Long context degradation covers how and where that happens.

Treat 262K as headroom, not a target. Fill it only when the material earns its place, because every token in it is paid for again on every turn.

Reasoning you cannot switch off

The quieter change is that 3.5 has no none setting. The lowest effort is low, and the default is high. Every reply carries some thinking before the visible text, and by default quite a lot.

For a narrator that is a reasonable bet. Planning helps with continuity, foreshadowing and keeping several characters' motives straight. The costs are latency before the first word appears, and output tokens. How reasoning tokens are reported and billed for this model is something to check in the usage block of a few real responses rather than assume.

In practice, try low for rapid dialogue exchanges and keep high for scene transitions and turns where the plot moves. If your product needs near-instant replies with no thinking pass at all, 3.5 cannot provide it. Aion-3.0 and the DeepSeek-based models in the family can.

Cached reads decide the bill

A long campaign at 3.5's prices is an input-cost problem. Take one turn deep into a session, with 150,000 tokens of context and a 500-token reply:

• Uncached input: 150,000 x $3.00 per million = $0.45
• The same input read from cache: 150,000 x $0.75 per million = about $0.11
• Output: 500 x $6.00 per million = $0.003

Over a hundred such turns that is $45 against roughly $11. The reply is a rounding error either way. What you control is the cache hit rate, and that comes from keeping the front of the prompt stable: system prompt and character sheets first, unchanged between turns, with new material appended at the end.

The common mistake is rewriting a summary at the top of the prompt every turn. That invalidates the cache on every turn and puts you back on the $45 line. Prompt caching explained covers the mechanics.

3.5 or 3.5 Mini

The Mini has the same 262K window, the same effort levels and the same GLM base, at $0.70 / $1.40: under a quarter of the full model's price. Neither model has published quality figures, so the only honest way to choose is to compare them on your own material.

A sensible default is to run the Mini for ordinary turns and send the scenes that carry the story, such as openings, reveals and climaxes, to 3.5. If you cannot tell the difference in blind reading, stay on the Mini. The Aion 3.5 Mini guide covers that routing in more detail.

A short test plan

1. Take a real campaign log past 100,000 tokens and replay its last twenty user turns on Aion-3.0, with the log trimmed to fit 131K, and on 3.5 with it whole.
2. Check whether 3.5 uses the material only it could see: names, unresolved threads, earlier promises.
3. Run the same replay on 3.5 at low and at high, recording time to first token.
4. Check cache hits in the usage data for each run.

If the longer window does not show up in the story, the window is not your bottleneck, and the Mini at a quarter of the price deserves a serious look.