What is Aion 3.5?
Aion 3.5 is a role-play and storytelling model from Aion Labs, the brand of Deep Forge sp. z o.o., a company registered in Warsaw. It was released on 23 September 2026 alongside a cheaper Aion 3.5 Mini. Rather than one model, it is a small ensemble built on Z.ai's GLM family: several specialised models each contribute to a reply, which Aion Labs says gives stronger narrative structure, tension and conflict. Earlier versions were Aion-2.0 (February 2026, on DeepSeek) and Aion-3.0 (July 2026).
It takes text only, with a context window of 262,144 tokens and up to 32,768 output tokens, and supports reasoning, tool calls and structured output through an OpenAI-compatible API.
On OpenRouter, where Aion Labs is the only provider, Aion 3.5 costs $3 per million input tokens and $6 per million output tokens, with cached input at $0.75; Aion 3.5 Mini costs $0.70 and $1.40.
How much does Aion 3.5 cost?
| Plan | Price | What you get |
|---|---|---|
| Free | $0 | Daily credit allowance for API or browser chat; no card required |
| Aion 3.5 | $3 / $6 per 1M tokens | Input / output; cached input $0.75 |
| Aion 3.5 Mini | $0.70 / $1.40 per 1M tokens | Input / output |
Aion 3.5: $3 input / $6 output per 1M tokens, cached input $0.75; Aion 3.5 Mini: $0.70 / $1.40. Free tier ($0) with a daily credit allowance.
Is Aion 3.5 worth it?
Aion 3.5 is one of very few current models built only for role-play and fiction. Most people who role-play with AI use general models: on OpenRouter's role-play board the most used are DeepSeek V4.1 Flash, Tencent's Hy4 Preview and GLM 5.3 Flash, none of them made for it. Those models are tuned to be helpful and brief, which is the wrong instinct for a story. They agree with you, resolve conflict too soon and wrap up scenes you wanted to keep going. Aion Labs' answer is to split the job: several GLM-based models each contribute to a reply, aiming for structure, tension and conflict.
The specification suits long stories. A 262K-token context holds a novel's worth of chat history and character notes, and replies can run to 32,768 tokens. Tool calls and structured output mean a game can ask it for a scene and a JSON record of who is where.
The catch: there is no evidence it is better. Aion Labs publishes no benchmark, no sample comparisons and no description of how many models take part or what each does. MiniMax, by contrast, published a 300-session test for its own role-play model, M2-her. The ensemble also makes Aion 3.5 expensive for the category: $6 per million output tokens is twelve times DeepSeek V4.1 Flash's $0.50, and Aion Labs is its only provider on OpenRouter, so there is no fallback if it goes down.
Privacy is better than most. Aion Labs' policy says it does not store the content of API or chat requests, keeps only metadata such as time, model and token counts for up to 12 months, and does not train on your content. The service is for adults only. The policy page sets no rules on explicit content; those sit in the separate terms.
Our take: role-play is the one job where a model's personality matters more than its test scores, so the test that counts is your own. Run the same opening scene through Aion 3.5 Mini at $1.40 per million output tokens and through DeepSeek V4.1 Flash, read both, and pay for the full model only if the difference is obvious to you.
Where it earns its place
A solo developer of a text adventure game
Generate non-player characters who push back instead of agreeing with the player
Call Aion 3.5 through its OpenAI-compatible API with a character sheet in the system prompt and ask for a scene plus a JSON state update using structured output. At $6 per million output tokens, a 1,000-token reply costs about $0.006.
A novelist using AI to draft scenes between sessions
Keep a long story's characters and plot threads consistent over months
The 262K-token window holds the full chat history and notes without summarising; the free daily credit on Aion Labs' site is enough to judge the writing before paying.
A product manager at a companion-chat app
Decide whether to route premium users to a specialised model
A/B test Aion 3.5 Mini against the current default on retention and regeneration rates; the policy of not storing request content simplifies the privacy review, but a single provider means planning a fallback model.
Before you commit
- Setup
- Sign up at aionlabs.ai (no card for the free tier) or use an existing OpenRouter key with model aion-labs/aion-3.5; any OpenAI-compatible client works.
- Integrations
- OpenAI-compatible API; on OpenRouter, Venice and AI/ML API; supports reasoning, tools, structured output and streaming. Works in front ends such as SillyTavern through OpenRouter.
- Data and privacy
- Aion Labs' privacy policy: request and response content is not stored, metadata kept up to 12 months, no training on your content. Email Agents are the exception and store their content until deleted. Adults only. Not available with OpenRouter's EU in-region routing.
- Cost at scale
- At $3 input and $6 output per 1M tokens, a role-play session that resends a 50K-token history for 100 replies of 500 tokens costs about $15.30 before caching; cached input at $0.75 cuts the history cost by three quarters.
- Limits
- Text only; no published rate limits, benchmarks or ensemble details; one provider.
What’s good
- Built specifically for role-play and long-form fiction
- 262K-token context and up to 32,768 output tokens
- Policy says request content is not stored or used for training
- Cheaper Mini version at $0.70 / $1.40 per 1M tokens
What’s not
- No published benchmark or writing samples
- $6 per 1M output tokens, twelve times DeepSeek V4.1 Flash
- Single provider on OpenRouter, so no failover
- Closed weights; the ensemble's design is not documented
Best for
Writers and interactive-fiction developers who find general chat models too agreeable and want to test a model tuned for conflict and long scenes.
Aion 3.5 alternatives
Aion 3.5: common questions
- Is Aion 3.5 worth it?
- Aion 3.5 is a serious attempt at a model that writes stories rather than answers, but at twelve times DeepSeek V4.1 Flash's output price you are paying for a claim nobody has measured yet. We rate it 3.6 out of 5.
- How much does Aion 3.5 cost?
- Aion 3.5: $3 input / $6 output per 1M tokens, cached input $0.75; Aion 3.5 Mini: $0.70 / $1.40. Free tier ($0) with a daily credit allowance.
- What are the downsides of Aion 3.5?
- No published benchmark or writing samples; $6 per 1M output tokens, twelve times DeepSeek V4.1 Flash; single provider on OpenRouter, so no failover; closed weights; the ensemble's design is not documented.
- Who is Aion 3.5 best for?
- Writers and interactive-fiction developers who find general chat models too agreeable and want to test a model tuned for conflict and long scenes.
- How much does Aion 3.5 cost for a team or in production?
- At $3 input and $6 output per 1M tokens, a role-play session that resends a 50K-token history for 100 replies of 500 tokens costs about $15.30 before caching; cached input at $0.75 cuts the history cost by three quarters.
- Is Aion 3.5 safe for company data?
- Aion Labs' privacy policy: request and response content is not stored, metadata kept up to 12 months, no training on your content. Email Agents are the exception and store their content until deleted. Adults only. Not available with OpenRouter's EU in-region routing.
- What are the best Aion 3.5 alternatives?
- The closest alternatives we have reviewed are DeepSeek-V4.1-Flash, GLM-5.3-Flash, Hy4 preview and MiMo-V2.6-Pro.
Answers checked Nov 5, 2026.