Aion Labs API Review 2026: The Roleplay and Storytelling Specialist
Aion Labs is not trying to be the next OpenAI. It is a Polish AI company (Deep Forge sp. z o.o., Warsaw) that found a narrow, profitable wedge: roleplaying and storytelling models, delivered through an OpenAI-compatible API, priced per token, with a second product line of email agents that work over plain SMTP. It is live on OpenRouter with six models, a published rate-limit ladder, and a pricing page that matches its own documentation to the cent. For a category where half the new providers cannot decide what they charge, that consistency is worth something.
We verified every price, limit and endpoint in this review against the Aion Labs documentation, the live OpenRouter catalog, and the Terms of Service on November 3, 2026.
What Aion Labs actually sells
The core product is a multi-model system built for narrative quality. Aion 3.5 and Aion 3.0 are described as collaborative generation systems where multiple specialized models each contribute to every response. The stated goal is stronger narrative structure, more compelling tension and conflict, and a nuanced treatment of mature and darker themes. Aion 3.5 Mini and Aion 3.0 Mini use the same architecture at a lower price point and context window. The result is a catalog that covers two capability tiers (flagship and Mini) across two model generations (3.5 and 3.0), plus two legacy models that remain online for backward compatibility.
The flagship Aion 3.5 was released on September 23, 2026, alongside its Mini sibling. Both use the GLM model family as the base. Aion 3.0 and 3.0 Mini, released in May 2026, use a mix of GLM and DeepSeek families. Aion 2.0, from December 2025, is a DeepSeek V3.2 variant optimized specifically for immersive roleplaying — it is the only model in the catalog that Aion Labs describes as a variant of a single named base model rather than a multi-model system. The oldest model, Aion-RP 1.0 (8B), is a fine-tuned Llama 3.1 8B released in November 2024 that still ranks among the best in the RPBench-Auto character evaluation benchmark, a roleplaying-specific variant of Arena-Hard-Auto where LLMs evaluate each other's responses.
What all six models share is mandatory reasoning. Every response goes through the reasoning pipeline, and on Aion 3.5 you can control how hard it thinks: the supported effort levels are low, high, and max, with high as the default. This means the effective cost per turn depends on which effort level you use. At max effort, a 500-token response on Aion 3.5 generates 500 output tokens at $6.00/M plus whatever reasoning tokens the model produces internally — and those reasoning tokens are billed as output. Budget for that.
The second product is Email Agents. Each agent gets its own dedicated email address, its own isolated computer with full internet access, and a system prompt that defines what it does. You send an email with a task — research, drafting, summarisation, data extraction — and the agent replies with the result. Attachments are supported; the agent can read, edit, and return documents. This is available to businesses on request, not self-serve.
Pricing
Aion Labs publishes a single per-token price list. There are no tiers, no request minimums, no platform fee, and no markup difference between buying direct and buying through OpenRouter. The prices below are the prices.
| Model | Context | Input / 1M | Cached / 1M | Output / 1M | Released |
|---|---|---|---|---|---|
| Aion 3.5 | 256K | $3.00 | $0.75 | $6.00 | 2026-09-23 |
| Aion 3.5 Mini | 256K | $0.70 | $0.18 | $1.40 | 2026-09-23 |
| Aion 3.0 | 128K | $3.00 | $0.75 | $6.00 | 2026-05-05 |
| Aion 3.0 Mini | 128K | $0.70 | $0.18 | $1.40 | 2026-05-14 |
| Aion 2.0 | 128K | $0.80 | $0.20 | $1.60 | 2025-12-21 |
| Aion-RP 1.0 (8B) | 32K | $0.80 | — | $1.60 | 2024-11-30 |
The flagship tier at $3.00/$6.00 per 1M tokens is priced like a mid-range model — more expensive than DeepSeek V4 Pro ($0.90/$1.50 on Relace endpoints) but cheaper than Claude Opus 5.5 ($4/$20). The Mini tier at $0.70/$1.40 is genuinely competitive for 256K context: it undercuts GLM 5.3 Flash ($0.0285/$0.0923 on OpenRouter, though that is a 1M-context model at a very different price point) and is roughly in line with what you would pay for a good open-weight model on a budget provider. The cache discount is 75%, which is substantial if your workload re-reads the same system prompt or world-building document on every turn. For a roleplay application where the world bible stays constant across hundreds of turns, the cache rate is the number that matters: on Aion 3.5 Mini, re-reading 100K tokens of world-building costs $0.018 instead of $0.07.
There is no batch API, no fine-tuning endpoint, and no committed-spend discount. You pay the listed rate for every token, and the rate does not change based on volume. That simplicity is a feature for small teams that do not want to model-tier their way through a pricing page, but it also means there is no lever to pull if your usage grows past a few hundred dollars per month. At that point, the Business tier with monthly invoicing becomes the relevant conversation.
All six models are listed on OpenRouter at identical prices. We verified this directly: the OpenRouter catalog prices match the Aion Labs own documentation to the cent for every model. That is rarer than it should be in this category.
Rate limits and tiers
Aion Labs uses a lifetime top-up tier system. Once you reach a tier, it is never reduced.
| Tier | Requirement | Requests/min | Tokens/min | Daily Token Cap |
|---|---|---|---|---|
| Free | Default on signup | 15 | 20,000 | 20,000 |
| Tier 1 | Any credit top-up | 50 | 1,000,000 | Unlimited |
| Tier 2 | $100 lifetime | 100 | 2,000,000 | Unlimited |
| Tier 3 | $250 lifetime | 200 | 5,000,000 | Unlimited |
| Tier 4 | $500 lifetime | 500 | 10,000,000 | Unlimited |
| Tier 5 | $1,000 lifetime | 1,000 | 20,000,000 | Unlimited |
The free tier is enough to evaluate the models but not to run a production workload. One dollar of credit — the minimum top-up — immediately removes the daily token cap and raises throughput by 50x. For comparison, a $100 lifetime top-up (Tier 2) gets you 100 requests per minute and 2 million tokens per minute, which is sufficient for a single production application serving a few hundred concurrent users on the Mini tier.
The tier structure is worth understanding because it is lifetime-based, not monthly. Once you top up $100, you are at Tier 2 forever, even if your balance drops to zero. This is a retention-friendly design that rewards early adoption — a team that topped up $100 in November 2026 will still have Tier 2 limits in 2027 without spending another dollar. It also means there is no recurring subscription: you top up when you need to, and the tier persists.
For a roleplay application with 500 concurrent users averaging 20 turns per minute at 500 tokens per turn on Aion 3.5 Mini, you would need roughly 5,000 requests per minute and 5 million tokens per minute — solidly Tier 4 territory ($500 lifetime). That is a one-time $500 top-up for a production workload, not a monthly fee. Compare that to a subscription-based provider where the same volume might cost $500/month, and the economics favor Aion Labs for sustained, predictable usage.
Model details
Every Aion model has mandatory reasoning enabled. The Aion 3.5 models support three effort levels (low, high, max) with high as the default. The older models have reasoning enabled without configurable effort. None of the models currently expose tool calling through the API, which means Aion Labs is not competing for the agentic-coding market that Relace, Portkey, and OpenRouter itself are fighting over. This is a deliberate focus on narrative quality over tool use.
The 256K context window on Aion 3.5 and 3.5 Mini is the largest in the catalog, and it matters for the roleplay use case: a long-running interactive fiction or collaborative storytelling session can easily accumulate tens of thousands of tokens of world-building, character notes, and plot threads. At $0.18/M cached input on the Mini, re-reading a 100K-token world bible costs about 1.8 cents per turn.
How Aion Labs compares to alternatives
The most direct comparison for narrative-quality APIs is not a general-purpose model but a fine-tuned open-weight model served through an aggregator. A Llama 3.1 8B fine-tune on OpenRouter might cost $0.10-$0.30/M input and $0.15-$0.50/M output, which is cheaper than Aion 3.5 Mini at $0.70/$1.40. But that comparison breaks down on three axes. First, context: most fine-tuned 8B models are capped at 4K-32K tokens. Aion 3.5 Mini gives you 256K. Second, consistency: a fine-tune is a snapshot; Aion Labs updates its models and documents the release dates. Third, the multi-model system: Aion 3.5 is not a single model but a pipeline of specialized models contributing to each response, which is a different architecture from a single fine-tuned checkpoint.
Against general-purpose providers, the comparison is about price-performance for narrative tasks. Claude Opus 5.5 at $4/$20 per 1M tells excellent stories, but at nearly 3x the input price and 3.3x the output price of Aion 3.5, it is not the economical choice for a high-volume roleplay application. GPT-6.1 Terra at $1/$5 is closer on price but is a generalist — it will not have the narrative-specific tuning that Aion builds into every model. For a workload that is 90% storytelling and 10% everything else, the specialist premium is real but the generalist inefficiency is also real.
The gap in Aion's catalog is tool use. None of the six models support function calling, which means if your application needs the model to look up a database, call an external API, or control a game engine, you will need to build that orchestration yourself outside the Aion API. This is a deliberate trade-off — Aion Labs is selling narrative quality, not agent infrastructure — but it does mean that Aion is a complement to a general-purpose provider, not a replacement for one.
Email Agents
The Email Agents product is what makes Aion Labs structurally different from every other provider in this catalog. Instead of exposing an API endpoint, each agent is an email address. You send a task; the agent replies. The agent runs on an isolated computer with full internet access, can browse the web, run code, and edit documents. Custom agents and agent swarms — multiple agents each with their own email address, collaborating on multi-step workflows — are available as a professional service.
The use case Aion Labs describes is straightforward: you need research on a topic, you send an email to your agent with the question, and it replies with a structured answer. You need a document edited, you attach it, and the agent returns the edited version. You need a multi-step workflow — say, competitive analysis across five websites — you describe it in the system prompt and the agent handles it. Because the interface is email, there is no SDK to install, no webhook to configure, and no infrastructure to maintain. For teams that live in their inbox, this is a genuinely low-friction way to deploy an AI worker.
This is not a product you can self-serve. Email agent access requires contacting Aion Labs directly and having it enabled for your account. The Terms of Service mention Email Agents as a distinct service from the LLM API, priced separately. Aion Labs offers a free use-case audit where they review your workflow and show you how agents could save time before you commit.
The Email Agents product does not affect the API pricing in this review. It is a separate line item, and the per-token rates in the table above apply only to the LLM API.
Getting started
The API is OpenAI-compatible. Point any OpenAI SDK at https://api.aionlabs.ai/v1 and it works.
from openai import OpenAI
client = OpenAI(
base_url="https://api.aionlabs.ai/v1",
api_key="YOUR_AION_API_KEY"
)
completion = client.chat.completions.create(
model="aion-labs/aion-3.5",
messages=[
{"role": "system", "content": "You are a collaborative storytelling assistant."},
{"role": "user", "content": "Continue the story from where we left off."}
]
)
print(completion.choices[0].message)Or with curl:
curl "https://api.aionlabs.ai/v1/chat/completions" -H "Content-Type: application/json" -H "Authorization: Bearer $AION_API_KEY" -d '{
"model": "aion-labs/aion-3.5-mini",
"messages": [
{"role": "system", "content": "You are a helpful assistant."},
{"role": "user", "content": "Write a short story about a lighthouse."}
]
}'JavaScript works the same way:
import OpenAI from 'openai';
const openai = new OpenAI({
apiKey: process.env.AION_API_KEY,
baseURL: 'https://api.aionlabs.ai/v1',
});
const completion = await openai.chat.completions.create({
model: 'aion-labs/aion-3.5',
messages: [
{ role: 'system', content: 'You are a helpful assistant.' },
{ role: 'user', content: 'Write a joke about programming.' }
]
});
console.log(completion.choices[0].message);The GET /v1/models endpoint is public and requires no authentication, so you can always verify the current model list and pricing without an account.
curl https://api.aionlabs.ai/v1/modelsThis returns the full model catalog with per-token pricing, context lengths, and release dates. Because the endpoint is unauthenticated, you can also use it to monitor for new model releases or price changes without burning API credits.
Who is Aion Labs for
Aion Labs is a good fit if you are building interactive fiction, roleplay applications, narrative games, or any product where the quality of the storytelling is the primary differentiator and tool use is not required. The 256K context on the 3.5 models, the 75% cache discount, and the consistent pricing across direct and OpenRouter channels make it a practical choice for production workloads in that niche.
The Mini tier is the sweet spot for most independent developers and small teams. At $0.70/$1.40 per 1M with 256K context and a $0.18 cache rate, you can run a roleplay application with hundreds of concurrent users for under $100/month in API costs. The flagship tier makes sense when you need the collaborative generation quality on long documents — novel-length fiction, detailed campaign settings, or multi-session story arcs where narrative coherence across 200K+ tokens is the hard problem.
It is not the right choice if you need tool calling, function calling, or agentic workflows — the models do not support those. It is also not the right choice if you need self-serve enterprise pricing, SOC 2 compliance documentation, or regional data residency guarantees beyond the standard Terms of Service. The Email Agents product is interesting but gated behind a sales contact, so you cannot evaluate it without a conversation.
The competitive landscape for narrative-quality APIs is thin. Most general-purpose providers can tell a story, but few optimize for it. Aion Labs' closest analog is not OpenAI or Anthropic — it is the collection of fine-tuned open-weight models on OpenRouter that target the same roleplay benchmark. What Aion Labs offers over those is a single API, a single price list, and the 256K context window that most fine-tuned 8B models cannot reach.
For the roleplay and storytelling niche, Aion Labs is a legitimate specialist with transparent pricing and a live product. If that is your market, it belongs on your shortlist alongside the general-purpose providers that happen to also tell stories.
Sources: Aion Labs documentation (introduction, quickstart, models, pricing, rate limits, agents, API reference), Aion Labs Terms of Service (version 2026-10-17), Aion Labs homepage and contact page, OpenRouter model catalog. All verified November 3, 2026.