Worth a look if you build fiction or character-chat apps and want an OpenAI-compatible way to try tuned GLM and DeepSeek variants, including its newest flagship with a 262,144-token context. Skip it if you need published benchmarks, compliance certificates or a named team: none were found on 3 October 2026.
Aion Labs is the brand of Polish company Deep Forge sp. z o.o., which has published roleplay-tuned language models since at least November 2024. It sells them through an OpenAI-compatible API and also runs email agents for legal and tax research. It is not the Israeli pharma venture studio AION Labs.
HQ: Warsaw, Poland
About Aion Labs
Aion Labs is the trading brand of Deep Forge sp. z o.o., a Polish limited-liability company entered in the Warsaw commercial register (KRS 0001036658). The company does not name its founders, executives, founding year or headcount on its site, so none are listed here. It is unrelated to AION Labs, the Israeli pharmaceutical venture studio at aionlabs.com. The earliest public model on record is Aion-RP Llama 3.1 8B, which appears on Hugging Face under the aion-labs organisation from 9 November 2024 with an Apache 2.0 licence tag.
The core product is an OpenAI-compatible LLM API at api.aionlabs.ai/v1, with a browser chat workspace beside it. The model list on 3 October 2026 holds Aion-RP 1.0 (an 8B Llama 3.1 fine-tune), Aion 2.0 (a DeepSeek V3.2 variant, 21 December 2025), Aion 3.0 (GLM-based, 5 May 2026), Aion 3.0 Mini (DeepSeek-based, 14 May 2026) and the two newest, Aion 3.5 and Aion 3.5 Mini, both released on 23 September 2026 with a 262,144-token context window. The base models come from Z.ai (the GLM family) and DeepSeek; Aion Labs adapts them for fiction and states that its 3.x models use several specialised models per response. HokAI lists related models from both makers, such as GLM-5.2 and DeepSeek-V4-Pro-0813, though Aion Labs does not say which release each Aion model starts from. HokAI's Aion Labs directory page lists the Aion models it tracks.
The second product line is email agents. Each agent gets its own address, takes tasks and attachments by email and can run on a schedule. The agents page is written for Polish legal and tax work, with plugins for Polish legislation (the Sejm ELI API), administrative court rulings and Supreme Court rulings. No per-agent price is published; the pricing page covers prepaid credits, a free daily allowance and a custom Business plan.
Usage evidence comes from OpenRouter, where Aion 3.5 listed SillyTavern (357 million tokens), Janitor AI (200 million) and Chub AI (66.7 million) as its top apps on 3 October 2026, with Hermes Agent fifth and the time window unstated. Janitor AI is the one with a HokAI record, and consumer chat products such as Character.ai are the nearest alternative for the same audience, though they are finished products rather than APIs. Against general frontier labs, Aion Labs wins on narrow tuning and loses on transparency: no benchmark scores, model cards or compliance certifications were found for any Aion model on 3 October 2026.
On data, the privacy policy says Aion Labs keeps account, billing and request metadata, does not store the content of prompts and responses, does not train on customer content and sets no analytics or advertising cookies. The terms add that requests are forwarded to upstream AI providers and third-party hosting infrastructure, so retention at those hosts sits outside Aion Labs' own statement. The service is for adults only.
Mission
Intelligence On Demand: models built for creative and production workloads, and agents that act over email.
Products
- Aion LLM API (Developer API (OpenAI-compatible)): https://www.aionlabs.ai/docs/
- Aion Email Agents (Email-based AI agents): https://www.aionlabs.ai/agents/
- Aion-RP Llama 3.1 8B (Open-weights roleplay model): https://huggingface.co/aion-labs
Aion Labs Models on HokAI
Links
Frequently Asked Questions
Has Aion Labs raised outside funding?
No funding round, investor or valuation has been published for Aion Labs as of 3 October 2026, and HokAI does not estimate one. The only financial figure on record is the PLN 10,000 share capital of its operating company, Deep Forge sp. z o.o., which is registered capital and not money raised. Revenue comes from per-token API sales and prepaid credits, according to the pricing page.
What does Aion Labs sell, and what does it cost?
The LLM API bills per token from prepaid credits: Aion 3.5 costs $3.00 per 1M input and $6.00 per 1M output tokens, Aion 3.5 Mini $0.70 and $1.40, and Aion 2.0 and Aion-RP 1.0 $0.80 and $1.60 (vendor model list, 3 October 2026). A free tier gives a daily credit allowance with no card, and a Business plan with monthly invoicing is quoted on request. Email agents have no published price.
How does Aion Labs handle prompts and personal data?
Its privacy policy says prompts and responses are processed but not stored or used for training, while request metadata such as model, token counts and IP address is logged for up to 12 months. Upstream model providers and hosting partners also see the prompts, some outside the EEA and mainly in the United States, and that statement does not cover them. The site names no security audits or certifications. Email Agent messages are kept until the owner deletes the agent.
Which companies compete with Aion Labs?
Z.ai competes directly because Aion 3.0 and 3.5 are built on its GLM family: Z.ai wins on published context and open weights, Aion Labs on roleplay tuning. DeepSeek, whose V3.2 underlies Aion 2.0, wins on price and open weights. For end users rather than developers, Character.ai and Janitor AI compete for the same audience as finished products.
How do you start using Aion Labs?
Create an account at aionlabs.ai (adults only), then call https://api.aionlabs.ai/v1 with any OpenAI-compatible client or use the browser chat workspace. The free tier needs no card and allows 15 requests per minute and 20,000 tokens per day; any paid top-up moves you to tier 1 with 50 requests per minute and no daily token cap. The model list at /v1/models needs no key.
Top Alternatives
- Z.ai: Pick Z.ai if you want the GLM models directly, such as GLM-5.2 with a 1M-token context; pick Aion Labs if you want roleplay tuning on GLM behind one OpenAI-compatible endpoint.
- DeepSeek: Pick DeepSeek if you want general open-weight models at low per-token rates; pick Aion Labs if you want fiction-tuned variants of DeepSeek and GLM bases.
- Mistral AI: Pick Mistral AI if you want open weights you can self-host, as with Mistral Large 3; Aion Labs publishes no weights for its newer models that we found.