AionLabs

This vendor sells roleplay models on a router, and the router is the only place most people meet it. That turns out to be worth checking, because AionLabs runs its own catalogue — an OpenAI-compatible API at https://api.aionlabs.ai/v1 whose GET /v1/models 'endpoint is publicly accessible without authentication'source, accessed 2026-09-06 — and the two listings of the same four models do not say the same things.

Start with the clock. The vendor stamps each model with a release date and the router stamps each row with a creation timestamp, and the four pairings run 66 days for aion-labs/aion-rp-llama-3.1-8b (30 November 2024 to 4 February 2025), 64 for aion-labs/aion-2.0 (21 December 2025 to 23 February 2026), then 63 and 54 for Aion-3.0 and Aion-3.0-Mini, released nine days apart in May 2026 and listed one second apart on 7 July. Nothing on either site explains the delay. Four samples is not many, but a twelve-day band across seventeen months of releases is a tighter spread than "whenever someone filed the paperwork" usually produces.

The router's oldest row carries the more interesting drift. Its listing still leads with a ranking: 'Aion-RP-Llama-3.1-8B ranks the highest in the character evaluation portion of the RPBench-Auto benchmark, a roleplaying-specific variant of Arena-Hard-Auto, where LLMs evaluate each other’s responses. It is a fine-tuned base model rather than an instruct model, designed to produce more natural and varied writing.'source, accessed 2026-09-06 RPBench-Auto is a real benchmark with a public leaderboard, and that leaderboard is 17 rows, headed by Higgs-Llama-3 70B V2 at 68.25% on Character and closed by Character.AI at 12.47%; no Aion model appearssource, accessed 2026-09-06. It has not moved since d1cca01f, 'rename to V1 for consistency', committed 2024-08-06 — the most recent change to the leaderboard filesource, accessed 2026-09-06 — 95 days before this model's weights repository was created, so the model could never have been on it. Models join that leaderboard by opening a pull request, so an absence proves nothing about how the model would score; what it means is that the claim cannot be checked where it points. The live leaderboard linked at the top of the benchmark's own README is gone: it redirected to https://www.boson.ai/rpbench and returned HTTP 404 when requested on 2026-09-06source, accessed 2026-09-06. AionLabs' own catalogue does not repeat the ranking: its entry for the same model reads, in full, 'A variant of LLama 3.1 8B optimized for immersive roleplaying and storytelling.' — the whole of it; no benchmark is namedsource, accessed 2026-09-06. The weights' model card is out of step too, in the other direction — it promises 'This is a completely uncensored model that has been trained to excel at roleplaying and creative writing. It can utilize the full 131K context.'source, accessed 2026-09-06, where the router's row for that model serves 32768tokensopenrouter-models, last checked 2026-09-14 and the vendor's own docs table lists the aion-labs/aion-rp-llama-3.1-8b row at Context 32Ksource, accessed 2026-09-07. Here the card is the outlier and the router matches the vendor.

The newer rows drift the other way, and this one lands on the choice a reader is actually making. The vendor describes Aion-3.0 as 'A multi-model system built on the GLM family of models. Aion 3.0 employs a collaborative generation process in which multiple specialized models contribute to each response. The result is stronger narrative structure, more compelling tension and conflict, and a nuanced treatment of mature and darker themes.'source, accessed 2026-09-06 OpenRouter's paraphrase is 'Aion-3.0 is a multi-model roleplaying and storytelling system from AionLabs, built on the GLM family of models. It uses a collaborative generation process in which multiple specialized models each contribute to a response, producing stronger narrative structure and more compelling tension and conflict.'source, accessed 2026-09-06 The closing clause is gone — from Aion-3.0 and from its Mini, though the equivalent sentence survives intact on the listing for Aion-2.0. For anyone picking between these rows that clause is usually the deciding one, and it is missing from the page where the picking happens.

Two last things the board cannot show. The first is who this is: the terms of service are 'a legally binding agreement between you and Deep Forge sp. z o.o. (NIP: 1231533694), the company operating Aion Labs', and they 'shall be governed by and construed in accordance with the laws of the Republic of Poland'source, accessed 2026-09-06 — a Polish company, where OpenRouter's provider record for aion-labs gives headquarters "IL" and an empty datacenters listsource, accessed 2026-09-06. The plain two-word name is not a linkable alias here for a related reason: an unrelated AION Labs owns that name in another field entirely, selling itself on "the future of pharma" and "drug discovery and development". The second is that models are at most half the business. The front page gives equal billing to alongside the model API, email agents — 'Deploy agents with a dedicated email address', each of which 'Runs on its own isolated computer with full internet access' and can 'edit and return documents in the reply'source, accessed 2026-09-06 — a product line with no row in any catalog, so the model rows are all any board can ever show of the company that sells it.