Aion 3.5: What AionLabs' Flagship Ensemble Is, What It Costs, and How to Call It

Aion 3.5 is the flagship model from AionLabs — a commercial, API-only text system published under the identifier aion-labs/aion-3.5 and dated 23 September 2026 on the vendor's own model list. Its smaller sibling, Aion 3.5 Mini, is a different model with a different identifier (aion-labs/aion-3.5-mini), a different rate card and its own page in the AionLabs lineup. The two are not one model offered in two sizes, and everything below that says "the model" means Aion 3.5 specifically unless Aion 3.5 Mini is named.

AionLabs shipped this generation on 23 September 2026, which makes it a few weeks old rather than a launch story — so this page is the reference for the name, not an announcement. It exists because people type the name in three spellings and land nowhere useful: Aion 3.5, the hyphenated aion-3.5, and the vendor-prefixed AionLabs: Aion 3.5 that appears in the provider's own model metadata. All three are the same model, and the identifier to use in code is aion-labs/aion-3.5.

The split with Aion 3.5 Mini, stated plainly

AionLabs runs two tiers at the 3.5 generation, exactly as it did at 3.0. Aion 3.5 is the expensive one; Aion 3.5 Mini is the cheap one. On the vendor's published spec sheet they are identical everywhere except price — same 262,144-token context window, same 32,768-token output ceiling, same "multi-model system" description, built on the GLM-family of models, same mandatory reasoning behaviour. If you are choosing between them on published facts alone, you are choosing on price and nothing else.

The family lineage is the one place they historically diverged, and it is worth knowing even though that split has now closed. AionLabs' 3.0-generation small tier was described by the vendor as built on the DeepSeek-family, while the 3.0 flagship was built on the GLM-family. At 3.5 both tiers are described as GLM-family systems. Whatever the ensemble's internal division of labour was, the two tiers no longer draw on different lineages — and that is a change to the 3.0-to-3.5 delta, not a difference between the current tiers.

Practically: if you are searching for Aion 3.5 Mini's price, its page and its comparisons live elsewhere. This page is the flagship.

Size, architecture and licence

This is the thinnest part of the public record, and the honest version is short.

• Size — AionLabs publishes no parameter count for Aion 3.5, and no active-parameter figure. It is not a disclosed dense or mixture-of-experts model with a published config; it is a commercial endpoint.

• Architecture — the vendor's own description, repeated in its docs and in its model-discovery endpoint, is "a multi-model system" built on the GLM-family of models that "employs a collaborative generation process in which multiple specialized models contribute to each response." AionLabs does not say how many models are in that system, which models they are beyond naming the GLM-family, or whether every member runs on every call or a router selects among them per prompt.

• Licence — none is published. There is no licence text on AionLabs' model documentation, and no open-weights release: the only AionLabs repositories on Hugging Face are the 2024 Aion-RP 1.0 checkpoints built on Llama 3.1 8B, which are a different product line entirely. Aion 3.5 is a proprietary hosted model. You can call it; you cannot download it, quantise it, or self-host it.

That unnamed-ensemble detail is the most consequential unknown on this page, because it is the one with money behind it. A system that runs several constituent models per request costs its operator several times what a single forward pass costs, and a flat per-token rate is a bet that the blend averages out. A system that routes each prompt to one constituent costs roughly what one model costs. From the outside the two are indistinguishable, and nothing AionLabs has published distinguishes them.

Context window and maximum output

• Context window — 262,144 tokens. The vendor's discovery endpoint reports it as context_length: 262144; the docs page rounds the same figure to "256K". Both are the same number written two ways.

• Maximum output — 32,768 tokens. The docs page says "Max output: 32K"; the endpoint reports max_completion_tokens: 32768.

Those two figures are identical on Aion 3.5 and Aion 3.5 Mini, and identical to nothing else you might assume — they doubled from the 3.0 generation, which carried a 131,072-token window at the same rate card. Worth noting for planning: 262,144 tokens of input and 32,768 tokens of output is an unusual ratio, and a long-form narrative workload that wants a large window is still capped at a 32K reply per call.

Modalities

Aion 3.5 is text in, text out. That is not an inference from the model's name: AionLabs' model-discovery endpoint declares the architecture as text->text for Aion 3.5, Aion 3.5 Mini, Aion 3.0 and Aion 2.0 alike. There is no image, audio or video input path in the published metadata, and no vision variant of the model. The docs page itself does not restate modalities — if you are checking, the vendor's /v1/models response is the surface that carries it.

Two behavioural flags come from the same endpoint and matter more than they look. The model is listed as a reasoning model, and the effort levels it exposes are low, high and max with high as the default — there is no "none" or "off" level in the list, unlike the 3.0-generation models, which expose five levels including an off switch. Reasoning is not optional on Aion 3.5. Tool calling is flagged as supported, and the endpoint also reports the model as not moderated, which is consistent with a vendor whose stated target is roleplay, storytelling and mature themes rather than a general-purpose assistant.

Price

Both tiers carry the same rate card their 3.0-generation predecessors carried. Per million tokens, as published by AionLabs today:

• Aion 3.5 — $3.00 input, $0.75 cached input, $6.00 output.

• Aion 3.5 Mini — $0.70 input, $0.18 cached input, $1.40 output.

The cache-read line is the one teams under-read. On Aion 3.5 a cache read is 25% of the input rate, so a workload that re-sends a long fixed prompt — a character bible, a style guide, a world state — is paying $0.75 rather than $3.00 for that portion, and the same 4:1 ratio holds on the Mini. On a long-context narrative workload that is the difference between the rate card and the invoice.

Whether that is expensive depends on what you are comparing it to, and the comparison is genuinely hard here: Aion 3.5's rate is several times what a measured general-purpose small model charges for input. Where a vendor does move a price, a pass-through gateway that bills the provider's list price with nothing added per token reflects the change the same day rather than at your next billing cycle — which is the only reason pricing pages are worth reading twice.

What AionLabs claims — and what anyone has measured

The vendor's claims are architectural and qualitative. AionLabs credits the ensemble with "stronger narrative structure, more compelling tension and conflict, and a nuanced treatment of mature and darker themes," and its 2.0-generation description adds depth on darker material. Those are the claims of a specialist system aimed at interactive fiction and roleplay, and they are stated without numbers.

AionLabs publishes no benchmark table for Aion 3.5 — no coding, reasoning, or knowledge evaluation of any kind. Independent measurement is absent too: as of 12 October 2026, Artificial Analysis's model index does not list Aion 3.5, Aion 3.5 Mini, Aion 3.0 or Aion 2.0, and no AionLabs provider entry appears in its comparison set.

So there is no head-to-head number here to quote, and it is worth being precise about why that matters. Aion 3.5 has not been measured against any other model at matched effort on a matched harness, publicly, by anyone. Its $3.00 input rate therefore has no denominator: you can price a token, but you cannot yet price a finished job, because no standardised job has been run on this model in public. Anyone who tells you otherwise is quoting a vendor claim as a measurement, or a score from a different model.

How to call it

AionLabs runs an OpenAI-compatible API, so there is no SDK to learn and no bespoke request format. The pieces, all from the vendor's own quickstart documentation:

• Base URL — https://api.aionlabs.ai/v1, with chat completions at POST /v1/chat/completions and discovery at GET /v1/models. The discovery endpoint needs no key, which is how the spec figures on this page were read.

• Authentication — an API key created in the AionLabs dashboard, sent as Authorization: Bearer YOUR_API_KEY.

• Model identifier — aion-labs/aion-3.5 for the flagship. The vendor's own printed examples use exactly that string, so it is the one to paste rather than the display name.

• Reasoning — set effort to low, high or max in the request parameters; high is the default. There is no off switch.

• Streaming is offered over the API and the browser workspace, and tool calling is flagged as available in the model metadata, though the vendor's quickstart samples show only a basic non-streaming call.

Access is tiered by lifetime credit top-ups, and the vendor's rate-limits page states the numbers. The Free tier — described as the default on signup — allows 15 requests per minute, 20,000 tokens per minute and a 20,000-token daily cap. Tier 1, unlocked by any top-up, moves to 50 requests per minute, 1 million tokens per minute and no daily token limit. From there the ladder runs to 1,000 requests per minute and 20 million tokens per minute at $1,000 in lifetime top-ups. AionLabs' pricing page lists monthly billing and a business plan with custom rates but publishes no per-token figure beyond the model rate card above.

That free ceiling is worth doing the arithmetic on before you plan a test: 20,000 tokens per day will not carry a 262,144-token context experiment, so a serious evaluation of this model means topping up. The cheapest route to a first honest read is the cache-read line on the flagship or the Mini tier at $0.70 input — and note that the Mini's published specs are otherwise identical, which makes it a legitimate way to sample the architecture before paying flagship rates for it.

What is still unanswerable about Aion 3.5

Three questions would change how a buyer reads this page, and none of them is answerable today.

The first is what the ensemble actually is. Without a constituent count or a routing description, the $3.00 rate cannot be judged as generous or as thin — only as unverified.

The second is any measurement at all, from AionLabs or from anyone independent. Until one exists, Aion 3.5 sits in the small category of production models whose price is public and whose performance is not, and the honest treatment of a model in that category is to say so rather than to rank it.

The third is what happens to the 3.0 generation now that the small tier has moved families. AionLabs has kept 3.0, 3.0 Mini and 3.5 in the catalogue simultaneously, all three priced exactly as before, so nothing has been forced on existing callers — but a lineup with two generations at the same rate card and no published benchmark is a lineup where the newer identifier is a choice you have to reason about rather than a default you can take on faith.

The defensible summary is this: Aion 3.5 is a proprietary GLM-family ensemble published by AionLabs, dated 23 September 2026, with a 262,144-token window, a 32,768-token reply ceiling, text-only I/O, mandatory reasoning and a $3.00 / $6.00 rate card that has not moved from the previous generation. Its public record is unusually complete on the spec sheet and completely empty on the scoreboard, and that is the shape of the decision: you are buying a documented architecture with an undocumented capability, at a price point set for a specialist rather than a commodity model.