MiniMax M3.1 Flash vs Aion 3.5: Twenty Times the Input Price for a Model Nobody Has Scored

Aion 3.5 asks $3.00 for a million input tokens. MiniMax-M3.1-Flash asks $0.15. The model charging twenty times more is the one nobody has scored — not Artificial Analysis, not any other neutral evaluator we can find — and it is also the one that cannot accept an image, cannot be downloaded, and stops generating at 32,768 tokens. Every published measurement in this matchup points the same direction, and the only argument on the other side is one AionLabs has not put a number behind. That is worth saying before either vendor's framing gets a hearing.

The comparison is not close on paper, which makes it more useful than a close one. Aion 3.5, released 23 September 2026 by AionLabs, is a text-only reasoning model with a 262,144-token window and a flat $3.00 / $6.00 rate card. MiniMax-M3.1-Flash, released 19 October 2026 by MiniMax, is a 428-billion-parameter mixture of experts with 23 billion active per token, native image and video input, a one-million-token window, a $0.15 / $0.60 rate card, and weights on Hugging Face since 23 October. One of those is a 2026 frontier model priced like a commodity. The other is a general-purpose reasoning endpoint priced like a frontier model from two years ago.

What each side actually is

Aion 3.5 is AionLabs' flagship, model ID aion-labs/aion-3.5, text in and text out. The context window is 262,144 tokens, doubled from Aion 3.0's 131,072, with a 32,768-token cap on the reply. Reasoning is mandatory rather than optional and defaults to high, with low and max available; there is no non-reasoning path. AionLabs describes the architecture as a multi-model ensemble in the GLM family rather than a single trained network — the most interesting claim in its documentation and the one that is hardest to verify from outside, because no evaluation has been published that separates the ensemble's output from any single member's. Cached input reads at $0.75 per million, a 75% discount on the input rate.

MiniMax-M3.1-Flash is MiniMax's 428B-total, 23B-active sparse MoE, trained natively in low precision and shipped that way: the open checkpoint is 252.4 GB across 48 safetensors shards plus a 2.3 GB DSpark speculative-decoding drafter, rather than a BF16 master copy. It takes text, images and video and returns text, carries a 1,000,000-token input window with up to 524,288 output tokens, and prices flat at $0.15 per million input, $0.60 output and $0.015 on cached reads. Reasoning is always on with five effort levels; there is no way to turn thinking off, and the API rejects a disable request outright.

The dimensions that decide it

• Input price — MiniMax-M3.1-Flash $0.15 per million against Aion 3.5 $3.00. A 20× spread.

• Output price — $0.60 per million against $6.00. A 10× spread.

• Cached input — $0.015 per million against $0.75. A 50× spread, and the line that dominates a reasoning-heavy bill.

• Context window — 1,000,000 tokens against 262,144.

• Maximum output — 524,288 tokens against 32,768.

• Input modalities — text, image and video against text only.

• Independent score — Artificial Analysis Intelligence Index 43, fifth of 174 models, at $0.22 per index task; nothing published for Aion 3.5.

• Output speed — 95.6 tokens per second on the tracker against roughly 52 measured by third-party aggregators for Aion 3.5.

• Open weights — a Hugging Face checkpoint under the MiniMax Community License against a closed, API-only model.

• Reasoning control — five effort levels, always on, against three levels, mandatory.

Both rate cards are flat, so flatness cancels

The tempting defence of Aion 3.5 is predictability: $3.00 per million across the whole 262K window, no long-context cliff, no tier that doubles your per-token cost halfway through a document. That is a genuine virtue, and it is also not a differentiator here, because MiniMax-M3.1-Flash is flat too — $0.15 flat, from 4,000 tokens to a million. Whatever a predictable rate card is worth, MiniMax is offering it at one-twentieth the price and with a window four times larger. Flatness separates these two models from tiered vendors. It does not separate them from each other.

What remains is the actual invoice. A 240,000-token prompt costs about $0.036 on MiniMax-M3.1-Flash. The same prompt on Aion 3.5 costs about $0.72, and the reply costs ten times as much per token on top of that. If your workload is document-scale — long contracts, whole repositories, multi-hour transcripts — the price difference compounds at exactly the point where token counts are largest, which is the opposite of where you want it.

The scored-versus-unscored problem is the whole story

Artificial Analysis scores MiniMax-M3.1-Flash at 43 on its Intelligence Index, which places it fifth of 174 models, at a measured $0.22 to complete one index task. The tracker also notes that it is unusually verbose — it consumed 180 million output tokens across the index, well above the median — and since output bills at four times the input rate, verbosity is already baked into that $0.22. You are looking at a number that has survived the model's worst habit.

Aion 3.5 has no equivalent. AionLabs publishes no benchmark table we can find, no independent lab has run it, and the sibling comparison pages that list it carry the same note: not evaluated. That absence is not proof the model is weak. It is proof that a buyer has nothing to compare against except the rate card, and the rate card is the one piece of evidence that argues against it.

Where Aion 3.5 still has a real case

Three arguments survive contact with the numbers. The first is failure diversity. If a large share of your traffic already runs on one family of models, an ensemble built from a different lineage is an uncorrelated second opinion, and uncorrelated errors are worth paying for on the small fraction of calls where a wrong answer is expensive. You would route maybe two per cent of traffic to it. At two per cent, a 20× input price is a rounding error against the cost of one bad judgement.

The second is that mandatory high reasoning removes a class of silent regression. There is no effort parameter to tune, so there is no low setting quietly degrading a task nobody re-tested after the last prompt change. That is a maintenance argument, not a capability one, but maintenance arguments are real.

The third is Aion 3.5 Mini. At $0.70 per million input, $0.18 cached and $1.40 output for the same 256K window, it is a plausible cheap tier, and any comparison that looks only at the flagship is comparing the wrong Aion model. Even so, the Mini is still roughly 4.7× MiniMax-M3.1-Flash on input, still text only, still unscored, and still capped at 32K of output.

Where MiniMax-M3.1-Flash wins without an argument

Modality is the shortest route to a decision. Aion 3.5 is text only. If the workload involves a screenshot, a scanned page, a diagram, a chart or a video frame, the comparison ends before pricing is mentioned, because one model cannot be given the input at all.

Output length is the second. A 524,288-token ceiling against 32,768 is a difference of sixteen times, and it decides every task whose deliverable is itself long: a full file rewrite, a migration across a large repository, a structured extraction over a substantial document. A 32K cap forces chunking, and chunking forces you to write orchestration code that has nothing to do with the model's quality.

Open weights are the third, with a caveat that matters. MiniMax published the checkpoint on 23 October under the MiniMax Community License. That is open weights, not open source: non-commercial use is free, commercial use requires displaying "Built with MiniMax M3" prominently, a one-time notice to MiniMax if annual revenue is at or below $20 million, and prior written authorisation above that line. It is a far more permissive arrangement than an API-only model, and a far less permissive one than MIT. Read the licence before you plan a deployment around it.

Putting both in one request path

Neither model is on OrcaRouter's catalogue today, and this article is not a claim that either is. Aion 3.5 is reachable through AionLabs' own API and several third-party platforms; MiniMax-M3.1-Flash through MiniMax's own API and the open checkpoint if you have the GPUs for 252 GB. What OrcaRouter routes from the same neighbourhood is MiniMax-M3 and DeepSeek V4.1 Flash, both at provider list price with 0% markup, so a vendor rate change reaches our side the same day it reaches theirs.

That matters for the architecture this matchup actually implies. The sensible shape is not a migration from one model to the other; it is a small, permanent slice of traffic sent to a differently-built model for a second opinion. Two providers, two keys and two billing surfaces for a two-per-cent workload is precisely the overhead a gateway removes, and automatic failover means the expensive leg degrading does not take the pipeline with it. When the models you are actually comparing are on the same key, the A/B is a model string rather than a second integration.

What would settle it

One independent evaluation of Aion 3.5. That is the entire missing piece, and its absence is doing more damage to AionLabs' case than any competitor could. Score it near 43 at $0.22 per task and the premium collapses to a niche second-opinion story. Score it materially above 43 and the calculus changes: you would still not move a pipeline to a $3.00 input rate, but you would have a defensible reason to spend it on the calls that matter, and the ensemble claim would finally have evidence behind it rather than a description.

Until that number exists, the choice is a measured model at one-twentieth the price with four times the context and image input, against an unmeasured model at a premium whose main documented advantages are flat pricing that its rival also offers and mandatory reasoning that its rival enforces by rejecting the alternative. There is only one defensible default in that pair — and the interesting question is not whether to pick it, but which two per cent of your traffic deserves the exception.