MiniMax M3.1 Flash vs Aion 3.5 Mini: The Price Gap Shrinks, the Evidence Gap Doesn't

Aion 3.5 Mini is the first AionLabs model in this series where price alone does not settle the argument. MiniMax M3.1 Flash bills $0.15 per million input tokens and $0.60 per million output; Aion 3.5 Mini bills $0.70 and $1.40. That is 4.7x and 2.3x — a real premium, but nothing like the 20x and 10x its full-size sibling, Aion 3.5, asks against the same MiniMax M3.1 Flash. The gap is close enough on cost that the decision moves to the places where the two are not close at all: the output ceiling, the input modality, and whether anyone outside the vendor has ever measured the thing.

Both models are current. Aion 3.5 Mini shipped on 23 September 2026 alongside the flagship; MiniMax M3.1 Flash shipped on 19 October 2026 and put its weights on Hugging Face four days later. The Aion model is closed, text-only, and always reasons before it answers. MiniMax M3.1 Flash is a 428-billion-parameter mixture of experts with 23 billion active per token, a one-million-token window, and image and video input. On paper the comparison is lopsided; the interesting part is why the bill does not always look that way.

What each model actually is

MiniMax M3.1 Flash is MiniMax's sparse MoE, roughly 428B total parameters with 23B activated per token, served with a 1,000,000-token context window. The rate card is $0.15 per million input tokens, $0.60 per million output, and $0.015 per million cached reads. MiniMax's model card lists text, image and video inputs with text output; Artificial Analysis records text and image. Reasoning is always on, with five effort levels — low, medium, high, xhigh and max — and the API rejects a request to disable thinking outright, returning an error rather than silently ignoring it. The weights are downloadable under the MiniMax Community License, which is open weights rather than open source: attribution is required, and commercial use above $20M in annual revenue needs prior written authorisation. The vendor's own benchmark table includes 62.8 on Automation-Bench, 73.7 on DeepSWE v1.1, 61.7 on OSWorld 2.0 and 82.2 on CyberGym. Those are vendor-run numbers and have not been reproduced independently.

Aion 3.5 Mini is AionLabs' cheaper tier, identifier aion-labs/aion-3.5-mini, and it is a different kind of object. AionLabs describes it as a multi-model system built on the GLM family — several models contributing to one response — and does not publish a parameter count, an ensemble size, or which constituents run on any given call. The context window is 262,144 tokens, doubled from Aion 3.0 Mini and identical to the flagship's, with a hard 32,768-token cap on the reply. Input is text only. Reasoning is mandatory: low, high and max effort levels, defaulting to high, with no non-reasoning path. Pricing is unchanged from the previous generation at $0.70 input, $1.40 output and $0.18 cached input. The weights are not available, and the model has no published benchmark results of any kind.

Two things follow from that description before any measurement. The Mini is not a crippled Mini in the way the name suggests — its context window matches the flagship's — and the thing it is missing is output length, not input length. And the cost story is not just the rate card, because both models think before they answer and neither lets you switch that off.

The rate card, and the three multipliers that set the bill

Per-token prices first, because they are the only numbers both vendors publish cleanly:

• Input — MiniMax M3.1 Flash $0.15 per million against Aion 3.5 Mini $0.70. A 4.7x spread.

• Output — $0.60 against $1.40. A 2.3x spread, and the narrower of the two.

• Cached input — $0.015 against $0.18. A 12x spread, and the line that dominates agent bills.

• Context window — 1,000,000 tokens against 262,144. A 3.8x spread.

• Maximum output — 524,288 tokens against 32,768. A 16x spread, and the largest single gap in this matchup.

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

• Independent score — Artificial Analysis Intelligence Index 43 for MiniMax M3.1 Flash; no entry for Aion 3.5 Mini.

The ratio people quote is usually the input ratio, and it is the least useful one. A model that reads ten tokens for every one it writes is priced mostly by its input rate, so a 10M-input, 1M-output job costs $2.10 on MiniMax M3.1 Flash and $8.40 on Aion 3.5 Mini — a 4x difference, close to the headline. A balanced 1M-in, 1M-out job costs $0.75 against $2.10, or 2.8x. The ratio moves with the workload, which is why "4.7x cheaper" is a slogan rather than a budget line.

Cached reads move it much further. An agent that carries a 100,000-token system prompt across 1,000 calls reads 100 million cached tokens; that is $1.50 on MiniMax M3.1 Flash and $18.00 on Aion 3.5 Mini. If your workload is a long, stable prefix with short turns, the cache rate is the number to compare, and the gap there is 12x rather than 4.7x.

Then there is the multiplier neither rate card shows: thinking tokens. Both models always reason, and thinking tokens are generated tokens billed at the output rate on both. If a model spends 2,000 tokens thinking to produce a 500-token answer, you are billed for 2,500 output tokens — five times the visible text. At those numbers the effective output cost per million visible tokens is $3.00 for MiniMax M3.1 Flash and $7.00 for Aion 3.5 Mini: the 2.3x ratio survives, but every absolute number is five times the sticker. And the defaults are not the same setting. MiniMax's API documentation says an omitted effort level defaults to max, the most expensive setting; Aion defaults to high. Out of the box, the model with the cheaper rate card is also the one spending more per answer. The rate card is a floor, not a forecast, and neither vendor publishes a thinking-token budget you could use to build one.

Context is generous. Output is where the Mini stops.

Aion 3.5 Mini's 262,144-token window is the generous part of its spec sheet. It is twice the previous Mini's window, it matches the flagship exactly, and it is enough for a novel, a large repository sweep, or an hour of transcript. Against most models it would not be a limitation at all. Against MiniMax M3.1 Flash it is 3.8x smaller, and for the growing class of workloads that stuff a whole codebase or a document set into one call, that is a real ceiling.

The harder ceiling is the reply. Aion 3.5 Mini stops at 32,768 output tokens; MiniMax M3.1 Flash accepts up to 524,288, with MiniMax recommending 131,072 in practice. That is a 16x difference, and it is the dimension most likely to disqualify the Aion model for a real job rather than merely disadvantage it. A 40,000-token generated artifact — a long migration, a full report, a large refactor emitted in one pass — does not fit in Aion 3.5 Mini at all. You would chunk it, which changes the task, adds orchestration, and usually costs more than the per-token difference you were trying to capture. For chat turns and chapters, 32K is plenty. For generation-heavy pipelines, it is a hard stop.

Neither vendor publishes a long-context surcharge on these two models; the rate cards above are flat across the window. That matters because it means the comparison does not flip at 200,000 tokens — MiniMax M3.1 Flash is cheaper at every length, and the Aion model simply cannot go past its ceiling.

One model has evidence. The other has a positioning statement.

This is the part of the matchup that does not narrow. MiniMax M3.1 Flash has a vendor benchmark table, a third-party score and a measured speed. Aion 3.5 Mini has none of the three. AionLabs publishes no SWE-bench, Terminal-Bench, GPQA or MMLU-Pro result for any 3.5 model, and Artificial Analysis has no entry for Aion 3.5 Mini, Aion 3.5, or either 3.0 generation. There is no independent capability score and therefore no cost-per-completed-task figure, which is the number that actually decides procurement.

What exists instead is gateway telemetry. Trackers that score it from live traffic cluster around 56 tokens per second at roughly 1.5 seconds to first token — half the 116.2 tokens per second Artificial Analysis measures for MiniMax M3.1 Flash, and those are aggregated production numbers rather than a controlled run. Treat the direction as informative and the exact figure as soft.

The MiniMax side is better documented but not pristine. Its benchmark table is vendor-run, several rows use specific agent harnesses, and the numbers are not apples-to-apples across models. Artificial Analysis's independent score of 43 — fifth of 173 models — is the check on it, and 43 is a good-not-frontier result. The more useful independent detail is verbosity: Artificial Analysis notes the model emitted 180 million output tokens on its index against a 100 million median. A model that talks that much makes the output rate, and the effort setting, more important than the input rate everyone quotes.

Absence of evidence is not evidence of weakness, and it matters less for some jobs than others. If the work is prose — roleplay, fiction, character voice — benchmark suites measure it poorly anyway, and AionLabs has positioned this model squarely there. But it means you cannot buy Aion 3.5 Mini on the strength of a number. You can only buy it on the strength of your own evaluation, which you will have to run before you commit.

Where Aion 3.5 Mini is the right buy

It is the cheap seat in an Aion stack. At $0.70 and $1.40 it is 4.3x cheaper than full Aion 3.5 on both input and output, while carrying the same 262,144-token window and the same GLM-family ensemble description. If you already route to Aion 3.5 for hard work, the Mini is the tier to put the easy turns on, with the same API shape and the same reasoning semantics.

Its output cap is not a problem for conversational work. 32,768 tokens is roughly 24,000 words — more than any single chapter, support answer or analysis most applications return. The cap only bites when the product is the long artifact itself.

Text-only is a feature if you never send images. There is no multimodal surcharge to pay for and no image preprocessing to build. That advantage is thinner than it sounds, because the multimodal model is also the cheaper one per token, but it is real for teams that want the simplest possible pipeline.

Style is the unmeasured claim worth testing. A multi-model ensemble tuned for narrative voice can beat a general agent model on prose even when it loses every benchmark. That is a plausible reason for the model to exist and the one claim in its favour that no rate card can settle. It is also exactly the kind of claim that compounds when it is wrong, so test it on your own prompts rather than trusting either vendor's framing.

Who should pick which

• Pick MiniMax M3.1 Flash if the workload is agentic — tool loops, computer use, long-horizon tasks — where the vendor's agent benchmarks are at least a directional signal and the 1M window keeps the transcript intact.

• Pick MiniMax M3.1 Flash if you generate long artifacts. Above 32,768 output tokens there is no decision to make; Aion 3.5 Mini cannot produce the response.

• Pick MiniMax M3.1 Flash if anything in the pipeline sends a screenshot, a scanned page or a video frame. Text-only rules the Aion model out before price is considered.

• Pick MiniMax M3.1 Flash if you need to self-host, pass a licence review, or avoid a single-vendor dependency. The MiniMax Community License has real conditions — attribution, and a $20M revenue trigger — but the weights are downloadable, which is more than Aion offers. Artificial Analysis still labels the model proprietary, which is a fair reading of a community licence that is not OSI-approved.

• Pick Aion 3.5 Mini if the work is prose under 32K, text in and text out, and you are willing to run your own A/B because nobody else has. Budget for the reasoning tokens and set effort to low or high deliberately rather than accepting the default.

• Pick Aion 3.5 Mini if you are already on Aion 3.5 and want one vendor, one API shape and one billing relationship for both tiers. The 4.3x step down to the Mini is the cheapest way to cut spend in that stack without changing anything else.

• Pick neither if the decision has to be defended with an audited score. Only one of these models has one, and even that one's detailed benchmarks are vendor-run.

What routing does and does not solve

OrcaRouter does not route MiniMax M3.1 Flash or Aion 3.5 Mini today, so this is not a piece where one key gets you either model. What it does route is the previous MiniMax generation and the closest price-comparable alternative: MiniMax-M3 and DeepSeek V4.1 Flash, both at provider list price with 0% markup. If the open question is whether a cheaper model can carry your workload, those are the two you can test this afternoon without a new contract.

The routing argument that does apply here is about the decision you are making. A 4.7x price gap is small enough that the right answer may differ per task, and the honest way to handle that is not to pick one model for everything. It is to send the hard turns to the model with evidence and the cheap turns to the one with fit, fail over when a provider degrades, and measure cost per completed task rather than cost per token. That is a routing problem before it is a model problem — and it is the reason the "which is cheaper" question, on its own, keeps producing the wrong answer.

What would change this verdict

Two things, and both are cheap for AionLabs to do. Publish any benchmark result for Aion 3.5 Mini — even a vendor-run one — and the evidence half of this comparison becomes arguable rather than one-sided. Get an Artificial Analysis entry, and the cost-per-task question becomes answerable at all. Until one of those happens, buying the Aion model is a bet on fit, made without a scoreboard.

On the MiniMax side, the number to watch is the promised Pro variant, and the number to distrust is the 200+ tokens-per-second throughput claim in the launch material. Artificial Analysis measures 116.2, which is a good result and half the headline. Vendor throughput figures are usually best-case, single-stream, short-context numbers; the honest planning figure is the measured one.

The verdict for now is unglamorous. For agent work, long outputs, multimodal input or self-hosting, MiniMax M3.1 Flash wins on capability and on price at the same time, which is rare and does not require much deliberation. For prose under 32K tokens, Aion 3.5 Mini is a defensible specialist with a plausible reason to exist and no evidence behind it — buy it because the writing fits, not because the price does, and count the thinking tokens before you commit to the bill.