MiniMax-M3.1-Flash vs MiniMax-M3.1-Flash-Preview: Same Model, Different Contract

If you have been calling MiniMax-M3.1-Flash-Preview since late September, the question the 19 October launch raises is not whether the new model is better. It is whether the thing you were already testing and the thing M​iniMax now sells you are the same artefact — and if they are, why one of them had no price, no benchmark sheet and no independent score for three weeks while the other arrived with all three. Short answer: MiniMax-M3.1-Flash-Preview was the subscription-only, unpublished version of an unnamed model that had been running in public for weeks under the codename Space Bunny. MiniMax-M3.1-Flash is that model turned into a product, with a rate card of $0.15 per million input tokens and $0.60 per million output, a published benchmark table, and an Artificial Analysis Intelligence Index score of 42. The documented AI failures and agent incidents behind releases like this one are collected in the OrcaRouter AI Incident Archive where each entry carries its own sources, severity, confidence grade and disputed flag.

Everything that changed between the two is commercial or documentary. Very little of it is behavioural. That distinction is the whole point of this page, because a team that migrated to the preview on behaviour is now running a model whose model id may not exist in three months, and a team that avoided it because of the missing price has a reason to look again.

Three dates, and what each one actually did

The two model ids sit at the end of a short chronology, and reading it in order is the only way to tell a rename from a release.

• 23 September 2026 — an anonymous model appears in a third-party catalogue under the tag Space Bunny, free to call, no vendor documentation anywhere.

• 27 September 2026 — MiniMax-M3.1-Flash-Preview goes live inside M​iniMax Code and on the Token Plan. Subscription key only. No per-token price, no benchmark table, no weights, no model card.

• 19 October 2026 — MiniMax-M3.1-Flash launches on the pay-as-you-go API alongside M Plan and M​iniMax Code, with published pricing, six benchmark scores, and weights promised "later". Artificial Analysis publishes 42 the same day.

What that chronology rules out is the reading that these are two different models. The preview was not a smaller or earlier checkpoint; it was the same model with the commercial pages switched off. M​iniMax's own launch post refers to the model having processed more than 200 trillion tokens during testing, and the preview period is where most of that traffic came from. What the chronology also rules out is the reverse reading — that the GA is just a label change. Three things became true on 19 October that were not true on 27 September, and they are the ones worth auditing.

The three things that are genuinely different

A price exists. The preview had none — not on the pricing page, not on the model page, not in the API reference. The only way to buy it was a Token Plan seat at $22, $55 or $132 a month, and the only way to estimate its cost was to divide your monthly fee by whatever you managed to run. MiniMax-M3.1-Flash is metered: $0.15 per million input tokens, $0.60 per million output, $0.015 per million cached reads. Against MiniMax-M3's $0.30/$1.20 that is a straight halving on the standard tier, with no large-context surcharge — M3 doubles its rate above 512,000 input tokens, and M3.1-Flash does not.

An independent score exists. The preview had no entry on Artificial Analysis for the three weeks it was live, which is unusual for a model in public hands. The GA scored 42 on launch day, ranked fourth of 173, at $0.22 per Intelligence Index task and 125.5 output tokens per second. That last figure is the one to keep: M​iniMax's own launch post claims "200+ TPS" everyday output and "170+ TPS" at P50, and the independent measurement lands 40% below the headline claim. Neither number is wrong — the vendor figure is presumably a best-case serving path — but only one of them was produced by someone who does not sell the model.

A benchmark table exists, and it is vendor-reported. Six evaluations, none independently reproduced. BankerToolBench 76.2, DeepSWE v1.1 at 73.7%, OSWorld V2 at 61.7, Agents' Last Exam at 29.6, CyberGym at 82%, and AutomationBench-AA described only as a top performer with no number attached. The preview shipped with none of this, so the table is new information even though its provenance is the same company that wrote the press release.

What did not change, and is easy to assume did

Behavioural parity is an inference, not a published guarantee. M​iniMax has not said the preview checkpoint and the GA checkpoint are bit-identical, and it has not published a migration note. What it has published is a parameter table that reads the same on both ids, and the identical restrictions. If you are deciding whether your preview evaluation carries over, these are the things you can bank on:

• Context window — 1,000,000 tokens on both, unchanged

• Thinking — always on. thinking: {"type": "disabled"} returns HTTP 400 on the GA exactly as it did on the preview, and reasoning_effort: "none" is rejected the same way

• Effort ladder — low, medium, high, xhigh, max, defaulting to max when the field is omitted, on both

• Reasoning output — returned on reasoning_content or as a thinking block depending on protocol; reasoning_split: false is not accepted on either

• Admission tiers — service_tier: "priority" is unsupported on both, so the preview gave you no fast lane and neither does the GA

• Sampling defaults — temperature 1, top_p 0.95, on both

• Open weights — none on either. M​iniMax says weights for the GA will be released "later", with no date and no licence named

• Parameter count — undisclosed on both. M​iniMax describes the model as using "more aggressive sparsity" without saying what is being sparsified

Two of those deserve more than a bullet. The effort default is max, and thinking tokens are billed as output and count against your max_tokens. A workload ported from the preview keeps that default, which is fine if it was already tuned there and expensive if it was not. And the loss of service_tier is the same loss it was in September — if your architecture assumes a priority admission lane for latency-sensitive calls, neither id gives you one, and the fallback is low effort rather than a faster queue.

The migration question nobody has answered yet

M​iniMax has not published a retirement date for MiniMax-M3.1-Flash-Preview. The id still resolves in the vendor's own tooling pages, and the documentation that describes it has not been taken down. But the preview's own documentation said the model "can change without notice", and a preview id that survived the GA launch with no stated deprecation schedule is a worse position than a deprecated one, because you cannot plan around a date that does not exist.

The practical read: treat the preview id as a testing surface and move production traffic to the GA. The move costs nothing behaviourally — same context, same thinking constraints, same ladder — and it buys you a published rate, a published benchmark sheet and an independent score you can point at in a procurement review. The one thing it does not buy is any warning about the preview's removal, so if you cannot complete the move this week, log the responded model id rather than only the requested one, so a silent substitution shows up in your monitoring rather than in your users' output.

For teams running several models, that monitoring problem is a routing problem. OrcaRouter puts a fleet behind one API key at provider list prices with 0% markup, so a model id change or a rate card revision is a routing-table edit rather than a code change — and automatic failover catches the case where an id stops answering at three in the morning. MiniMax-M3.1-Flash is not on our catalogue today; the MiniMax models we do route are MiniMax-M3 and the M2.7 family, and neither of those is a substitute for this one, so the honest advice for the GA is the vendor's own API until we carry it.

Which id to use, by situation

There is one situation where the preview id is still the right call: you hold a Token Plan seat, your usage is steady and low-volume, and the subscription is cheaper than metered tokens would be. At $0.15/$0.60 and a demonstrated 180 million output tokens consumed across a 1,600-task index run, heavy agentic usage meters out fast, and a $132 seat can beat it. For everyone else — bursty traffic, production SLAs, anything you have to cost-model — the GA id is the one to build on, because it is the only one of the two with a number attached that survives a finance review.

The thing to remember is that this was never a model comparison. It was a product being finished in public, with the pricing page arriving three weeks after the API did. The model that was called MiniMax-M3.1-Flash-Preview in September and the model called MiniMax-M3.1-Flash in October are close enough that your evaluation results should carry over. The contract around them is what changed, and that contract is what you are now signing.