MiniMax-M3.1-Flash vs MiniMax-M3: Half the Price, None of the Weights

MiniMax-M3.1-Flash costs exactly half what MiniMax-M3 does per token, scores thirteen points higher on the Artificial Analysis Intelligence Index, and is the model M​iniMax is now routing its own legacy subscribers into. It is also the first model in this line you cannot download. MiniMax-M3, released 1 June 2026, is a 428-billion-parameter mixture-of-experts with 23 billion active parameters, published weights on Hugging Face, and a rate card you can read from a table; MiniMax-M3.1-Flash, released 19 October 2026, is a closed model with a better benchmark score, a faster serving path, and a promise that the weights will arrive "later." That sentence is the entire decision, and the price gap is a distraction from it — though a large enough distraction that it is worth doing the arithmetic before you decide. 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.

What the two rate cards actually say

Start with the money, because the money is where the marketing claim and the invoice diverge.

• Standard input — MiniMax-M3.1-Flash $0.15 per 1M vs MiniMax-M3 $0.30 per 1M up to 512,000 input tokens

• Standard output — $0.60 per 1M vs $1.20 per 1M, same 512,000-token threshold

• Above 512,000 input tokens — MiniMax-M3 doubles to $0.60 in and $2.40 out; MiniMax-M3.1-Flash publishes no long-context surcharge

• Cached reads — $0.015 per 1M vs $0.06 per 1M

• Priority admission — not offered on MiniMax-M3.1-Flash; MiniMax-M3 sells it at 1.5× the standard rate, so $0.45/$1.80 below the threshold and $0.90/$3.60 above it

• Intelligence Index — 42 vs 29, per Artificial Analysis

• Cost per Index task — $0.22 vs $0.51, a 57% cut rather than the 50% the rate card implies

• Context window — 1,000,000 tokens on both, though our catalogue lists MiniMax-M3 at 1,048,576

• Output speed — 125.5 tokens per second vs 106.4, both independently measured

• Weights — promised later vs published, open, on Hugging Face

• Parameters — undisclosed, described only as "more aggressive sparsity" vs 428B total and 23B active

Two lines in that list carry more weight than their length suggests. The first is the long-context surcharge. If your workload routinely pushes past 512,000 input tokens — a repository-wide coding pass, a long transcript, a document set — the effective gap is not 2× but 4× on input and 4× on output, because MiniMax-M3.1-Flash does not tier its price at all. The second is priority admission. MiniMax-M3 can buy a faster-serving lane for 1.5×; MiniMax-M3.1-Flash cannot buy one at any price. Its API reference lists service_tier as unsupported. If latency under load was the reason you paid for M3, the newer model does not sell you that lever — it sells you a faster underlying serving path and then takes the lever away.

Reaching half the cost is easier than reaching half the bill

Halving the rate card does not halve the invoice, and this is the single most common error in migration planning for this pair. MiniMax-M3.1-Flash thinks on every request at a default effort of max, cannot be told to stop, and is demonstrably verbose: it consumed 180 million output tokens across the Intelligence Index where the median model consumed 99 million. MiniMax-M3 lets you turn thinking off entirely and defaults to it being off on the Anthropic-compatible endpoint.

Put a number on it. Take a document-tagging job that sends 2,000 input tokens of schema and instructions per call and returns 80 tokens of JSON. On MiniMax-M3 with thinking disabled you pay 2,000 × $0.30 per million plus 80 × $1.20 per million — about $0.0007 per call. On MiniMax-M3.1-Flash the same job, with no way to disable thinking, spends the input at $0.15 per million but then emits a reasoning block that runs to several hundred tokens before the answer. At 500 thinking tokens plus 80 answer tokens, the output bill is 580 × $0.60 per million — roughly $0.00065. Barely cheaper, and that is with a conservative thinking estimate. At 1,500 thinking tokens the newer model costs more than the older one for identical output, because the older one's ability to switch reasoning off is worth more than the newer one's cheaper rate.

Now change the workload. A multi-step agent run that already used 40,000 output tokens on M3 for tool calls and reasoning, with thinking enabled because it needed to think. M3 bills 40,000 × $1.20 per million, or $0.048. M3.1-Flash at the same volume bills $0.024 — and in practice spends less, because the trajectories are shorter, which is why the independent cost-per-task figure falls 57% rather than 50%. That is where the savings live: long reasoning-heavy work, not short deterministic calls.

The rule that falls out of this: if you can describe your workload as "many small answers," MiniMax-M3.1-Flash may be the more expensive model. If you can describe it as "few long answers," it is unambiguously the cheaper one.

The weights are the part you cannot route around

Every number above is reversible. M​iniMax could cut MiniMax-M3's price tomorrow; it could raise M3.1-Flash's. What is not reversible is the licence. MiniMax-M3 is downloadable — 428 billion parameters total, 23 billion active, on Hugging Face and mirrored in an MXFP8 build — which means you can quantise it, fine-tune it, run it on your own hardware, keep serving it after the vendor moves on, and audit what is inside it. MiniMax-M3.1-Flash is a hosted endpoint that M​iniMax says will get weights later, with no date, no licence, and no parameter count.

That asymmetry shows up in three places a cost comparison will not capture. Compliance: a regulated team that needs to run inference inside its own perimeter, or document the model artefact it depends on, cannot use MiniMax-M3.1-Flash today at any price. Continuity: M​iniMax's own retirement schedule has already folded two entire model families — M2.5 and M2.7 — into MiniMax-M3.1-Flash on 19 October, with full retirement on 18 November. Models do get switched off, and the only guaranteed way to survive that is to own a copy. Reproducibility: a benchmark score you cannot re-run against a frozen checkpoint is a score you have to take on trust, and every MiniMax-M3.1-Flash number in this article is of that kind.

If M​iniMax ships the weights, this section is obsolete and the memory of a 428B downloadable model at $0.30 per million becomes the more interesting footnote. Until it does, "later" is a promise, and the older model is the one you can hold.

The capability gap is real, and narrower than the index implies

Thirteen points on the Intelligence Index is a large jump — one model generation at this price point, roughly. But part of that gap measures the change in how the models are allowed to run rather than what they know. MiniMax-M3.1-Flash is scored with reasoning forced on at maximum effort; MiniMax-M3's index result is produced under its own defaults, which on one protocol means thinking off. Some of those thirteen points are the ladder, not the weights.

The vendor-side evidence points the same way. M​iniMax's own launch numbers for MiniMax-M3.1-Flash — BankerToolBench 76.2, DeepSWE v1.1 at 73.7%, OSWorld V2 at 61.7, Agents' Last Exam at 29.6, CyberGym at 82% — are all agentic evaluations, every one of them a setting where forcing maximum reasoning helps. There is no published comparison to MiniMax-M3 on the same harnesses, so the honest statement is that the newer model is better on agentic work by a margin nobody outside M​iniMax has measured against its predecessor, and that the index suggests a genuine improvement of something less than thirteen points once the reasoning policy is held constant.

Where the difference is not in dispute is operational. 125.5 tokens per second against 106.4 is an 18% faster serving path, measured by a third party, and it holds regardless of workload. For an interactive coding tool, that is the difference a user feels.

Where MiniMax-M3.1-Flash is the wrong answer

Three workloads should stay on MiniMax-M3 even after this launch, and it is worth naming them because the launch coverage does not.

The first is anything inside your own network. There is no version of the newer model you can run, and no indication of when there will be.

The second is high-volume, low-token, deterministic work — classification, extraction, routing, formatting — where the ability to disable thinking is the deciding cost feature and MiniMax-M3 has it.

The third is anything that depends on priority admission. If your application bought MiniMax-M3's 1.5× tier to keep p99 latency inside an SLA, MiniMax-M3.1-Flash removes that option; you would be trading a controllable lever for a faster baseline you cannot tune.

Everyone else — agentic pipelines, computer-use loops, long-context coding sessions, teams whose workloads are long rather than many — should be moving. The newer model is faster, cheaper per completed task, and better on exactly the evaluations that describe that work.

Running both without running two integrations

The realistic answer for most teams is not a switch but a split: MiniMax-M3 for the short deterministic calls and anything that has to stay on-premises, MiniMax-M3.1-Flash for the long agentic runs. Both are the same vendor, the same protocol family, and in the case of the Anthropic-compatible endpoint the same request shape — so the split is a routing rule rather than two SDK integrations.

That is what OrcaRouter is for: one API key across 200+ models with provider list prices passed through at 0% markup, so MiniMax-M3's $0.30/$1.20 is what you pay and a vendor revision lands on your invoice the day it lands on theirs. MiniMax-M3 is on our catalogue today. MiniMax-M3.1-Flash is not — call it through M​iniMax's own API, M Plan or M​iniMax Code — so the split above runs as a routing rule with one leg outside us until we carry it. Automatic failover still does real work in that shape: if the newer endpoint is rate-limited or an id changes under you, the request falls back to MiniMax-M3 rather than to an error page.

What makes the split worth building rather than theorising about is that the two models genuinely differ in shape, not just in score. A router that sends short calls to the model that can stop thinking, and long calls to the model that cannot, is doing something the benchmarks already justify.

The short version

MiniMax-M3.1-Flash is the better model and the worse asset. It scores 42 against 29, runs 18% faster, and costs less per completed task — 57% less on an index built from long, reasoning-heavy work. MiniMax-M3 is downloadable, cheaper for short calls, and something you own. For the next few months, the sensible posture is to run both, put the long work on the new one, and watch for the weights release, because the day M​iniMax publishes the checkpoint the only argument left for the old model is the priority lane.