MiniMax M3.1 Flash vs Gemini 3.8 Flash: The Price Gap Doubles on January 1

The most expensive thing about Gemini 3.8 Flash is not its price today. It is the date attached to it. Google shipped Gemini 3.8 Flash on 2 September 2026 at $0.75 per million input tokens and $3.75 per million output, with cached input at $0.075, and has said that rate runs only through 31 December 2026. On 1 January 2027 it doubles to $1.50 and $7.50, with cached input at $0.15. MiniMax M3.1 Flash launched on 19 October 2026 at $0.15 and $0.60, flat, with cached input at $0.015, and its weights went up on Hugging Face on 23 October. Today the input price ratio is five to one and the output ratio is just over six to one. In January they are ten to one and twelve and a half to one, and nothing about either model will have changed except the calendar.

That makes this an unusual head-to-head. On capability the two are close: Artificial Analysis scores MiniMax M3.1 Flash at 43 and Gemini 3.8 Flash at 41 on the same v4.3.2 Intelligence Index. On cost per completed index task they are not close at all — $0.22 against $1.24. The interesting question is not which model is better. It is which of the two differences you can build a budget around, and which one expires.

Two price sheets, two kinds of expiry

Put the structures beside each other before the numbers. MiniMax's card is a flat rate: the same $0.15 and $0.60 at every hour, with a 90% discount on cached input. Gemini's card is a promotional rate with an announced end date and a cache read that costs five times MiniMax's. Neither is a trick; they are different commercial bets. MiniMax is buying adoption with a low list price. Google is buying adoption with a temporary discount on a model it expects to keep selling at the higher rate.

• MiniMax M3.1 Flash — $0.15 input, $0.60 output, $0.015 cached input per million tokens, flat, no time-of-day tier

• Gemini 3.8 Flash today — $0.75 input, $3.75 output, $0.075 cached input per million tokens, through 31 December 2026

• Gemini 3.8 Flash from 1 January 2027 — $1.50 input, $7.50 output, $0.15 cached input per million tokens

• Cache ratio — 5x in MiniMax's favour today, 10x after the change

• Context — 1,000,000 input tokens on both

• Max output — 524,288 tokens on MiniMax against 65,536 on Gemini

• Weights — MiniMax Community Licence and a 252 GB download on MiniMax; closed on Gemini

Run a support-assistant workload through both: 20 million input tokens a month at a 60% cache-hit rate, plus 3 million output tokens. On Gemini 3.8 Flash today that is $6.00 of fresh input, $0.90 of cached input and $11.25 of output — $18.15. From January it is $12.00, $1.80 and $22.50 — $36.30. On MiniMax M3.1 Flash the same month is $1.20, $0.18 and $1.80 — $3.18. The gap is 5.7x today and 11.4x after the change. Gemini's cache storage is billed separately by the hour, so the cached line is a floor rather than the whole cost; the direction of the comparison does not change.

That is also where a routing layer earns its place. OrcaRouter passes provider list prices through at 0% markup, so the January adjustment will appear in your routing costs the day it appears in Google's documentation, with no new contract and no renegotiation. The same pass-through applies to every route on the catalogue, which means the Gemini side of the price comparison above can be re-run on your own traffic rather than on a hypothetical month.

What the independent index says, and what it does not

Artificial Analysis measures both models on the same suite, which makes its numbers the only like-for-like capability comparison in this article. The two-point gap is small enough that it should not decide a procurement on its own, but the cost-per-task gap is large enough to decide one on its own.

• Intelligence Index — 43 for MiniMax M3.1 Flash, 41 for Gemini 3.8 Flash at high effort, both v4.3.2

• Cost per index task — $0.22 against $1.24, a 5.6x difference

• Output speed — 92.5 tokens per second against 146.9

• Time to first token — 1.52 seconds against 24.70 seconds at high effort

• Output tokens on the index run — roughly 180 million for MiniMax, 170 million for Gemini, both well above the median

• Input modalities — text, image and video on MiniMax; text, image, speech and video on Gemini

The index is a capability average, not a workload. Gemini's 41 is measured at high reasoning effort, and a team that runs it at medium will see a different score and a different bill. MiniMax's 43 is measured at its default effort, which is max, because thinking cannot be switched off. Those are not the same operating point, and the two-point gap should be read with that in mind. It is also why the cost-per-task number is the more honest comparison: it already includes each model's default verbosity and reasoning behaviour.

The output ceiling is the real divide

Forget price for a moment. The structural difference between these two models is how much they can write in one call. Gemini 3.8 Flash caps output at 65,536 tokens. MiniMax M3.1 Flash allows 524,288. That is an eight-to-one difference, and it decides entire workload categories before the rate card is consulted.

A 65K ceiling is generous for chat, extraction, summarisation and ordinary code generation. It is not generous for a full-repository refactor, a generated report with tables and appendices, a long transcript rewrite, or an agent that writes a plan and then executes it in one response. Above 65K output tokens, Gemini does not lose on price; it stops being an option unless you build a chunking pipeline around it. MiniMax's 524K ceiling is the capability that makes its long-context window useful rather than decorative.

The input side is more even. Both models accept a million tokens of context, and both accept video. Gemini also accepts audio and PDF documents; MiniMax's model card lists text, image and video. If your pipeline transcribes meeting audio or reads scanned contracts, that is Gemini's territory, and no amount of price arithmetic changes it. If your inputs are screenshots, video frames and code, both models can see them.

Reasoning effort is a latency dial, and the dials are not the same

Both models think before they answer, and both charge for the thinking as output. The controls differ. Gemini exposes low, medium and high effort, with medium as the default and thinking on by default. MiniMax exposes low, medium, high, xhigh and max, with max as the default when the field is omitted, and rejects any attempt to disable thinking with HTTP 400.

That asymmetry shows up in the latency numbers. Artificial Analysis measures MiniMax at a 1.52-second median time to first token and Gemini at 24.70 seconds at high effort. The Gemini figure is not a serving defect; it is what a high-effort reasoning model does before it emits the first token on a benchmark harness. But it is a product decision. A streaming interface that waits 25 seconds before the first token feels broken, and a team that drops Gemini to low effort to fix that is no longer comparing the model the index measured.

MiniMax's mandatory thinking has the same problem in reverse. A classification or formatting job that used to cost sixty output tokens now pays for a reasoning trace before the answer, even at low effort. The per-token price is lower; the token count is higher. That is why the $0.22 cost per completed task is the number to quote, not the $0.60 rate.

The reliability caveat that neither price sheet shows

Cheap and capable is not the same as safe to automate. Artificial Analysis's AA-Omniscience evaluation put MiniMax M3.1 Flash at 16 overall with a 93% hallucination rate: when the model did not know an answer, it answered anyway almost every time rather than abstaining. That is the single most important number in this comparison for anyone building an agent that acts on the output. A model that is five times cheaper per task and rarely says "I don't know" can be far more expensive once a human has to check its work.

Gemini 3.8 Flash has its own failure modes — it is a verbose reasoner, and Google's promotional pricing suggests the company is still buying usage rather than defending a mature price — but the abstention behaviour is the documented asymmetry. The practical response is not to avoid MiniMax. It is to route around the weakness: keep a second model behind it for factual lookups, require citations or tool-grounded answers on high-stakes paths, and measure the rate at which each model declines to answer on your own evaluation set rather than trusting either lab's benchmark table.

Where OrcaRouter fits

OrcaRouter routes Gemini 3.8 Flash today. It does not host MiniMax M3.1 Flash, and this article is not a claim that it does. What the router gives you is the Gemini side of the comparison on the same key as more than 200 other models, at provider list price with 0% markup, so the January price change lands at the same time it lands at Google. It also gives you automatic failover: if the Gemini route is slow or erroring, the request can land on another model without the caller changing anything.

If you are testing MiniMax M3.1 Flash against Gemini, the practical setup is to call MiniMax through its own API or a third-party platform, keep Gemini routable through OrcaRouter, and compare the two on your own prompts and your own monthly token mix. The price gap is real, but the workload that produces it is yours, not the benchmark's.

The January decision, brought forward

The temptation is to wait until December to decide. That is the wrong order. The Gemini price change is already published, the MiniMax price is already flat, and the workload arithmetic can be run today. A text-only pipeline that spends $1,000 a month on Gemini output today will spend roughly $2,000 in January for the same tokens. Moving it to MiniMax now costs a migration and saves the January bill; moving it in January costs both. A pipeline that needs audio input or Google's ecosystem integration should stay, budget for the increase, and stop pretending the current rate is the real one.

There is also a middle path that neither price sheet advertises: run both. Gemini for the requests that need audio, video or a wider document read; MiniMax for the text-heavy majority where the price gap is largest. That split is a routing decision, not a rewrite, if your gateway can address both.

The short version: on a text workload, MiniMax M3.1 Flash is roughly five times cheaper today and roughly ten times cheaper in January, with a similar independent score and a much larger output ceiling. On a multimodal workload that needs audio, Gemini 3.8 Flash is the only one of the two that can do the job, and the January increase is the price of that capability. Neither answer is close. The mistake is choosing on today's rate without reading the date next to it.