A generated hero card titled 'MiniMax M3.1 Flash vs Gemini 3.6 Flash', with three chips reading 'Output ceiling: 524,288 tokens against 65,536', 'DeepSWE v1.1: 73.2% measured against 49% vendor-reported', and 'Price: $0.15/$0.60 against $0.75/$3.75'. A footer reads that Gemini 3.6 Flash prices are Google's promotional standard-tier rates through 31 December 2026, and the DeepSWE figures are vendor-reported for Gemini and independently measured for MiniMax M3.1 Flash.
Guides & Insights

MiniMax M3.1 Flash vs Gemini 3.6 Flash: The Model You Already Run Is Two Generations Old

Author

jinhao.song

Date Published

Latest models · 20View all models →
Benchmarks: Artificial Analysis · updated daily
Back to all posts

Most teams comparing MiniMax M3.1 Flash with Gemini 3.6 Flash are not choosing between two new models. They are holding a model that shipped on 21 July 2026 — two Gemini Flash generations ago, with Gemini 3.7 Flash in August and Gemini 3.8 Flash in September — and wondering whether the replacement should come from Google at all. Google has an answer ready: upgrade to Gemini 3.8 Flash, same price, better scores. The reason to look sideways instead is that Gemini 3.6 Flash and Gemini 3.8 Flash cost exactly the same, so "stay put" is not a price argument, and the interesting question is what MiniMax M3.1 Flash does that neither Gemini does.

That question has a more specific answer than the benchmark tables suggest. It is not the headline coding score, where the two models are not even measured on the same harness. It is the output ceiling and the input modality list, which are the two places where switching changes what your pipeline can do rather than just what it costs.

Three generations, one price

Google's Flash cadence through 2026 has been roughly one release every three weeks: Gemini 3.6 Flash on 21 July 2026, Gemini 3.7 Flash on 13 August, Gemini 3.8 Flash on 2 September. All three sit at the same standard-tier list price, $0.75 per million input tokens and $3.75 per million output, from a promotional schedule that doubles to $1.50 and $7.50 on 1 January 2027. Identical price across three generations is unusual and it changes the decision: there is no cost reason to remain on Gemini 3.6 Flash, only a migration reason.

MiniMax M3.1 Flash entered at $0.15 input and $0.60 output per million tokens on 19 October 2026, with cached input at $0.015. That is a fifth of Gemini 3.6 Flash's input rate and roughly a sixth of its output rate, and unlike Google's promotional schedule it carries no published expiry.

• Price — MiniMax M3.1 Flash $0.15 in / $0.60 out per million vs Gemini 3.6 Flash $0.75 / $3.75 through 31 December 2026, then $1.50 / $7.50

• Cached input — $0.015 vs $0.075, doubling to $0.15 in January

• Max output per response — 524,288 tokens vs 65,536

• Context window — 1,048,576 tokens vs 1,048,576, a genuine tie

• Inputs — text, image and video vs text, image, video, audio and PDF files

• Release — 19 October 2026 vs 21 July 2026

A generated two-column scoreboard for MiniMax M3.1 Flash and Gemini 3.6 Flash across six shared dimensions: input price $0.15 against $0.75, output price $0.60 against $3.75, maximum output 524,288 tokens against 65,536, input modalities text, image and video against text, image, video, audio and PDF, DeepSWE v1.1 73.2% measured against 49% vendor-reported, and reasoning effort low to max against configurable thinking. A footer reads that Gemini prices double on 1 January 2027 and that the two DeepSWE figures come from different harnesses.

The output ceiling is the difference nobody advertises

Gemini 3.6 Flash caps a single response at 65,536 tokens. MiniMax M3.1 Flash caps it at 524,288 — eight times higher, and the gap is not academic for the work both models are sold for. A full-file refactor of a large module, a generated migration with its tests, a multi-thousand-line report assembled in one pass: at 65K you either split the work across calls and stitch the result, or you accept truncation and detect it downstream. At 524K the call completes.

Gemini 3.8 Flash carries the same 65,536-token ceiling, so upgrading within Google's line does not fix this. If long single-pass generation is the actual constraint, the comparison that matters is not 3.6 against 3.8 — it is either Gemini against a model that will emit the whole artifact.

Sixty-five thousand tokens is still a lot of output, and most production calls never approach it. The teams that feel this limit are the ones running code-generation or document-assembly jobs where the natural unit is the whole artifact, and they usually discover the ceiling as a mysterious quality regression rather than a clean error.

Two DeepSWE numbers that are not comparable

The most quoted figure for this matchup is that MiniMax M3.1 Flash beats Gemini 3.6 Flash on software engineering by a wide margin. The wide margin is probably real; the number behind it is not a fair comparison.

Google reported 49% for Gemini 3.6 Flash on DeepSWE v1.1 at its July launch, on its own harness. MiniMax reported 73.7% for MiniMax M3.1 Flash on the same benchmark in October, also on its own harness. A study published on 17 November 2026 then measured MiniMax M3.1 Flash independently at 73.2% and found Gemini 3.8 Flash — not 3.6 — at 61.7% when run by someone other than Google. Gemini 3.6 Flash has not been run through that harness at all.

So the honest reading is directional: the two vendor numbers, 49% and 73.7%, were produced by different labs under different scaffolds and do not subtract cleanly. What the independent study adds is the correction on the Google side — its own newer model came in about ten points under Google's claimed range when measured outside the lab. Applying that correction to the older, weaker model's 49% is guesswork, not arithmetic, and anyone quoting a 24-point gap as measured fact is quoting two marketing pages at each other.

For the practical version: both models are plausible coding agents, MiniMax M3.1 Flash has the only independently reproduced score in the pair, and Gemini 3.6 Flash's real coding behaviour is best judged on your own repository rather than on either vendor's chart.

What MiniMax M3.1 Flash cannot read

Here is where the older Gemini wins outright, and it is the reason a blanket "switch" recommendation is wrong. Gemini 3.6 Flash ingests audio and PDF files alongside text, images and video. MiniMax M3.1 Flash takes text, image and video only.

If your pipeline transcribes a call recording, parses a scanned contract, or feeds a slide deck through as a document, Gemini 3.6 Flash does that natively and MiniMax M3.1 Flash does not — you would be adding a transcription or PDF-extraction step in front of it, with its own cost and its own failure modes. That is not a benchmark gap; it is a capability that is either present or absent.

Video is the mirror case. MiniMax M3.1 Flash reads video natively, which Gemini 3.6 Flash also does, so neither model has an advantage there — but it matters for the migration itself: a team moving screen recordings or UI captures into an agent loses nothing by leaving Google for this specific input type.

Where the two models land in an agent loop

MiniMax M3.1 Flash runs reasoning always on, with selectable effort from low through medium, high, xhigh and max. Gemini 3.6 Flash exposes configurable thinking and a documented reasoning-effort parameter on its API. Both give you a latency dial; neither gives you a model that can turn thinking off entirely, so interactive latency comes from the effort setting rather than from a separate non-reasoning mode.

The measured latency profiles differ in a way worth flagging. Artificial Analysis reports Gemini 3.6 Flash at 182 output tokens per second with 12.9 seconds to the first answer token, and MiniMax M3.1 Flash at 132 tokens per second with 50.6 seconds to the first token against a peer median of 19.2. On a batch job that 50-second think is invisible. In an editor where a human waits for the first line, it is the difference between a tool that feels responsive and one that feels broken, and no amount of throughput fixes it.

Reasoning effort is the lever: MiniMax M3.1 Flash at low effort will not spend fifty seconds thinking, and the coding scores everyone quotes were not produced at low effort. If you are switching for the DeepSWE number, you are buying the high-effort configuration and its latency with it.

One caveat on Gemini 3.6 Flash's status

Google's own model documentation lists Gemini 3.6 Flash as stable, described as a previous-generation Flash model, with no deprecation date and no shutdown notice. Artificial Analysis labels the same model deprecated and points users toward Gemini 3.7 Flash. Both can be true — a model can be stable on the vendor's API and marked superseded by a tracker — but if deprecation risk is part of your reasoning, the vendor's page is the one that carries a date, and it currently carries none.

The practical version of that risk: Gemini 3.6 Flash is two generations behind with no published expiry and an identical price to its successor. Even a team with no interest in MiniMax should not be planning a 2027 roadmap on it.

Running both from one endpoint

The migration question usually stalls on contracts rather than models. Both of these are on the OrcaRouter catalogue — minimax/minimax-m3.1-flash and google/gemini-3.6-flash — behind one key, priced at each provider's list rate with 0% markup added. That last part is what makes a January decision cheap to defer: when Google's promotional rate doubles on 1 January 2027, a pass-through shows the new number on your invoice that day, so the comparison you run in December is still the comparison in February.

A screenshot of the OrcaRouter model page for google/gemini-3.6-flash, showing the Gemini 3.6 Flash heading, a 1,048,576-token context window, a 65,536-token maximum output, text, image, video, audio and file input with text output, the $0.75 input and $3.75 output list price per million tokens, and capability tags for tool calling, structured outputs, reasoning and context caching.

MiniMax M3.1 Flash is also reachable over three API shapes — Chat Completions, OpenAI Responses and Anthropic Messages — which means an Anthropic-style client can point at it without a rewrite. For a team testing whether the coding gains hold on its own repository, that is the difference between a spike this week and a migration project next quarter.

A screenshot of Google's Gemini API pricing page for Gemini 3.6 Flash, showing the standard tier at $0.75 per million input tokens and $3.75 per million output tokens for prompts up to 200k tokens, alongside a batch tier at $0.375 and $1.875, a flex tier at $0.50 and $2.50 and a priority tier at $1.35 and $6.75. A second block covering prompts above 200k tokens lists standard at $1.35 and $6.75 flagged as a 50% or greater price increase, batch at $0.675 and $3.375, flex at $0.90 and $4.50 and priority at $2.70 and $13.50.

Who should pick which

If your pipeline ingests audio or PDFs, Gemini 3.6 Flash is doing work MiniMax M3.1 Flash cannot do without a preprocessing step, and the price gap has to be measured against that step's cost. Stay, or move to Gemini 3.8 Flash at the same price for the better scores.

If you are generating long artifacts in a single pass, the 524,288-token output ceiling is a capability difference that no Gemini Flash model currently closes. That alone justifies a trial.

If you are running long-horizon coding agents against a repository with a test suite you can trust, MiniMax M3.1 Flash has the only independently reproduced score in this pair, at a fraction of the token price, and the retry loop that gets an agent from 73% to 90% on a 113-task suite costs pennies at these rates.

And if you are on Gemini 3.6 Flash purely by inertia, the correct comparison is not really against MiniMax at all — it is against Gemini 3.8 Flash, which Google will hand you at the same price today and double it in January. Whichever side of that decision you land on, the deadline is the same date, and it is six weeks away.

Compared in this article3

Detected from this article · Benchmarks: Artificial Analysis · updated daily