MiniMax M3.1 Flash vs Gemini 3.5 Flash-Lite: Half the Price, Twice the Score, and Still Not the Obvious Pick

Gemini 3.5 Flash-Lite has been on sale since 21 July 2026. MiniMax M3.1 Flash arrived on 19 October 2026 at $0.15 per million input tokens and $0.60 per million output — half and roughly a quarter of Gemini 3.5 Flash-Lite's $0.30 and $2.50. On Artificial Analysis' Intelligence Index the newer model scores 43 and the Google model scores 22. So MiniMax M3.1 Flash is cheaper on every token line and twice as capable on the independent index. Then you reach the cost-per-task row, and Gemini 3.5 Flash-Lite wins it, $0.19 against $0.22. That inversion is the only honest way into this comparison, because it tells you the decision is not about rate cards at all.

Two rate cards that only look similar

Start with what each vendor actually publishes, because the structures differ as much as the numbers.

• Input — $0.15 per million tokens on MiniMax M3.1 Flash against $0.30 on Gemini 3.5 Flash-Lite.

• Output — $0.60 against $2.50, a four-to-one gap.

• Cached input — $0.015 against $0.03, both a 90% discount on the input rate.

• Cache storage — Google meters it at $1.00 per million cached tokens per hour. MiniMax's launch rate card does not itemise a storage charge.

• Batch and Flex — Google cuts input to $0.15 and output to $1.25 on both tiers. MiniMax's launch pricing is a flat three-line card, with no batch tier in the launch material.

• Priority — Google sells a priority tier at $0.54 input and $4.50 output. MiniMax's launch post announced no equivalent.

• Context and output ceiling — both accept 1 million input tokens. MiniMax allows up to 524,288 output tokens; Gemini 3.5 Flash-Lite stops at 65,536.

• Weights — MiniMax published a 252.4 GB download on 23 October under its community licence. Gemini 3.5 Flash-Lite is proprietary.

The batch tier deserves its own sentence, because it is the one place the two cards nearly touch. A team that can tolerate batch latency gets Gemini 3.5 Flash-Lite input at $0.15 — exactly MiniMax's standard input price — but output at $1.25, still more than double MiniMax's $0.60. Batch narrows the input gap and leaves the output gap intact. If your workload is output-heavy, which agent and coding workloads are, the batch discount does not rescue the comparison.

The capability gap is not subtle, but read the benchmark version

Artificial Analysis runs both models through the same suite, and the results are not close:

• Intelligence Index — 43 for MiniMax M3.1 Flash, ranked fifth of 173 models; 22 for Gemini 3.5 Flash-Lite, ranked 44th.

• Terminal-Bench 4.0 — 26% against 1%.

• AutomationBench-AA — 73% against 25%.

• GDPval-AA — 1,458 against 985.

• SciCode — 53% against 41%.

The Terminal-Bench number is the one that will be quoted out of context, so it needs context. Google's own model card for Gemini 3.5 Flash-Lite cites Terminal-Bench 2.1 at 54.0%. Artificial Analysis runs Terminal-Bench 4.0 and records 1%. Both figures can be correct, because they are different benchmark generations testing different things. What the 1% means is not that Flash-Lite cannot use a terminal; it means it almost never completes the current hard end-to-end terminal tasks without help. What the 54% means is that it did well on an earlier version of the benchmark. Anyone comparing the model to MiniMax M3.1 Flash on "Terminal-Bench" without naming the version is comparing two different tests.

The vendor-versus-independent distinction matters just as much. Google's card also reports SWE-Bench Pro at 54.2%, OSWorld-Verified at 74.0% and GDM-MRCR at 72.2% on 128K contexts — all vendor-reported, all on Google's chosen settings. MiniMax's launch post reports BankerToolBench at 76.2, DeepSWE v1.1 at 73.7% and OSWorld V2 at 61.7, also vendor-reported. The only rows in this article that put both models through the same harness are the Artificial Analysis ones, which is why they carry the argument.

Why the pricier model can still cost less per task

Artificial Analysis measures cost per completed Intelligence Index task at $0.19 for Gemini 3.5 Flash-Lite and $0.22 for MiniMax M3.1 Flash. That looks impossible beside a four-to-one output-price gap until you remember what the metric is: tokens actually consumed, multiplied by the provider's rates. It is not a rate-card comparison. It is a measurement of how much work each model does to finish the same task.

MiniMax M3.1 Flash is a verbose reasoner. Artificial Analysis recorded roughly 180 million output tokens across the index for M3.1 Flash, against a median of about 100 million for the models it tracks. Gemini 3.5 Flash-Lite's $0.19 cost per task, at roughly a third of MiniMax's blended token price, implies it gets through the same index with far fewer tokens. That is an inference from two published numbers rather than a vendor claim, but it points the same way: on tasks where M3.1 Flash's extra deliberation buys a correct answer, the token spend is a good trade. On tasks where the answer was never in doubt, the deliberation is waste, and the cheaper-per-token model becomes the more expensive one per task.

Run the invoice on a workload you might actually have: 20 million input tokens a month at a 60% cache-hit rate, plus 3 million output tokens.

• MiniMax M3.1 Flash — 8 million fresh input at $0.15 is $1.20; 12 million cached at $0.015 is $0.18; 3 million output at $0.60 is $1.80. Total $3.18.

• Gemini 3.5 Flash-Lite — 8 million fresh input at $0.30 is $2.40; 12 million cached at $0.03 is $0.36; 3 million output at $2.50 is $7.50. Total $10.26, before cache storage.

That is a 3.2x difference in MiniMax's favour. Shrink the output to half a million tokens, the shape of a classification or extraction job, and it becomes $1.68 against $4.01 — still 2.4x. Put Gemini on its batch tier with the 3-million-token output and it lands at $5.19, still 1.6x MiniMax. The per-token arithmetic does not flinch.

So the cost-per-task inversion is real but narrow, and it is a statement about token efficiency rather than price. If your workload resembles the Intelligence Index — short, well-specified tasks where a smaller model can finish quickly — Flash-Lite's efficiency can erase the rate-card gap. If your workload resembles the invoice above, it will not. The only reliable forecast is your own token counts on your own prompts, measured on both.

Speed is two measurements pointing opposite ways

The speed story is more interesting than "one is faster," because the two models win different halves of it.

• Time to first token — 1.48 seconds on MiniMax M3.1 Flash against 6.94 seconds on Gemini 3.5 Flash-Lite. The MiniMax model starts answering nearly five times sooner.

• Output throughput — 95.6 tokens per second against 430.2. Once it starts, Flash-Lite streams about four and a half times faster.

Which number matters depends on the shape of the call. In a chat interface, first-token latency is what the user feels; a model that pauses for seven seconds before the first word feels broken regardless of how fast it types afterwards. In a long generation — a 100,000-token report, a big refactor — throughput dominates the wall clock and the slow starter can finish first. In an agent loop, where each step is a short tool call and the loop runs twenty times, time to first token is paid twenty times and throughput barely matters.

Two caveats. Artificial Analysis measured 95.6 tokens per second for M3.1 Flash, while MiniMax's own material claims a median above 170 and peaks above 200 — a reminder that serving path and measurement conditions move this number a lot. And Flash-Lite's 430 tokens per second is genuinely fast, but its 65,536-token output cap means the long-generation case where that speed would pay off most is bounded. MiniMax's 524,288-token cap is eight times larger.

The provider spread is wider than the gap between the models. OrcaRouter's own page for Gemini 3.5 Flash-Lite reports a p50 time to first token of 3.59 seconds and a p95 of 10.00 seconds across the providers it routes to, against Artificial Analysis' 6.94-second figure. That is not a contradiction; it is what a routed model looks like when different backends serve it. It also means the honest way to compare these two on latency is to measure your own route, not to trust a single number from either vendor.

The row that should decide regulated work

Artificial Analysis' AA-Omniscience evaluation gives MiniMax M3.1 Flash a score of 16: 56% of factual questions answered correctly, and a 93% hallucination rate. That last figure does not mean 93% of answers were wrong. It means that of the responses that were not correct, 93% were outright wrong rather than partial answers or refusals — the model usually answers when it should decline.

Gemini 3.5 Flash-Lite's Artificial Analysis page carries no AA-Omniscience score, so the two cannot be compared on abstention. That absence is not a clean bill of health; it is missing data. What can be said is that a model with a lower capability ceiling and a strong bias toward short answers may fail differently — by declining or by producing obviously thin output — but there is no published measurement that establishes it. Do not assume the smaller model is safer just because it is smaller.

For coding and agentic tasks, AA-Omniscience is a poor predictor: the model is not being asked to recall facts, it is being asked to act on files and tool results. For anything that resembles knowledge work — summarising a contract, answering a customer's policy question, extracting a figure from a filing — the 93% figure is the first thing to design around, not the last. Retrieval, citation checks and a human review step are not optional there, and a cheaper token price does not offset the cost of a confident wrong answer.

A decision rule that survives your own workload

• High-volume classification, extraction, routing, translation with short outputs — Gemini 3.5 Flash-Lite. The 22 index is sufficient for narrow tasks, 430 tokens per second is the point, and 65,536 output tokens is ample. Its batch tier at $0.15 input makes the volume case stronger.

• Repo-scale coding, multi-step agents, long documents, long-form output — MiniMax M3.1 Flash. Terminal-Bench 4.0 at 26% against 1% is the relevant gap, 524,288 output tokens removes a hard ceiling, and the input price is half.

• On-premises or weight access — MiniMax M3.1 Flash, and only it. The 252.4 GB download is real, but it is governed by the MiniMax Community Licence: attribution is required on commercial deployments, and prior written permission is required above $20 million in annual revenue. "Open weights" is not "open source," and procurement should read the licence before the parameter count.

• Latency-sensitive chat — test both. MiniMax wins time to first token by a wide margin; Gemini wins streaming throughput. Which one feels faster depends on answer length.

• Unsupervised factual question answering — neither without retrieval and verification. MiniMax M3.1 Flash has a measured abstention problem; Gemini 3.5 Flash-Lite has no published measurement at all.

Where each one is actually callable

There is an asymmetry worth stating plainly. Gemini 3.5 Flash-Lite is in OrcaRouter's catalogue today, so it can be called through one API key at the provider's list price with 0% markup, with automatic failover across providers and the routing DSL available for composing it with other models. MiniMax M3.1 Flash is not in the catalogue; the MiniMax line there stops at MiniMax-M3. Calling M3.1 Flash means MiniMax's own API or one of its serving partners — including Tenstorrent Cloud, which announced the deployment on 22 October.

That matters for evaluation more than for production. If you want to settle the cost-per-task question on your own data, the useful comparison is often MiniMax-M3 against Gemini 3.5 Flash-Lite on one key, because both are routable and metered the same way; the newer model then has to beat that baseline before you take on a second vendor. OrcaRouter passes provider list prices through unchanged, so a MiniMax price change or a Google batch tier shows up on your invoice the same day it changes, not at the end of a negotiated contract.

The pick depends on which number you can afford to be wrong about

If you choose Gemini 3.5 Flash-Lite and you are wrong about capability, the failure is visible: 1% on Terminal-Bench 4.0 and a 25% AutomationBench-AA score do not hide. You will see the task fail and you will move it to another model.

If you choose MiniMax M3.1 Flash and you are wrong about cost, the failure is smaller than the rate card suggests: on the token-metered example above it is two to three times cheaper, and on Artificial Analysis' task metric it is about 16% more expensive. Neither outcome is catastrophic.

If you choose MiniMax M3.1 Flash and you are wrong about abstention, the failure is silent. That is the asymmetry that should drive the decision for knowledge work, and it is the reason this comparison does not end with "the cheaper, smarter model wins." For agentic and coding workloads, where the model is acting on artifacts you can inspect, MiniMax M3.1 Flash is the stronger pick at a lower token price. For high-volume, short-answer processing, Gemini 3.5 Flash-Lite's speed and task efficiency still earn their place. For factual work with no verification layer, the honest answer is that neither model has earned it.