MiniMax M3.1 Flash vs DeepSeek V4 Flash: One Name, Two Models

M​iniMax M3.1 Flash and DeepSeek V4 Flash look like the same product sold twice: both are cheap, both advertise a million-token context, both come from Chinese labs that have spent 2026 undercutting each other on price. The first thing to know is that one of those two names no longer points at a fixed model. M​iniMax M3.1 Flash shipped on 19 October 2026 as a closed, always-thinking coding and agent model at $0.15 in and $0.60 out per million tokens. DeepSeek V4 Flash is an MIT-licensed open-weight model from April 2026, refreshed on 31 July — and since 10 September, anyone typing deepseek-v4-flash into D​eepSeek's own API has been talking to DeepSeek V4.1 Flash instead. D​eepSeek's documentation now says the retired names are still accepted but that "the corresponding models have been retired." So there are two honest versions of this comparison, and which one you need depends on whether you call D​eepSeek's endpoint or download the weights. 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.

Both are worth pricing, because both are genuinely cheap and they are cheap in different ways. What follows is the availability trap first, then the spec lines that actually decide the purchase, then the arithmetic — because on this pairing the per-token price and the per-task bill point in opposite directions.

Which DeepSeek V4 Flash you are actually comparing

Three things share the name, and they behave differently at checkout.

• The original open weights. DeepSeek V4 Flash is a 284-billion-parameter mixture-of-experts with 13 billion active per token, published under MIT on Hugging Face in FP4 and FP8. The July refresh is the version third parties still serve, and the version our catalogue lists: $0.22 per million input, $0.66 per million output, $0.007 per million on cache reads, 1,048,576-token context, 384,000 maximum output, text input only.

• The vendor card it used to carry. $0.14 in and $0.28 out per million was D​eepSeek's own rate for V4 Flash. You cannot pay it today, because D​eepSeek retired the SKU. Quoting it against M​iniMax's $0.15/$0.60 is the most common mistake in this comparison, and it overstates D​eepSeek's advantage by roughly two times on output.

• The identifier that still answers. deepseek-v4-flash now routes to DeepSeek-V4.1-Flash-1017 and bills at V4.1 Flash rates — $0.15 in and $0.60 out off-peak, doubling to $0.30 and $1.20 during D​eepSeek's weekday peak windows (01:00–04:00 and 06:00–10:00 UTC), with cache hits at $0.003 and $0.006. That is a different architecture at a different price, wearing the old name as a convenience alias.

The practical consequence is narrower than it sounds and easier to act on than to remember. If you self-host or buy DeepSeek V4 Flash from a third-party serving the open weights, you get the April/July model at a stable flat rate. If you point at D​eepSeek's own URL, you get the successor, whose price doubles for part of the working day. Neither of those is M​iniMax's actual rival on a scoreboard — the successor is, and it is two points ahead. That shows up in the numbers below.

The six lines that decide it

Set the two spec sheets side by side and most rows are a wash. The rows that are not a wash are decisive enough to plan around.

• Price per million — M​iniMax M3.1 Flash $0.15 in, $0.60 out, $0.015 cached. DeepSeek V4 Flash $0.22 in, $0.66 out, $0.007 cached as we serve it; $0.15 / $0.60 through D​eepSeek's own alias, doubling at peak.

• Context window — 1,000,000 tokens against 1,048,576. A 48,576-token difference that will never be the reason a job fails.

• Maximum output — not published by M​iniMax. 384,000 tokens on DeepSeek V4 Flash. This is the line that matters most and it takes a section of its own.

• Inputs — text, image and video in, text out, on M​iniMax M3.1 Flash. Text only on DeepSeek V4 Flash, whose vision work moved to a separate experimental build that D​eepSeek also retired in September.

• Thinking control — an effort ladder from low to max that you cannot switch off, defaulting to max, on M​iniMax M3.1 Flash. Thinking on by default but genuinely disableable on DeepSeek V4 Flash.

• Weights — closed, with M​iniMax promising release "later," against MIT-licensed weights on Hugging Face today.

Two more rows belong here even though they are operational rather than technical. M​iniMax's rate card has no peak windows at all — the price is the price, whatever hour you run — while D​eepSeek's successor doubles on weekday peaks that fall squarely on a European morning. And M​iniMax publishes 200 requests per minute and 10 million tokens per minute shared with MiniMax-M3, against a 2,500 concurrent-request ceiling on D​eepSeek's Flash SKU.

The number one side will not publish

A million tokens of context is a promise about reading. The maximum output is a promise about writing, and it is the one that decides whether a task fits in a single call. Ask M​iniMax M3.1 Flash to read a repository and emit a migration plan, a full test suite, or a 60-file refactor, and it will do it across turns, because the vendor publishes no ceiling — the documentation exposes a max_tokens parameter and warns that thinking tokens count against it, and stops there. Ask DeepSeek V4 Flash the same thing and you have 384,000 tokens of room, which is enough for the whole plan in one response.

That is not a flaw in M​iniMax's model so much as a hole in its documentation, and there is a live cost to the hole: you cannot size a budget, set a retry policy, or predict where a long generation truncates without a number. If M​iniMax publishes one later, this paragraph ages badly in the model's favour. Until then, any workload whose value depends on one very long completion should be routed to the model that states its ceiling.

The only same-harness score, and the ladder around it

Vendor benchmark tables are marketing with decimal places, so the useful question is which two numbers were produced by the same harness on the same version. On this pairing there is exactly one clean pair: DeepSWE v1.1, where M​iniMax reports 73.7% and D​eepSeek reports 74.2% for V4.1 Flash. Half a point apart, both vendor-reported, neither independently reproduced. Every other row a marketer will show you compares different harness versions or different generations and cannot be subtracted.

M​iniMax's launch post adds BankerToolBench 76.2, OSWorld V2 61.7, Agents' Last Exam 29.6 and CyberGym 82%, all vendor-reported, none with a D​eepSeek column beside it. It also claims "top performers" on AutomationBench-AA without publishing the score, which is the sort of claim that should be read as absent until someone else fills it in.

The independent picture is thinner and more useful. Artificial Analysis scores M​iniMax M3.1 Flash at 43 on its Intelligence Index, fourth of 173 models, at a cost of $0.22 per completed index task. The retired DeepSeek V4 Flash 0731 sits at 34 and $0.17. DeepSeek V4.1 Flash 1017 — the model the old name actually reaches — sits at 45 and $0.26. Read as a line rather than as a ranking, that is remarkably orderly: nine index points for five cents a task, then two more points for four cents more. M​iniMax is the middle rung, and the top rung belongs to the vendor it is supposedly competing with.

Speed is where the vendor claim and the measurement part company most sharply. M​iniMax's post promises "200+ TPS in everyday use" with a P50 above 170. Artificial Analysis measures 104.7 output tokens per second with a 1.28-second time to first token, while DeepSeek V4.1 Flash 1017 measures 210.6 tokens per second at 1.07 seconds. Treat those as live figures rather than fixed: the same Artificial Analysis page read 125.5 tokens per second and an index score of 42 on launch day and reads 104.7 and 43 today, which is what a one-day-old entry looks like before it settles.

What a month of agent traffic costs on each

Per-token prices are a poor proxy here, because the two models spend output tokens at very different rates. M​iniMax M3.1 Flash cannot stop thinking: send thinking: disabled and the API returns HTTP 400, and the default effort is the most expensive rung. That means every turn carries reasoning tokens the invoice counts and the user never sees.

Take a coding agent with a 40,000-token stable prefix — system prompt, tool schemas, repository map — running 60 turns, each turn adding 1,200 new input tokens and producing 900 tokens of visible output. Assume M​iniMax's default effort roughly triples its output per turn with reasoning and D​eepSeek runs in non-thinking mode:

• M​iniMax M3.1 Flash, at $0.15 / $0.60 / $0.015 — $0.036 on cached prefix reads, $0.011 on new input, $0.097 on 162,000 output tokens. About $0.14 per session.

• DeepSeek V4 Flash, at our served $0.22 / $0.66 / $0.007 — $0.017 cached, $0.016 new input, $0.036 on 54,000 output tokens. About $0.07 per session.

• DeepSeek V4.1 Flash 1017, at $0.15 / $0.60 / $0.003 off-peak — $0.007 cached, $0.011 input, $0.032 output. About $0.05, and about $0.10 if the run lands in a peak window.

The assumption doing the work is the thinking-token volume, and it is an assumption — M​iniMax publishes no figure for it, and the ratio moves with the effort setting you choose. What is not an assumption is the direction: the model with the equal-or-cheaper sticker produces the larger bill on an agent loop, because it reasons on every turn whether the turn needs it or not. Lower the effort to low or medium and the gap narrows sharply; leave the default in place and it does not.

Caching widens it further. M​iniMax's cache discount is 90% off input, with matching that starts at 512 input tokens and runs tool list, then system prompt, then user messages. D​eepSeek's is 98% off and applied automatically, with the vendor's own caveat that it is best-effort with no fixed TTL. On a long session that difference is worth more than the headline rate.

This is the section where routing earns its keep rather than being a slogan: on one key across 200+ models with provider list prices passed through at 0% markup, both D​eepSeek SKUs are callable today — deepseek/deepseek-v4-flash for the pinned open weights and deepseek/deepseek-v4.1-flash-1017 for the fixed October snapshot — and a vendor price cut lands here the same day it lands there. M​iniMax M3.1 Flash is not on that list, and the honest version of the availability line is that it is reachable only through M​iniMax's own API, its M Plan subscription and its M​iniMax Code product, with weights promised and no date attached.

The switch you cannot turn off

The effort ladder on M​iniMax M3.1 Flash is low, medium, high, xhigh and max, and the default when you omit the field is max. There is no off. This is a design decision with two faces. A routine rename at low and a repository-wide refactor at xhigh are the same model making a sensible trade, which is exactly what a Flash tier should offer. But a caller who never sets the field — which is most callers, on the first integration — pays the most expensive rung on every request, and the effect is invisible until the bill arrives.

Two operational details follow from it. Thinking tokens count against max_tokens, so a cap sized for a short answer can be consumed entirely by reasoning and return an empty message with a max_tokens stop reason. And reasoning_split cannot be set to false, so on the OpenAI-compatible path the thinking always arrives in the reasoning_content field rather than inside the content string. If you have a parser that scrapes thinking tags out of older M​iniMax output, it has nothing to do here.

DeepSeek V4 Flash gives you the other arrangement: thinking on by default, but a real non-thinking mode, and effort selection on the V4.1 successor. For classification, extraction, formatting and short tool calls — the bulk of most agent traffic — that is the difference between paying for reasoning and not.

Who should pick which

Neither of these is the right default for everything, and the split is unusually clean.

Pick M​iniMax M3.1 Flash when the work is agentic and multimodal — screens in the loop, video as input, computer-use or cybersecurity evaluation — and when you will actually set the effort level per call. Its vendor numbers on OSWorld V2 (61.7) and CyberGym (82%) are exactly the kind of capability that DeepSeek V4 Flash's text-only interface cannot express at all, and the flat rate card is a genuine advantage for a team whose load lands in D​eepSeek's peak windows.

Pick DeepSeek V4 Flash when you need open weights you can quantise, self-host, fine-tune and keep after a vendor retires it; when you need a published output ceiling for long single-pass generation; when you need to switch reasoning off; or when your traffic is exactly the throughput-sensitive, non-reasoning kind that the cheaper cache and non-thinking mode are built for. The MIT licence is not a footnote — it is the only line on either spec sheet that a price change cannot take away.

What would change the answer

Three things, in order of how much they would move it. M​iniMax publishing an output ceiling would remove the single biggest functional gap on the sheet. M​iniMax releasing the weights it has promised would delete D​eepSeek's licence advantage outright and turn this into a straight price-and-speed argument. And a second independent measurement of M​iniMax M3.1 Flash at a lower effort setting would settle the question the cost arithmetic currently has to assume — how much of its 43 costs $0.22, and how much of that $0.22 is thinking nobody asked for.

Until then the defensible position is narrower than either launch post would like. On capability, the independent index puts M​iniMax M3.1 Flash nine points above the D​eepSeek model whose name it is usually compared against and two points below the D​eepSeek model that name now reaches — and the two-point gap costs four cents a task. On everything else the decision is not about scores at all. It is about whether you need weights, an output ceiling, and an off switch, or whether you need a flat price and eyes.