MiniMax M3.1 Flash charges DeepSeek Flash's off-peak $0.15 and $0.60 at every hour. Artificial Analysis prices DeepSeek at peak, and off-peak DeepSeek scores two points higher for $0.13 a task to MiniMax's $0.22.

TL;DR

• Same $0.15 and $0.60 as DeepSeek V4.1 Flash off-peak, with no peak doubling, but cache reads cost $0.015 against $0.003.
• On a cached agent month the two land $1.27 apart if traffic is spread over the week. US daytime traffic favours DeepSeek; Beijing and European mornings favour MiniMax.
• AA's DeepSeek cost per task uses the peak card. Off-peak it is about $0.13, against M3.1 Flash's $0.22 for a score two points lower.

Four cards that start at $0.15

Per million tokens, from each provider's own pricing page, read October 22, 2026. DeepSeek's peak runs 01:00-04:00 and 06:00-10:00 UTC, Monday to Friday, except Chinese public holidays. Qwen's cache read is $0.015 if you create the cache explicitly, at 125% of input.

The model ID is MiniMax-M3.1-Flash. It has a 1M context window and one price at every length, so there's no 512K cliff like the one M3 launched with. MiniMax hasn't published a cache-write rate, a batch rate or a Priority row for it. M3 does have a Priority tier, at 1.5x.

One thing changed on the same page. M3's half price, sold as a seven-day launch promo in June, is now labelled "Permanent 50% off". So M3 sells for $0.30 and $1.20, and M3.1 Flash costs half of that on input and output and a quarter on cache reads.

On a cached agent, the clock decides

Matching DeepSeek's input and output prices doesn't make them the same bill, because agent traffic is mostly cache reads, and that's where MiniMax charges five times as much. We priced a month of agent traffic: 1 billion input tokens, 85% of them cache hits, and 50 million output tokens.

• DeepSeek V4.1 Flash, off-peak only$55.05
• MiniMax M3.1 Flash, any hour$65.25
• DeepSeek V4.1 Flash, spread evenly$66.52
• Qwen3.8 Flash$71.50
• GLM-5.3-Flash$73.00
• DeepSeek V4.1 Flash, peak only$110.10
• MiniMax M3$156.00

MiniMax M3.1 Flash: 150M fresh input at $0.15 ($22.50), 850M cached at $0.015 ($12.75), 50M output at $0.60 ($30.00). Qwen assumes implicit caching, billed at 20% of input. The even spread puts 35 of 168 weekly hours at peak.

DeepSeek off-peak is $10.20 cheaper, and all of that is the cache line. At peak it costs $44.85 more. MiniMax wins on this mix once more than 18.5% of your traffic lands in peak hours. Traffic spread evenly across the week puts 20.8% there, so the two come out $1.27 apart, which is close enough to call a tie.

Where your users are matters more than the average. A US team working 9 to 5 never touches DeepSeek's peak, since both windows fall in the Pacific evening and overnight. A team in Beijing hits peak from 9 to 12 and 2 to 6, which is most of the working day. Europe gets the morning, roughly 8 to noon in Berlin until the clocks change on Sunday. For those last two, M3.1 Flash is the cheaper of the pair.

With no caching at all, the comparison is simple. The cards are identical off-peak and MiniMax is half price at peak, so it can never cost more. Qwen3.8 Flash is $0.13 cheaper on output but charges twice MiniMax's rate on implicit cache hits, so on this mix it lands $6.25 above. It has no Artificial Analysis score yet either.

Artificial Analysis bills DeepSeek at peak

Rate cards don't tell you how many tokens a model burns to finish a task. Artificial Analysis does, as a cost per Intelligence Index task:

Artificial Analysis model pages, read October 22, 2026. Output tokens are the total used to run the index.

Look at the price AA uses for DeepSeek: $0.30 in and $1.20 out. That's the peak card. Every DeepSeek line is exactly double at peak, so the off-peak cost per task works out to about $0.13, and a run spread evenly over the week to about $0.16. Off-peak, DeepSeek scores two points higher than M3.1 Flash for 59% of its cost per task. Even AA's peak figure is only four cents more than MiniMax's.

Claude Haiku 5.5 is the closer match. It ties M3.1 Flash at 43, costs a cent less per task and runs two and a half times as fast. It writes 2.4x as many tokens to get there, but at $0.50 per million output it can afford to. As we noted when it launched, Haiku's price jumps 5x on prompts over 100,000 tokens. A long-context agent can lose that cent quickly, and MiniMax has no such step.

Against its own predecessor, M3.1 Flash is a big step: 14 points above M3 for 43% of the cost per task. If you're on M3 today, switching is the easy part of this decision.

What the launch post leaves out

Speed first. MiniMax says 200+ tokens a second in everyday use and a median above 170. AA measured 92.8, which ranks 87th of 152 and is less than half of DeepSeek's 203.4. On a coding agent that waits on every turn, that gap shows up as minutes.

Second, accuracy on facts. VentureBeat reports a 93% hallucination rate on AA-Omniscience, meaning that when the model doesn't answer correctly, it almost always guesses rather than declining. AA's data puts DeepSeek V4.1 Flash 1017 at about 67%. For retrieval-backed or tool-heavy work this matters less. For anything that answers from memory, it matters a lot.

Third, the benchmarks MiniMax leads with are its own: BankerToolBench 76.2, DeepSWE v1.1 73.7, OSWorld V2 61.7. On DeepSWE, the chart puts DeepSeek V4.1 Flash 1017 ahead at 77.

Last, weights. MiniMax says they "will be released later", and VentureBeat says the plan is October 23, with no licence terms yet. AA lists the model as proprietary for now. Until the weights ship, MiniMax is the only paid host. OpenRouter carries a free variant but no paid listing yet, so nobody else can undercut the $0.15.

Who should switch

If your traffic runs during Asian or European business hours and you're on DeepSeek V4.1 Flash today, M3.1 Flash at a flat $0.15 and $0.60 is worth testing. You get a predictable bill and lose two index points. If you're a US team, stay on DeepSeek off-peak. It's cheaper on cached work, faster and higher scoring.

If you're on MiniMax M3, move. Plug your own cache hit rate and peak share into our calculator before you choose between the other two. On this mix the difference is a few dollars a month either way.

Pages we read on October 22

• MiniMax pricing and MiniMax M3.1 Flash announcement - rate cards for M3.1 Flash and M3, vendor benchmarks, speed claim, weights
• DeepSeek pricing - peak and off-peak cards, peak windows
• Alibaba Model Studio (with its context cache page), and Z.ai pricing - Qwen3.8 Flash and GLM-5.3-Flash rates
• Artificial Analysis model pages - Intelligence Index, cost per task, output tokens, speed and the rates AA priced each model at
• VentureBeat - M3.1 Flash's AA-Omniscience hallucination rate, DeepSWE chart values, weights date