Cost calculator: where the bill actually goes
Three sliders. Prices are each provider’s list rate card. Cached input is the cheapest line on any rate card, so agent loops that re-send the same context save real money. Reasoning always runs on M3.1 Flash and bills as output. Slide output up to price the higher effort levels. MiniMax publishes no tokens-per-effort table, so treat lower effort settings as something to test on your own workload, not as a discount we can quote.
DeepSeek peak hours: 01:00–04:00 and 06:00–10:00 UTC on weekdays (Michael Leung). Off-peak, DeepSeek matches MiniMax’s headline rates; the cache line is where they differ ($0.015 vs $0.003). Prices: MiniMax docs via Michael Leung; DeepSeek docs; GLM via Michael Leung; GPT-6 Luna via vendor comparison.
Cost per finished task (Artificial Analysis)
Artificial Analysis ran the models through its weighted task set and counted input, caching, reasoning, and answer tokens. The result: MiniMax M3.1 Flash costs $0.22 per task, DeepSeek V4.1 Flash-1017 costs $0.26, and the older MiniMax M3 costs $0.51 (WinBuzzer).
What is MiniMax M3.1 Flash?
MiniMax M3.1 Flash is a mixture-of-experts model: 428 billion parameters in total, about 23 billion active per token (MiniMax, via MarkTechPost). It reads text, images, and video, and writes text. The context window is 1 million tokens.
Reasoning is always on and cannot be switched off. You pick one of five effort levels: low, medium, high, xhigh, or max, and max is the default. Reasoning tokens come out of the output allowance and bill as output, so the effort setting is a price knob, not a free quality dial. Requests that send thinking.type=disabled or effort=none return an error: an old M3 habit to unlearn (MarkTechPost).
The API speaks both the Anthropic and OpenAI formats, including Chat Completions and Responses, so most existing code needs only a base URL and key change. The model is live on platform.minimax.io, the M Plan subscription, and the MiniMax Code desktop app. MiniMax says inference runs above 200 tokens per second in everyday use, with a median above 170 (Tech Insider; AI-TLDR).
This is the general release of the model MiniMax first tested inside MiniMax Code as a preview. The preview was limited to M Plan and MiniMax Code; the full release adds pay-as-you-go API pricing and, MiniMax says, roughly three times faster inference.
How much does it cost?
| Model | Input / 1M | Cached input / 1M | Output / 1M |
|---|---|---|---|
| MiniMax M3.1 Flash | $0.15 | $0.015 | $0.60 |
| DeepSeek V4.1 Flash-1017, off-peak | $0.15 | $0.003 | $0.60 |
| DeepSeek V4.1 Flash-1017, peak | $0.30 | $0.006 | $1.20 |
| GLM-5.3-Flash | $0.15 | $0.03 | $0.50 |
| MiniMax M3 | $0.30 | $0.06 | $1.20 |
Rate cards via Michael Leung’s launch comparison (DeepSeek docs, MiniMax docs, Z.ai). MiniMax M3 input doubles above 512K context.
Benchmarks: strong agents, weak on saying “I don’t know”
The table mixes two kinds of numbers. MiniMax-reported scores come from the company’s own comparison table (which mostly compares against last-generation models, per the YouTube launch comparison). Independent scores come from Artificial Analysis. Read the label before the number.
| Benchmark | MiniMax M3.1 Flash | Rival | Reported by |
|---|---|---|---|
| AA Intelligence Index | 42 | DeepSeek V4.1 Flash-1017: 45; Claude Opus 5.5: 57.6 | Independent (AA) |
| DeepSWE v1.1 (coding) | 73.7% | DeepSeek Flash 1017: 74.2%; Opus 5: 74% | MiniMax table |
| Automation-Bench v1.0.6 | 62.2–62.8 | DeepSeek: 54.8 | MiniMax table |
| BankerToolBench (finance agents) | 76.2 | Ties Claude Opus 5 | MiniMax table |
| OSWorld V2 (computer use) | 61.7 | Opus 5: 70.6 | MiniMax table |
| Agents’ Last Exam | 29.6% | Opus 5.5: 38.2%; Gemini 4 Argon: 39.5% | MiniMax chart |
| Terminal-Bench 4.0 | 38.1 | Opus 5.5: 66.4 | Independent comparison |
| HLE (reasoning) | 48% | DeepSeek Flash: 43%; GLM-5.3-Flash: <48% | MiniMax (via Oossa) |
| CritPt (physics) | 22% | DeepSeek: 24% | MiniMax (via Oossa) |
| AA-Omniscience (factual) | 56% raw → 16 overall | — | Independent (AA, via VentureBeat) |
Space Bunny: the stealth launch
The launch was unusual: MiniMax let developers test the model for weeks before attaching a name to it.
- Late September 2026An anonymous model called “Space Bunny” appears on OpenRouter with no maker attached. Within three days it hits #1 in daily token volume, and stays there (OpenRouter, via Newscast.tech).
- October 2026Community testers guess a Chinese lab from the tokenizer; MiniMax later says the stealth run processed more than 200 trillion tokens. The company frames it as a public stress test: time to first token 1.95s → 1.22s, throughput 70 → 100 tokens/sec, error rate 1.4% → 0.7% (MiniMax, via Newscast.tech).
- October 19, 2026MiniMax confirms Space Bunny was an early build of M3.1 Flash and opens the general release: API, M Plan, and MiniMax Code.
- October 23, 2026Weights land on Hugging Face (MiniMaxAI/MiniMax-M3.1-Flash): 428 billion total, about 23 billion active per token, 1-million-token context, 252 GB, no access gate. The MiniMax Community license is free for non-commercial use; paid products must show the line “Built with MiniMax M3.1-Flash” and notify MiniMax, and revenue above $20 million a year needs written authorization. Self-hosting is real now; commercial plans need a license read first.
How to try it
Get an API key at platform.minimax.io (MiniMax warns never to share it), or use the M Plan subscription or the MiniMax Code desktop app. On OpenRouter it is listed as minimax/minimax-m3.1-flash, with a free tier reported at launch (Michael Leung). It also supports HTTP, the Anthropic SDK, and the OpenAI SDK (Wccftech).
One more consideration, separate from price and speed: MiniMax is a Chinese AI lab. Compliance teams should evaluate data residency and governance on their own terms; MiniMax’s benchmark release does not cover them (Shattered.io).
Related pages on this site: DeepSeek V4.1 Flash-1017 pricing and access guide, GPT-6 Luna pricing calculator, Claude Haiku 5.5 pricing and migration guide, Gemini 4 Argon pricing calculator.
FAQ
What is MiniMax M3.1 Flash?
MiniMax M3.1 Flash is a mixture-of-experts model from MiniMax, released October 19, 2026: 428 billion total parameters, about 23 billion active per token, 1-million-token context, built for agentic tasks, coding, computer use, and cybersecurity. It reads text, images, and video, and writes text.
How much does MiniMax M3.1 Flash cost?
$0.15 per million input tokens, $0.60 per million output tokens, $0.015 per million cached input tokens. Artificial Analysis puts the average finished agent-task cost at $0.22, down from $0.51 for the older M3.
What is Space Bunny, and is it really MiniMax M3.1 Flash?
Yes. Space Bunny was an anonymous model that appeared on OpenRouter in late September 2026 and quickly topped the daily token-volume charts. On October 19, 2026, MiniMax confirmed it was an early build of M3.1 Flash, and said it processed more than 200 trillion tokens while anonymous.
Is MiniMax M3.1 Flash open weights?
Yes, since October 23, 2026. MiniMax published the weights on Hugging Face (MiniMaxAI/MiniMax-M3.1-Flash): 428 billion parameters with about 23 billion active per token, 1-million-token context, 252 GB, no access gate. The MiniMax Community license is free for non-commercial use. Paid products must show the line “Built with MiniMax M3.1-Flash” and notify MiniMax; above $20 million in annual revenue, written authorization is required. It is neither MIT nor Apache: read the license before shipping a commercial product on these weights.
What are the five thinking levels?
Reasoning is always on. You pick low, medium, high, xhigh, or max effort; max is the default. Reasoning tokens bill as output, so the effort level directly changes the bill. Sending thinking.type=disabled returns an error.
How does it compare with DeepSeek V4.1 Flash-1017?
Headline rates match at off-peak ($0.15/$0.60). DeepSeek doubles its rates at peak hours while MiniMax is flat; MiniMax cached input ($0.015) costs more than DeepSeek’s off-peak cache ($0.003). On the AA Intelligence Index, DeepSeek scores 45 against MiniMax’s 42, at a $0.26 vs $0.22 cost per task.
Does MiniMax M3.1 Flash hallucinate?
Be careful with factual output. VentureBeat reported that Artificial Analysis found M3.1 Flash answers 56% of AA-Omniscience questions correctly in raw terms but scores only 16 overall, because 93% of its wrong answers were confidently incorrect rather than hedged or refused.
Where can I try MiniMax M3.1 Flash?
Through the MiniMax API at platform.minimax.io, the M Plan subscription, the MiniMax Code desktop app, or OpenRouter (minimax/minimax-m3.1-flash). The API supports both Anthropic-compatible and OpenAI-compatible formats.
What changed from MiniMax M3?
Halved per-token prices, a 1-million-token context, pay-as-you-go API access, five thinking effort levels, and roughly 3x faster inference at 200+ tokens per second. Agent-task cost fell from $0.51 to $0.22 in Artificial Analysis testing.
Can I use it with the OpenAI API?
Yes. MiniMax supports an OpenAI-compatible API including Chat Completions and Responses, plus Anthropic-compatible access over HTTP. Existing code mostly needs a base URL and key change.
Sources
- MiniMax announcement: M3.1 Flash (prices, speed, Space Bunny reveal, effort levels)
- Oossa ($0.22/task, AA index 42, HLE 48%, CritPt 22%, weights Oct 23)
- Pasquale Pillitteri (weights on Hugging Face Oct 23: 252 GB, no access gate; MiniMax Community license terms)
- MarkTechPost (428B/23B MoE, reasoning always-on, API formats, 150 tok/s product page)
- Michael Leung (rate-card table vs DeepSeek/Kimi/GLM, OpenRouter listing, migration notes)
- AI-TLDR (spec table: DeepSWE 73.7, OSWorld 61.7, 200+ tok/s)
- Shattered.io (AA-Omniscience 56% raw / 16 overall, 93% confidently wrong, via VentureBeat)
- Technology.org (57% vs 70% cost-cut caveat, stealth stats)
- Newscast.tech (Space Bunny stealth numbers: 56.7T tokens on OpenRouter, QA fixes)
- Madrobot.blog (MiniMax vs DeepSeek pricing and benchmark comparison)
- Tech Insider (speed figures, accuracy scrutiny)
- WinBuzzer (AA $0.22/task evaluation detail)
- YouTube launch comparison (prices and benchmarks vs Opus 5.5, GPT-6 Astra, Gemini 4 Argon)
- dev.to cost analysis (per-turn math and cache math vs DeepSeek Flash)
Ask an AI about this guide:
Agent-readable summary: /minimax-m3-1-flash/llms.txt