Editorial illustration for MiniMax M3.1 Flash at USD 0.15/M; Kimi K3 scores match table, cost chart diverges
MiniMax M3.1 Flash: $0.15/M Token Pricing
MiniMax put a new model on its API this week, M3.1 Flash, priced at $0.15 per million input tokens. That's roughly the cost tier of a lightweight assistant model, not a 428-billion-parameter mixture-of-experts system with a million-token context window. MiniMax also cleared up a question that had been floating around its developer channels: Space Bunny, the stealth model that quietly racked up more than 200 trillion tokens of usage before anyone outside the company knew what it was, turns out to have been an early build of this same Flash variant.
The parameter count, 428B total with about 23B active, matches MiniMax's earlier M3 release almost exactly, and the company hasn't published a technical report explaining what actually changed under the hood. What it has published is a scores table, claiming near-frontier results on agentic office tasks while landing mid-pack on harder coding benchmarks. Weights aren't available yet for anyone wanting to self-host; MiniMax says those arrive October 23, 2026. Until then, the API and MiniMax's own Code and M Plan products are the only way in.
MiniMax also confirmed that Space Bunny was an earlier M3.1 Flash build. That stealth model processed over 200T tokens. For developers, the message is near-frontier agentic scores at Flash-tier cost.
Why this matters
For developers pricing out agentic workloads, $0.15/M input tokens on a 428B MoE model with a 1M context window is the number that matters, not the parameter count. MiniMax is betting that cost-per-task beats cost-per-token as the metric people actually care about, and the Automation-Bench chart it's pushing is built around that argument. But the 62.8 vs 62.2 discrepancy between that chart and MiniMax's own table is the kind of detail worth flagging before anyone cites these numbers in a procurement doc.
Benchmark tables get screenshotted and repeated long after anyone checks the source, and a self-reported half-point gap on your own headline metric isn't a great look, even if it's just rounding. The Kimi K3 card matching MiniMax's DeepSWE and BabyVision figures is a useful cross-check and suggests the broader table holds up. Weights aren't landing until late October 2026, so anyone wanting to self-host or audit the model directly is stuck waiting more than a year.
Until then, this is an API-only bet, and the Space Bunny reveal is a reminder that MiniMax has been testing in the open longer than its announcements suggest.
Related Reading
Business & Startups Nordic pilot adds Gemini for Education, NotebookLM to boost AI literacy
Business & Startups Kling launches Video O1, all-in-one model with MVL bridge using transformer
Business & Startups DeepSeek Seeks More Capital Weeks After USD 7B Funding Round
From Archives Karnataka ElevenLabs partnership irks Indian voice AI startups over credibility
From Archives Anthropic cushions IPO with fund for workers its models might replace
Common Questions Answered
What is the pricing model for MiniMax M3.1 Flash and how does it compare to other models?
MiniMax M3.1 Flash is priced at $0.15 per million input tokens, which positions it in the lightweight assistant model cost tier despite being a 428-billion-parameter mixture-of-experts system with a million-token context window. This pricing strategy focuses on cost-per-task rather than cost-per-token, making it an economical option for developers building agentic workloads.
What was Space Bunny and how does it relate to M3.1 Flash?
Space Bunny was a stealth model that MiniMax had been running quietly, accumulating over 200 trillion tokens of usage before its existence became public knowledge. MiniMax later confirmed that Space Bunny was actually an earlier build of M3.1 Flash, demonstrating the model's real-world performance at scale before official release.
What are the key specifications of the MiniMax M3.1 Flash model?
M3.1 Flash is a 428-billion-parameter mixture-of-experts system that features a million-token context window, enabling it to handle large amounts of input data. Despite its substantial parameter count and context capacity, it maintains the low pricing of a lightweight assistant model at $0.15 per million input tokens.
Why does MiniMax prioritize cost-per-task over cost-per-token as a pricing metric?
MiniMax believes that developers building agentic workloads care more about the total cost to complete a task rather than the per-token expense, which is why they emphasize cost-per-task as the relevant metric. This approach is reflected in their Automation-Bench chart, which demonstrates near-frontier agentic performance at Flash-tier pricing levels.
Further Reading
- MiniMax M3.1 Flash: Redefining Agent Performance and Cost - MiniMax Research
- MiniMax M3.1 Flash Benchmarks, Pricing & Context Window - LLM Stats
- MiniMax M3.1 Flash: Price, Benchmarks, Space Bunny - CellCog
- Kimi K3 benchmarks and pricing, and why the open model market is changing - Hashnode
- Kimi K3 Pricing, Providers & Real-World Costs - DeepInfra
