MiniMax M3.1 Flash Explained
MiniMax's Flash-tier agentic model costs $0.15 in and $0.60 out per million tokens. MiniMax announced it October 19, 2026.
MiniMax M3.1 Flash is a Flash-tier language model announced by Chinese artificial intelligence (AI) lab MiniMax on October 19, 2026. It targets agent workflows, software engineering, Computer Use, and cybersecurity tasks. MiniMax bills the model under the identifier MiniMax-M3.1-Flash at $0.15 per million input tokens, $0.60 per million output tokens, and $0.015 per million cached input tokens on the MiniMax developer platform.
Before the official release, MiniMax tested an anonymous early version codenamed "Space Bunny", which processed more than 200 trillion tokens. MiniMax reports online inference nearly three times faster (vendor claim) and, separately, agentic task costs 70 percent lower than its predecessor. MiniMax has said weights will be released later and has not published a context window.
What MiniMax M3.1 Flash Is and What Changed
MiniMax M3.1 Flash is MiniMax's next-generation Flash-tier model designed for agentic execution, code generation, Computer Use, and security tasks. MiniMax announced the model on October 19, 2026, establishing a base price of $0.15 per million input tokens and $0.60 per million output tokens on pay-as-you-go billing.
MiniMax claims a generational leap in agentic tasks, coding, Computer Use, and cybersecurity. According to the MiniMax announcement, MiniMax says online inference is nearly three times faster (vendor claim). MiniMax also claims a 70 percent reduction in operational cost for agentic tasks relative to its predecessor, though the announcement text does not name the specific earlier model in that sentence.
The launch comes after testing through "Space Bunny", an anonymous early version that handled more than 200 trillion tokens prior to public availability. MiniMax lists direct hosted access in the announcement, and MiniMax stated that model weights will be released later. An earlier open-weight model is covered at MiniMax M2.
Pricing and Token Billing Structure
MiniMax M3.1 Flash charges flat rates of $0.15 per million input tokens, $0.60 per million output tokens, and $0.015 per million prompt-cached input tokens. MiniMax announced these rates for the model id MiniMax-M3.1-Flash in its launch announcement; the M3 comparison rates come from the MiniMax pay-as-you-go pricing page.
These rates reduce token expenses when compared against MiniMax M3. The pricing guide records the older MiniMax M3 model at $0.30 per million input tokens, $1.20 per million output tokens, and $0.06 per million cached input tokens for contexts up to 512,000 tokens. Comparing those figures shows that MiniMax M3.1 Flash cuts both input and output token rates by exactly 50 percent relative to MiniMax M3, while cached reads receive a 75 percent discount relative to M3 cache hits.
| Model | Input per 1M | Cached Input per 1M | Output per 1M |
|---|---|---|---|
| MiniMax M3.1 Flash | $0.15 | $0.015 | $0.60 |
| MiniMax M3 (up to 512k context) | $0.30 | $0.06 | $1.20 |
- Standard input: $0.15 per million tokens ($0.00015 per thousand tokens)
- Cached input: $0.015 per million tokens ($0.000015 per thousand tokens)
- Output text: $0.60 per million tokens ($0.0006 per thousand tokens)
- Input and output costs are half the recorded rates of MiniMax M3
MiniMax-Reported Benchmarks and Output Speed
MiniMax reports benchmark results, which are vendor-reported and not independently verified. In the MiniMax announcement, MiniMax publishes an OSWorld V2 score of 61.7 and an Agents Last Exam (ALE) score of 29.6. MiniMax claims both metrics rank first among Chinese models on those respective evaluations.
MiniMax reports a DeepSWE version 1.1 score of 73.7 percent. MiniMax reports 82 percent on CyberGym and 76.2 on BankerToolBench. MiniMax also states that MiniMax M3.1 Flash ranks among the top performers on the AutomationBench-AA benchmark suite.
In everyday deployment, MiniMax states that the model maintains an output speed exceeding 200 tokens per second (TPS). The median (P50) output speed is listed at 170+ TPS, which MiniMax attributes to more aggressive sparsity.
- OSWorld V2: 61.7 (MiniMax-reported; claimed first among Chinese models)
- Agents Last Exam (ALE): 29.6 (MiniMax-reported; claimed first among Chinese models)
- DeepSWE v1.1: 73.7 percent (MiniMax-reported)
- CyberGym: 82 percent (MiniMax-reported)
- BankerToolBench: 76.2 (MiniMax-reported)
- Everyday output speed: 200+ tokens per second; median 170+ tokens per second
Access Channels and Desktop Availability
MiniMax makes MiniMax M3.1 Flash available through three channels listed in the announcement. Developers call the model through the MiniMax API on the MiniMax developer platform under the id MiniMax-M3.1-Flash.
The model is available through the MiniMax M Plan. MiniMax also incorporates the model into MiniMax Code (desktop app).
MiniMax announced that downloadable model weights will be released later, without giving a fixed calendar date. MiniMax also stated that a Pro model that pushes intelligence further will arrive later (no date).
- Developer API via
MiniMax-M3.1-Flashon the MiniMax platform - Access via the MiniMax M Plan
- Desktop programming integration via the MiniMax Code application
- Model weights and a Pro model promised for future release
Unpublished Specifications and Evaluation Criteria
Several technical and commercial terms remain unpublished in the launch documentation. MiniMax has not published the total parameter count or the maximum context window supported on public API endpoints. While the pricing page lists MiniMax M3 for contexts up to 512,000 tokens (with a 2x multiplier above that), MiniMax has published no context window for M3.1 Flash as of October 19, 2026.
MiniMax also omits license terms for the future weights release, default organization rate limits, free usage tiers, and detailed M Plan pricing on the launch post.
MiniMax M3.1 Flash is not recommended for organizations that require on-premises model hosting immediately or teams that require third-party benchmark audits before procurement. That recommendation changes if MiniMax publishes a weights release date and license, or if independent evaluators verify the reported 73.7 percent DeepSWE score. To test the model on your own tasks, run your existing prompts through the MiniMax developer platform alongside whichever model you use now, if any. Price the runs with the listed rates to compare costs. Then evaluate task accuracy against your baseline before moving production traffic.
Frequently Asked Questions
- MiniMax M3.1 Flash is a Flash-tier language model announced by MiniMax on October 19, 2026. It is designed for agentic tasks, coding, Computer Use, and cybersecurity operations, priced at $0.15 input and $0.60 output per million tokens.
- MiniMax announced the model on October 19, 2026. Prior to launch, an anonymous version codenamed "Space Bunny" processed over 200 trillion tokens.
- MiniMax has said weights will be released later; no date and no license are published, so open-source status is unknown.
- MiniMax reports generation speeds exceeding 200 tokens per second in everyday use, with a median (P50) speed of 170+ tokens per second. MiniMax cites more aggressive sparsity.
- MiniMax reports benchmark scores of 73.7 percent on DeepSWE v1.1, 61.7 on OSWorld V2, 29.6 on Agents Last Exam (ALE), 82 percent on CyberGym, and 76.2 on BankerToolBench. These scores are vendor-reported and have not been independently verified.
Planning to Use MiniMax M3.1 Flash?
Book a free AI workflow audit with Layer3Labs to map where a model like this fits your team's work.
Book a Consultation