The Pulse
MiniMax Launches M3.1 Flash With 70% Lower Agentic-Task Costs
M3.1 Flash cuts usage costs for agentic tasks by 70% and makes online inference nearly three times faster, MiniMax says. The company also lists API prices and reports benchmark results across coding, computer use and cybersecurity.

MiniMax launched M3.1 Flash on October 19 with a sharp cost claim: agentic tasks, the company says, cost 70% less than they did with its predecessor. But its announcement gives fixed prices per million tokens without explaining the workload or calculation behind that broader comparison. The release pairs the savings claim with benchmark scores and a speed claim, all published by MiniMax.
MiniMax sets API rates at $0.15 per million input tokens
The company lists M3.1 Flash at $0.15 per million input tokens, $0.60 per million output tokens and $0.015 per million cached input tokens. Those are the model’s stated API rates; they are not, by themselves, a breakdown of the 70% reduction in total agentic-task costs. MiniMax does not specify a reference task or show how it calculated that overall comparison in the announcement.
“For agentic tasks, the model’s overall usage cost is 70% lower than its predecessor.”
MiniMax, the company
M3.1 Flash also makes online inference nearly three times faster, MiniMax says. Output speeds reach more than 200 tokens per second in everyday use, with median speeds consistently above 170 tokens per second, according to the company. The post attributes the efficiency improvements to more aggressive sparsity, but does not give a test setup for the speed figures.
Benchmark scores cover coding, computer use and security
MiniMax reports a score of 73.7% on DeepSWE v1.1, a benchmark for complex software tasks, and 76.2 on BankerToolBench, which it describes as testing multi-step planning and chained tool calls. On computer-use evaluations, the model scores 61.7 on OSWorld V2 and 29.6 on Agents’ Last Exam. Those results rank it first among Chinese models on both benchmarks, MiniMax says.
The model also scored 82% on CyberGym, which evaluates vulnerability discovery and validation, the company reports. Internal security experts tested the model on code auditing, vulnerability patching, penetration testing, malware analysis and reverse engineering, according to MiniMax. Those cybersecurity findings are MiniMax’s account of its own evaluations; the announcement does not provide further testing detail in its text.
Space Bunny was an early M3.1 Flash version
The launch identifies Space Bunny, an anonymously operated model that MiniMax says processed more than 200 trillion tokens, as an early version of M3.1 Flash. That disclosure connects the release to a model that had been used before MiniMax publicly named it. The company does not give a date range or additional operational details for Space Bunny in the announcement.
API and product access start now; weights come later
M3.1 Flash is available through the MiniMax API, M Plan and MiniMax Code, according to the company. MiniMax will release the model weights later, without giving a date in the launch post. A Pro model will also arrive later, the company says, positioning Flash as the current release in a pair of models with different performance aims.
Sources
- Original source · Blog post MiniMax minimax.io
Related stories
- MiniMax Publishes M3.1 Flash Weights Under Community License
- Vercel adds MiniMax H3 Image models to AI Gateway
- MiniMax Confirms Export Registration; CCTV-Affiliated Account Reports DeepSeek Claim
- Tencent releases Hy4 flash with lower API prices than Hy4
- Baidu previews ERNIE 5.5 Flash with 1M-token context