MiniMax M3.1 Flash: $0.15/M tokens cuts agent costs 70%
MiniMax M3.1 Flash pricing at $0.15 per million input tokens cuts agentic task costs by 70% versus M3. The 428B MoE model runs three times faster.
MiniMax Group Inc. released M3.1 Flash, a 428-billion-parameter mixture-of-experts model priced at $0.15 per million input tokens. The launch targets the cost structure of agentic AI systems, where completing a single task often requires dozens of model calls. According to MiniMax, the model reduces agentic task costs by 70% relative to its previous M3 model and achieves inference speeds nearly three times faster than its predecessor.
What changed
MiniMax rolled out M3.1 Flash as an update to the M3 model launched in June. The model is available through the MiniMax API, M Plan subscription, and MiniMax Code desktop app. Output pricing is set at $0.60 per million tokens, with cached input at $0.015 per million. The company indicated that model weights will be released at a later date and that a more capable Pro variant is in development.
The model handles text, image, and video input with support for prompts up to 1 million tokens. It includes reasoning capabilities with five selectable effort levels ranging from Low to Max. MiniMax credits the speed improvements to increased sparsity during inference, allowing peak output speeds of 200 tokens per second and median performance above 170 tokens per second.
The launch revealed that Space Bunny, an anonymous model that appeared on OpenRouter in late September, was an early M3.1 Flash build. The model processed over 200 trillion tokens before MiniMax confirmed its identity.
Why it matters
At $0.15 per million input tokens, M3.1 Flash pricing aligns with DeepSeek’s V4.1-Flash model during off-peak hours. For agentic workloads that require many model calls per task, the 70% cost reduction directly impacts operational expenses. This pricing dynamic places pressure on other developers competing in the market for both efficiency and inference speed in agent-heavy applications.
Organizations running agents at scale, whether for customer service automation or software engineering workflows, gain access to cheaper computational resources. The threefold increase in inference speed also reduces wall-clock time for multi-step tasks, a factor that matters for real-time agent deployments. For teams evaluating cost and speed trade-offs across models, M3.1 Flash enters a competitive pricing band that now includes multiple capable options.
MiniMax reported benchmark results showing the model performs comparably to leading models on specialized tasks. On DeepSWE v1.1, a software engineering test, M3.1 Flash achieved 73.7%, compared to 77% for DeepSeek’s V4.1-Flash. On CyberGym, a cybersecurity benchmark developed by UC Berkeley researchers to assess vulnerability discovery capabilities, the model scored 82%. For computer use, M3.1 Flash reached 61.7 on OSWorld V2 and 29.6 on Agents’ Last Exam, ranking first among Chinese models on both tests.
The model’s multimodal input support and million-token context window position it with other frontier options. Its emphasis on agentic performance signals a market shift toward models optimized for autonomous task execution rather than conversation alone.
What to test
Before adopting M3.1 Flash, independent verification of the reported benchmarks is advisable. The company claims no Chinese model has scored higher on OSWorld V2 and specific performance figures on BankerToolBench and CyberGym, but these results lack external verification. Your real-world performance on specific workflows may differ from benchmark numbers.
Test the pricing claim directly by confirming that your agentic workload actually costs 70% less than M3 in your production environment. Actual savings depend on how many tokens your agent consumes per task and whether the model’s reasoning capabilities suit your needs. The adjustable reasoning feature ranging from Low to Max effort affects both speed and cost, making it worth profiling that trade-off in your setup.
Evaluate the reported inference speed gains in your production deployment. While MiniMax reports peak speeds of 200 tokens per second, median performance at 170 TPS may be more representative of typical operation. Determine whether the threefold faster inference materially reduces latency for your agent loops, particularly if you run multiple tasks in parallel.
Confirm the timeline for model weights availability. The company stated weights will be released later without specifying when. If open-weight deployment is critical to your architecture, clarify that release schedule before committing resources. Similarly, verify that the Pro model release date aligns with your roadmap if higher capability is required.
Common questions
How does M3.1 Flash pricing compare to M3?
MiniMax set M3.1 Flash input pricing at $0.15 per million tokens and output at $0.60 per million. The company states that agentic tasks cost 70% less to run on Flash than on the prior M3 model, substantially reducing expenses for workflows involving multiple model calls.
What benchmarks demonstrate M3.1 Flash performance?
On DeepSWE v1.1, a software engineering test, M3.1 Flash scored 73.7%, versus 77% for DeepSeek V4.1-Flash. The model also achieved 82% on CyberGym cybersecurity tasks and ranks first among Chinese models on OSWorld V2 and Agents’ Last Exam, though these results have not been independently verified.
When will the model weights be available?
MiniMax indicated that M3.1 Flash weights will be released at a later date but did not provide a specific timeline. The company also confirmed that a more capable Pro model is coming, but no release date has been announced for that variant either.
The conclusion
M3.1 Flash enters a narrowing competitive space where pricing alone no longer serves as a differentiator among capable models. Performance depends on how well the model handles the specific tasks your agents need, whether code review, security scanning, business process automation, or computer use. The 70% cost reduction applies specifically to agentic workloads, making it worth testing if your current setup involves many model calls per task. The speed improvements are substantial enough to affect production latency, but measuring them in your own environment is essential.
The transition from M3 to Flash reflects the market’s focus on efficiency for agentic applications. Organizations deploying agent systems at scale should evaluate M3.1 Flash on their actual workflows before migrating. The pricing alignment with DeepSeek suggests neither vendor will easily undercut the other, indicating that agent-optimized models at competitive costs have become the market standard.
Related reading
AI Tool Herald may earn a commission from some links on this site. It never changes what we report or recommend. Affiliate disclosure
Related stories
GPT-6.1 Sol and Claude Opus 5.5: 75% price cuts reshape LLM market
OpenAI and Anthropic slashed LLM prices dramatically as open-weight models and inference provider fragmentation drive commodity-like competition in 2026.
AKV cache engineering cuts agent costs 27.7%
Tsinghua's AKV system lets LLM agents drop and reposition KV cache instead of summarizing context, cutting costs 27.7% while improving accuracy by 1.7 points.
Nanbeige4.3-33B scores 80.9 on Terminal-Bench
Nanbeige4.3-33B is a 33B open-weight coding model with 512K context and Apache 2.0 license. It scores 80.9 on Terminal-Bench 2.1, offering a local alternative to...