MiniMax just dropped M3.1 Flash, and I’m genuinely impressed by what they’ve managed here. After processing 200+ trillion tokens under the radar as Space Bunny, they’ve released a model that fundamentally changes the math on running agentic AI in production.
Let me be direct: this is the kind of efficiency breakthrough that actually matters. We’re not talking about marginal improvements or lab benchmarks that don’t reflect real usage. We’re talking about 70% lower costs for agentic tasks, output speeds hitting 200+ tokens per second, and competitive performance on the benchmarks that actually matter to builders.
The Real Cost Story
Pricing alone tells you something: $0.15 per million input tokens, $0.60 per million output tokens, with cached inputs at $0.015 per million. That’s aggressively priced, but what matters is the total cost-per-task metric. For agentic workloads, you’re looking at 70% savings compared to M3.1’s predecessor.
Here’s why that’s significant: agentic tasks are inherently expensive. Your agent thinks through multiple steps, reasons about problems, potentially calls external tools, and generates output iteratively. Every token compounds. A 70% reduction means workloads that were borderline uneconomical suddenly make sense at scale. You can run more sophisticated agents, more frequently, with tighter margins.
I’ve watched teams shelve agentic projects because the per-task cost was prohibitive. Flash changes that calculus immediately.
Performance Where It Counts
Speed isn’t just about user experience, though that matters. When you’re running agents in tight loops, latency compounds. Flash reaches median output speeds of 170+ TPS, with peak speeds hitting 200+. That’s meaningful for real-time agent interactions.
But I’m more interested in the benchmark results. Automation-Bench for complex agentic execution, DeepSWE for real-world coding tasks, and Agents’ Last Exam for computer use scenarios. These aren’t synthetic tests. They’re measuring what agents actually need to do: reason through complex problems, write code that works, and interact with computer interfaces to accomplish goals.
Flash performing at the level of leading models on these benchmarks while costing 70% less is the story. Performance parity at dramatically lower cost is how you shift an industry.
What This Means for Builders
If you’re working on AI agents or considering them, Flash is immediately worth evaluating. The combination of cost efficiency and competitive performance creates breathing room for experimentation. You can afford to run more complex agent architectures, longer reasoning chains, and more iterative refinement without economics forcing you back to simpler solutions.
For teams using inference APIs, this puts pressure on alternatives. Not just on pricing, but on the assumption that you need to choose between cost and capability. MiniMax is proving you don’t have to.
The rollout is also smart: Flash is available through the MiniMax API, M Plan, and MiniMax Code (desktop download available now). They’re meeting developers where they work, whether that’s through direct API calls or integrated desktop tooling. Model weights coming later suggests they’re planning for even broader adoption.
The Broader Implications
This release sits at an interesting inflection point. MiniMax is positioning Flash as the efficiency tier while promising Pro later as the capability tier. That’s a deliberate strategy: give developers an obvious cost-effective default, then offer a premium option for problem domains where you need maximum capability.
I think that’s the right approach for the next phase of AI infrastructure. We’ve spent cycles debating whether frontier models are necessary for every task. They aren’t. But until recently, the performance gap between frontier and efficient models made cost savings marginal. Flash closes that gap for entire categories of work, particularly agentic tasks where iterative reasoning and tool use matter more than raw scale.
Check out more on AI optimization to see how the industry is approaching similar efficiency challenges.
The question now isn’t whether Flash is worth testing. It’s whether your current setup is costing you 70% more than necessary while delivering comparable results.