MiniMax M3.1 Flash cuts token prices but agent costs depend on efficiency

MiniMax released its new M3.1 Flash model on Monday, October 20, 2026, announcing a fifty percent reduction in per-token pricing. The company claims this update reduces the total cost of running AI agents by seventy percent, though independent analysis suggests the actual savings for developers may be more modest depending on workload characteristics.

What happened

The launch introduces M3.1 Flash as a direct successor to the previous M3 model, with a pricing structure that significantly undercuts many competitors. MiniMax lists input tokens at $0.15 per million, output tokens at $0.60 per million, and cached input tokens at $0.015 per million. The company asserts that inference speeds are nearly three times faster than M3, contributing to the overall efficiency gains. However, the headline seventy percent cost reduction claim relies not just on cheaper rates, but on the assumption that the model completes tasks in fewer processing loops.

Independent firm Artificial Analysis tested the model using its Intelligence Index benchmark. Their results showed an average cost per task of $0.22 for M3.1 Flash, down from $0.51 for M3. This represents a fifty-seven percent reduction, which is substantial but falls short of MiniMax’s internal marketing figure. The discrepancy likely stems from differences in the specific workloads used for testing, as agent costs vary wildly based on how many tokens are consumed during repeated reasoning steps.

The release also clarified the identity of Space Bunny, a mystery model that had previously processed over two hundred trillion tokens. MiniMax confirmed that Space Bunny was an early build of M3.1 Flash. While the model supports a one-million-token context window, allowing large codebases to fit in a single prompt, developers who tested the preview version should not assume identical behavior in the final release. The model weights are expected to become available soon, reportedly around October 23, though the official announcement only states they will arrive later.

How it works

M3.1 Flash operates as a large language model optimized for agentic workflows, where the AI performs multi-step tasks like coding or data analysis. The cost efficiency comes from two factors: the lower price per token and the model’s ability to resolve tasks with fewer iterative calls. In agentic systems, the AI often loops, reading its own output and re-prompting itself until a task is complete. If M3.1 Flash requires fewer loops to reach a correct answer, the total token count drops, compounding the savings from the lower rate card.

However, input tokens typically account for the majority of usage in these workflows. This makes the caching mechanism critical. MiniMax charges $0.015 per million cached input tokens. When an agent repeatedly references the same context or codebase, caching prevents re-processing those tokens at the full input rate. The effectiveness of this feature depends heavily on the structure of the application and how much static context is reused across requests.

Key details

• Input tokens cost $0.15 per million, output tokens $0.60 per million, and cached input tokens $0.015 per million.
• Artificial Analysis measured a fifty-seven percent cost reduction per task, compared to MiniMax’s claimed seventy percent.
• The model scores 73.7 percent on DeepSWE v1.1, slightly behind Claude Opus 5 but ahead of GPT-5.6 Sol.
• On the AA-Omniscience benchmark, M3.1 Flash achieved fifty-six percent accuracy but exhibited a high hallucination rate, answering incorrectly rather than abstaining.
• Computer use capabilities scored 61.7 on OSWorld V2, trailing leading Anthropic and OpenAI models.
• Model weights are expected to be released around October 23, 2026, though no specific license terms have been published yet.

Why it matters

For software engineers building AI agents, the distinction between per-token price and per-task cost is vital. A cheaper rate card does not guarantee lower bills if the model requires more iterations to solve a problem. The finding that input tokens drive most agent costs means that caching strategies and context management will have a larger impact on the bottom line than the base output price. Teams must evaluate whether M3.1 Flash’s speed and lower rates offset any potential increase in loop counts or verification overhead.

The high hallucination rate identified by Artificial Analysis poses a significant risk for knowledge-heavy applications. The model attempted nearly every question in the Omniscience benchmark, with ninety-three percent of its non-correct responses being wrong answers rather than refusals. This behavior suggests that developers cannot rely on the model to self-regulate uncertainty. Implementing external verification layers, retrieval-augmented generation, or human-in-the-loop reviews becomes necessary to maintain reliability, which adds complexity and potential cost back into the system.

What you can do

• Run your own evaluations on representative agentic workflows to measure actual token consumption per task, rather than relying on headline pricing.
• Implement robust caching for static context and input data to maximize savings from the low cached input token rate.
• Add verification steps or guardrails to your pipeline to mitigate the risk of hallucinations in factual or security-sensitive tasks.
• Compare peak and off-peak pricing with competitors like DeepSeek, especially if your workload runs during UTC daytime hours.
• Wait for the official license terms before planning to self-host the model once weights are released.
• Monitor the performance of the upcoming Pro model, as no release date has been set yet.