MiniMax launches M3.1 Flash, claiming 70% lower AI costs

MiniMax launches M3.1 Flash, claiming 70% lower AI costs

MiniMax launched M3.1 Flash on October 19, 2026, claiming a 70% reduction in usage costs for agentic workloads and online inference nearly three times faster than its predecessor. The Chinese AI company is targeting developers and businesses running multistep AI tasks, where repeated model calls can quickly increase costs and delays.

The release adds another contender to a crowded market that includes DeepSeek, Moonshot AI’s Kimi line and Zhipu’s GLM series. But MiniMax’s headline savings and performance figures remain company claims, rather than independently verified results.

A push toward cheaper AI agents

AI agents can make dozens or hundreds of tool calls while completing a single assignment. For businesses deploying them at scale, the cost of finishing a task can matter more than performance on an individual benchmark.

MiniMax positions M3.1 Flash as an efficiency-focused update to MiniMax M3, emphasizing lower operating costs and latency. Its "Flash" line targets high-volume, latency-sensitive workloads rather than benchmark performance alone.

In its official announcement, the company said: "Today, we officially roll out MiniMax M3.1 Flash", presenting the release as a formal production launch rather than a limited preview.

MiniMax summarized its claims this way: "As a next-generation Flash model, M3.1 Flash delivers a generational leap in agentic tasks, coding, Computer Use and cybersecurity, while making online inference nearly three times faster and reducing usage costs by 70%" (MiniMax).

The company has not published every technical detail buyers would need to evaluate those claims, including the model’s parameter count or active-parameter count. A fully itemized, independently verified pricing breakdown was also not available alongside the cost-reduction figure.

sbb-itb-622dc24

Long context and reported speed gains

M3.1 Flash supports a context window of up to 1,000,000 tokens, according to MiniMax. That capacity allows long-running agents to retain more conversation history and tool outputs before needing to summarize or truncate earlier material.

The company reports output speeds exceeding 200 tokens per second in everyday use, with median speed consistently above 170 tokens per second.

Those figures underpin MiniMax’s pitch to teams running coding, computer-use and other multistep workflows. Faster responses and lower usage costs could reduce the expense of repeated model calls, although buyers would need to test the model on their own workloads to establish actual savings.

The release follows MiniMax’s progression from its early abab series to MiniMax M1, a reasoning-oriented model associated with extended "thinking", and then the coding- and agent-focused MiniMax M2 and M2.1 updates. M3.1 Flash builds on MiniMax M3 with a focus on efficiency.

What the benchmarks show - and do not show

MiniMax published results across five benchmarks covering financial tool use, software engineering, computer operation, professional tasks and cybersecurity.

These are MiniMax’s own reported results. Independent verification of the exact figures using a standardized test harness was not available at launch.

The company also has not published a head-to-head scorecard comparing M3.1 Flash with named Chinese or Western rivals under a single shared test setup. Direct rankings against competing models therefore remain unverified.

The computer-use and cybersecurity results carry particular significance for deployment. A model that operates software can click buttons, complete forms and take actions across applications, but mistakes can have consequences beyond an incorrect text answer. Vulnerability-discovery capabilities can support defensive security work while also posing dual-use risks.

Availability and competitive pressure

On access, MiniMax said: "M3.1 Flash is now available through the MiniMax API, M Plan, and MiniMax Code" (MiniMax). Exact regional availability and subscription-tier details beyond the company’s public statements were not independently confirmed at launch.

The company also described its intended development uses, writing: "Built for everyday development, it’s fast, reliable, and ready for real work, from quick bug fixes to full features" (MiniMax, via X).

MiniMax joins Chinese competitors emphasizing long context windows, lower costs and agentic capabilities, while Anthropic, OpenAI and Google continue developing their own frontier agentic models.

For enterprise buyers, the launch offers another option to evaluate alongside existing vendors. The central question is whether MiniMax’s reported speed and cost advantages hold up on real workloads - not simply whether the model posts competitive benchmark scores. Until independent testing is available, the claimed 70% savings remain a starting point for evaluation rather than a verified measure of production costs.

Read the source