MiniMax Launches M3.1 Flash, Claims 3x Faster Inference and 70% Lower Costs

On Monday, October 19, 2026, Chinese AI startup MiniMax launched an updated version of its M3.1 Flash model, telling customers the refreshed system delivers “generational” improvements in agentic tasks, programming, Computer Use and cybersecurity. The company says online inference speed has nearly tripled while usage costs have fallen by 70 percent, a combination that places M3.1 Flash directly in the industry’s escalating price war over fast, cheap models for high-volume AI workloads.

The Space Bunny prologue

Monday’s launch solved a mystery that had been circulating among developers for nearly a month. On September 23, 2026, an anonymous model called Space Bunny Alpha appeared on OpenRouter with its developer undisclosed, prompting speculation in the AI community. By the next day, September 24, it had reached the top of OpenCode’s daily token-usage rankings. MiniMax later confirmed that Space Bunny was an early version of M3.1 Flash being tested under a pseudonym.

Coverage of the launch reports that Space Bunny processed more than 200 trillion tokens during its anonymous testing period. OpenCode reported that the model held the leading position on its platform for 12 consecutive days and handled approximately 106 trillion tokens there. Those figures reflect reported platform usage rather than independently verified measures of model quality.

OpenRouter also reported measurable improvements while the anonymous model was live: time to first token fell from 1.95 seconds to 1.22 seconds, generation throughput rose from 70 to 100 tokens per second, and the error rate dropped from 1.4 percent to 0.7 percent. Actual performance may vary with workload, infrastructure and API demand.

Key details

The most concrete numbers in the announcement are the API prices: $0.15 per 1 million input tokens, $0.60 per 1 million output tokens, and $0.015 per 1 million cached input tokens. The model ships with a 1 million-token context window and supports multimodal inputs.

M3.1 Flash is available through the MiniMax API, the company’s M Plan subscription tier, and MiniMax Code, its coding-focused product line. Following the launch, OpenRouter announced a 72-hour promotional free-access period, while OpenCode announced a three-day offer.

On openness, the reporting diverges. TrendInTech reports that MiniMax confirmed the model weights will be released at a later date, without specifying when. BreakRead reports that MiniMax plans to release the model’s open weights on October 23, 2026, although the parameter count and licensing terms have not been disclosed. Either way, until weights ship, researchers cannot run the model locally or audit its behavior outside MiniMax’s hosted API.

MiniMax has not published a parameter count for the model. Neither the Reuters dispatch nor the company’s public statements, as reviewed in available reporting, included a tokens-per-second figure to back up the speed claim, so those numbers should be read as the company’s own characterization rather than an independently verified result.

MiniMax did publish benchmark comparisons, reported by BreakRead: on Automation-Bench the company reports 62.2 percent for M3.1 Flash against 60.5 percent for DeepSeek-V4.1-Flash-1017 and 50.3 percent for Claude Opus 5. The comparisons are not all favorable to MiniMax. On DeepSWE the company reports 73.7 percent against 77.0 percent for DeepSeek-V4.1-Flash-1017, and on BabyVision it reports 91.0 percent against 94.1 percent for Claude Opus 5. All of these are company-published figures and should not be treated as independently verified benchmark rankings.

MiniMax also confirmed plans for a higher-end Pro model, with no announced release date, pricing or detailed specifications.

Why it matters

M3.1 Flash arrives as price compression becomes the dominant competitive dynamic among mid-tier and “flash” class models. MiniMax’s bet is that developers running high-volume, latency-sensitive agent workloads will be swayed more by a 70 percent cost reduction than by a marginal benchmark gain. Cached-input pricing at $0.015 per million tokens, a steep discount on repeated context, is explicitly aimed at multi-turn, tool-calling agent loops rather than one-shot question answering.

The “Flash” naming itself signals the strategy. Google popularized the term with Gemini Flash as a lighter, faster, cheaper sibling to its flagship models, and DeepSeek, Mistral and now MiniMax have adopted similar naming for the same tradeoff: slightly reduced raw capability in exchange for dramatically lower cost and higher throughput.

The missing independent benchmarks are the caveat that should shape how buyers read this launch. Without third-party verification of the capability claims, enterprise teams evaluating M3.1 Flash for production workloads will need to run their own comparative tests before committing budget rather than relying on MiniMax’s self-reported multipliers.

Sources

• Reuters, “China’s MiniMax launches updated AI model M3.1 Flash, claims major improvements,” October 19, 2026 (reporting by Yukun Zhang, Shi Bu and Ryan Woo; editing by Louise Heavens), via KTWB
• TrendInTech, “What MiniMax Actually Announced on October 19,” trendintech.com
• BreakRead, “MiniMax M3.1 Flash Launches, Reveals Space Bunny AI,” breakread.com