China’s MiniMax rolled out an updated version of its flagship speed-tier AI model on Monday, October 19, 2026, telling customers that MiniMax M3.1 Flash delivers what the company calls a “generational leap” in agentic tasks, coding, Computer Use, and cybersecurity. Reuters reported the launch directly from Beijing, in a dispatch syndicated by WHTC, where MiniMax said the update makes online inference nearly three times faster than its predecessor while cutting usage costs by 70%.
The announcement lands in the middle of an already crowded year for Chinese AI labs. MiniMax, DeepSeek, Zhipu, Alibaba’s Qwen team, and Moonshot AI’s Kimi have all been racing to ship faster, cheaper models aimed at the same audience: developers building agents and coding tools who care more about latency and token price than raw benchmark bragging rights. M3.1 Flash is MiniMax’s latest entry in that race, and the way it reached the market, quietly, without an immediate price sheet, says almost as much about the state of competition as the launch itself.
Don't miss new tech stories on Google
Add Tech Insider once in the Google app and our stories appear in your news suggestions.
What MiniMax Actually Announced on October 19
MiniMax said M3.1 Flash is now available through three channels: the MiniMax API, the company’s M Plan subscription, and MiniMax Code, its developer-facing coding agent tool. In its official announcement, the company stated 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%, according to MiniMax’s own release materials.
MiniMax listed API pricing at $0.15 per million input tokens, $0.60 per million output tokens, and $0.015 per million cached-input tokens, putting it firmly in the budget tier of commercial LLM pricing. Developers can review the call parameters in MiniMax’s own API documentation. The model ships with a 1 million token context window, large enough to hold entire codebases, long chat histories, or multi-document research sessions in a single pass. MiniMax has not yet published a confirmed date for releasing the model’s open weights, a step the company has taken with earlier releases but did not commit to on launch day.
What makes the rollout notable is the path M3.1 Flash took to get here. A preview version, M3.1-Flash-Preview, had already appeared inside MiniMax Code on September 27, 2026, roughly three weeks before Monday’s formal announcement. According to Startup Fortune, that preview shipped without a public model card, a benchmark report, or even pricing information, an unusual way to introduce a model that MiniMax would later market as a flagship-tier upgrade.
The “Space Bunny” Mystery That Preceded the Launch
Before MiniMax put its name on the model, developers on OpenRouter had already been testing it under a different identity. An anonymous, free-to-use listing called Space Bunny Alpha began showing up on the router in late September, and community researchers started noticing that its outputs, latency profile, and coding behavior lined up closely with what would become M3.1 Flash, a pattern tracked in detail by Ground News’s aggregated coverage of the launch. MiniMax eventually confirmed the connection itself, posting on its official account that Space Bunny “has topped global leaderboards with 200T+ tokens” and was, in fact, an earlier version of M3.1 Flash.
This kind of stealth-testing pattern, releasing a model anonymously to gather real-world usage data and leaderboard standing before attaching a brand name and price tag to it, has become increasingly common among frontier labs racing to tune models against live traffic rather than synthetic benchmarks alone. For MiniMax, it meant M3.1 Flash entered the market with weeks of unlabeled production testing already behind it, a quieter and cheaper form of pre-launch validation than a traditional release cycle.
Inside MiniMax Code, the preview exposed five distinct reasoning levels, low, medium, high, xhigh, and max, giving developers granular control over how much compute the model spends thinking before it answers. That kind of tiered reasoning control has become a competitive expectation rather than a differentiator; rival labs including DeepSeek with its V4.1-Flash-1017 update have shipped similar adjustable-effort settings in recent months.
Pricing: Where M3.1 Flash Sits Against the Field
MiniMax’s stated pricing, $0.15 per million input tokens and $0.60 per million output tokens, places M3.1 Flash squarely in the low-cost tier that Chinese AI labs have used to compete with better-funded US rivals. Analysis from FourWeekMBA framed the pricing as landing close to DeepSeek’s off-peak discount rate, but available around the clock rather than only during specific hours, a distinction that matters for enterprise customers running continuous agent workloads rather than batch jobs scheduled around a discount window.
The 70% cost reduction MiniMax is claiming applies to the jump from its prior generation of models to M3.1 Flash, not necessarily to every competitor’s current pricing. Still, the broader trend is unmistakable: token prices for mid-tier coding and agentic models have been falling across the board throughout 2026, driven partly by Chinese labs undercutting each other and partly by efficiency gains in inference serving. MiniMax’s own roadmap illustrates the churn. Reports indicate that M Plan and Token Plan requests previously routed to its M2.5 and M2.7 models are now being redirected to M3.1 Flash, with the older models scheduled for retirement on November 18, 2026, a one-month runway for customers to migrate their integrations.
| Model | Input Price (per 1M tokens) | Output Price (per 1M tokens) | Context Window | Open Weights |
|---|---|---|---|---|
| MiniMax M3.1 Flash | $0.15 | $0.60 | 1,000,000 tokens | Not yet confirmed |
| MiniMax M2.7 (retiring) | Not disclosed | Not disclosed | Not disclosed | Not disclosed |
| DeepSeek V4.1-Flash-1017 | Varies by plan | Varies by plan | Not disclosed in this report | Yes, MIT-licensed |
| Cached input tokens (M3.1 Flash) | $0.015 | N/A | N/A | N/A |
Figures for MiniMax’s retiring M2.5 and M2.7 models were not disclosed in MiniMax’s public announcement materials. DeepSeek’s V4.1-Flash-1017 pricing varies by access plan, and this report does not independently confirm exact figures for that model beyond its earlier MIT license move, covered separately by this site in its report on the MIT-licensed update.
Why Agentic Tasks and Computer Use Are the Real Battleground
MiniMax’s choice to lead its announcement with agentic tasks and Computer Use, rather than general chat quality or creative writing, reflects where the commercial pressure in the LLM market has shifted. Enterprise buyers are no longer just asking which model writes the best paragraph; they are asking which model can reliably click through a web form, call an API, write and test a patch, and know when to stop and ask for help. Cybersecurity made MiniMax’s list of improvement areas too, a category that increasingly means a model’s ability to reason about vulnerabilities, triage alerts, and avoid introducing exploitable bugs into generated code, rather than just answering security trivia questions.
MiniMax has positioned the Flash line specifically for high-frequency, lower-latency developer workflows, like routine bug fixes and incremental feature work, rather than reserving it for the hardest, most complex agentic tasks that might justify a slower, more expensive flagship model. That positioning mirrors a split that has emerged across nearly every major AI lab’s product lineup in 2026: a fast, cheap “Flash” or “mini” tier for volume workloads, and a slower, pricier tier for tasks that genuinely need maximum reasoning depth. Anthropic, OpenAI, and Google have all built similar two-speed lineups, and MiniMax’s Flash branding borrows directly from that now-standard naming convention.
The company’s own messaging reinforced that framing. In a post introducing the preview, MiniMax wrote that its latest text model is “built for everyday development” and “fast, reliable, and ready for real work, from quick bug fixes to full features,” according to the company’s official account. That is a deliberately modest pitch compared to the “generational leap” language used in the broader launch statement, and the gap between the two messages hints at a company trying to serve two different audiences, enterprise buyers who want bold claims and individual developers who want a dependable daily-driver model, at the same time.
The Benchmark Gap: Claims Without Published Scores
One detail stands out against the confidence of MiniMax’s launch language: at the time of the Reuters report, no widely published, weighted benchmark score had been made available for M3.1 Flash. Startup Fortune’s earlier coverage of the preview noted the same gap, pointing out that the model shipped without a model card or benchmark report when it first appeared inside MiniMax Code in September. That means the “nearly three times faster” and “70% lower cost” figures MiniMax is promoting are company-reported claims rather than numbers verified against a third-party leaderboard like SWE-bench, LiveCodeBench, or any of the agentic-task benchmarks that have become standard reference points for coding models.
This is not unusual for a same-day launch, benchmark suites often lag a release by days or weeks as independent testers get API access, but it does mean buyers evaluating M3.1 Flash against competitors are, for now, working from MiniMax’s own numbers. Readers comparing cost and speed claims across vendors should treat any vendor-reported benchmark figure the same way: as a starting point for evaluation, not a settled fact, until independent test suites publish their own results.
MiniMax’s Position in the Crowded Chinese AI Market
MiniMax’s launch did not happen in isolation. Chinese AI labs have shipped a dense sequence of competing releases throughout 2026, each one narrowing the previous gap in price, speed, or context length. DeepSeek pushed its own V4.1-Flash-1017 update as a free, MIT-licensed release, a move The Information reported shook up the broader AI industry when it landed, emphasizing open distribution over the subscription-gated approach MiniMax has favored. Zhipu’s GLM family has taken a different tack, drawing attention less for pricing and more for unusual behavior in its reasoning traces, something GLM-5.3’s update-aware reasoning research highlighted when Apollo Research flagged the model reasoning about its own training process in a small number of test runs.
Alibaba’s Qwen team and Moonshot AI’s Kimi have also remained active participants in this market, though specific October 2026 performance or revenue figures for those two were not available in the sources reviewed for this report and should not be treated as confirmed without additional reporting. What is clear from the pattern of releases this year is that no single lab has managed to hold a pricing or speed lead for more than a few weeks before a rival undercuts it. MiniMax’s 70% cost-reduction claim for M3.1 Flash, even if ultimately validated by independent benchmarks, is likely to face a response from at least one competitor within the same quarter.
That dynamic also plays out against a backdrop of tightening policy scrutiny in the West. The US Department of Homeland Security recently invoked a 2018 federal law to restrict DeepSeek from federal IT systems, a move this site detailed in its coverage of the DHS DeepSeek ban. MiniMax has not been named in similar federal restrictions, but the broader regulatory climate around Chinese-developed AI models in government and enterprise settings remains unsettled, and any lab marketing aggressively to Western developers now operates under that shadow.
How M3.1 Flash Compares to the Open-Weight Field
Open-weight availability has become one of the clearest dividing lines in the current LLM market. Labs like Reflection AI have leaned hard into fully open releases, shipping models such as Beam, a 501-billion-parameter model with roughly 23 billion active parameters released under an Apache 2.0 license, a move covered in this site’s report on Reflection’s open-weight launch. DeepSeek has followed a similar open-distribution strategy with its Flash line. MiniMax, by contrast, has not committed to a confirmed date for releasing M3.1 Flash’s weights, keeping the model gated behind its API, M Plan, and MiniMax Code offerings for now.
That gating matters for a specific slice of the market: developers and enterprises that want to self-host a model for data residency, latency, or compliance reasons rather than call a hosted API. Every quarter that MiniMax delays an open-weight release of its Flash line is a quarter that self-hosting-focused buyers default to DeepSeek, Reflection, or another lab willing to publish weights immediately. Given how closely MiniMax has mirrored DeepSeek’s pricing structure, according to FourWeekMBA’s analysis, it would not be surprising if an eventual weight release follows the same competitive logic: match whatever the highest-profile open rival is offering, with a lag of weeks rather than months.
| Lab | Latest Model Named Here | Open Weights Status | Primary Positioning |
|---|---|---|---|
| MiniMax | M3.1 Flash | Not yet confirmed | Low-latency agentic, coding, Computer Use |
| DeepSeek | V4.1-Flash-1017 | Yes, MIT license | Free, open, coding-focused |
| Reflection AI | Beam | Yes, Apache 2.0 | Large open-weight general model |
| Zhipu | GLM-5.3 | Varies by release | Reasoning and agentic behavior research |
| Alibaba / Moonshot AI | Qwen / Kimi families | Varies by release | Not detailed in current sources |
Historical Context: MiniMax’s Path to M3.1 Flash
MiniMax has spent the past several release cycles building out a Flash-branded product line aimed at the volume end of the developer market, while reserving its flagship numbered models for customers who need maximum capability regardless of cost. The jump to M3.1 Flash follows the same incremental pattern the company has used before: ship a preview quietly inside its own tools, let real usage shape the final release, then announce broader availability with aggressive cost and speed claims once the model has already proven itself in production. The retirement of M2.5 and M2.7 by November 18, 2026, suggests MiniMax is consolidating its Flash lineup around a single current-generation model rather than maintaining multiple parallel versions, a cleanup move that should reduce support overhead on MiniMax’s side even as it forces existing customers to re-test their integrations against the new model within a one-month window.
That consolidation also reflects a broader maturing of the Chinese open and semi-open AI ecosystem. Two years ago, Chinese labs were largely playing catch-up on raw capability. By late 2026, the competition has shifted toward efficiency, latency, and total cost of ownership, metrics where Chinese labs have increasingly set the pace rather than followed it. MiniMax’s decision to headline its launch with inference speed and cost reduction, rather than a benchmark score, is itself evidence of that shift.
Market Impact: What This Means for Developers and Enterprises
For developers already using MiniMax Code or the M Plan, the practical impact is immediate: a faster, cheaper model is now the default for many requests, with the older M2.5 and M2.7 tiers on a retirement countdown. Enterprises evaluating MiniMax against Western alternatives like Grok, Claude, and DeepSeek now have another data point in the ongoing price war among coding-focused models, though the absence of independent benchmarks for M3.1 Flash means any head-to-head comparison should be treated as provisional.
The $0.15/$0.60 per-million-token pricing, combined with a cheap $0.015 cached-input rate, is clearly aimed at high-volume agentic workloads where the same context gets reused across many calls, a pattern common in coding agents that repeatedly reference the same codebase. That caching discount could matter more to total cost than the headline input and output rates for teams running sustained agent sessions rather than one-off queries. Enterprises that have already standardized on ChatGPT, Claude, or Gemini for coding workflows are unlikely to switch wholesale over a single price announcement, but MiniMax’s aggressive pricing adds further downward pressure on what every lab in this tier can charge.
What Experts and Official Statements Say
MiniMax’s own public statements are the clearest window into how the company wants M3.1 Flash understood. The company said: “Today, we officially roll out MiniMax M3.1 Flash,” in its launch blog post, a terse framing for a release the company elsewhere described in far bolder terms. Coverage from HIMIAO Horizon laid out the same speed and cost figures in more detail shortly after the announcement went live.
In its fuller capability statement, MiniMax said: “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%,” according to the company’s investor relations announcement.
On availability, MiniMax said: “M3.1 Flash is now available through the MiniMax API, M Plan, and MiniMax Code,” confirming the three channels through which developers can now access the model, per the same launch post.
Ahead of the formal launch, MiniMax’s official account had already told developers: “MiniMax-M3.1 Flash Preview is now live on the Token Plan! Faster, lighter, and built for teams running high-volume, latency-sensitive workloads, now available under your existing Token Plan subscription, no extra setup required,” in a post on its official X account. Financial outlet Futunn also tracked the rollout, reporting on the drive’s availability and feature set as it reached developers.
And describing the model’s developer focus specifically, MiniMax said: “MiniMax’s latest text model, M3.1-Flash-Preview, debuts today on MiniMax Code. Built for everyday development, it’s fast, reliable, and ready for real work, from quick bug fixes to full features,” in the same announcement thread.
Competitive Comparison: MiniMax Versus the Flash-Tier Field
Every major AI lab now fields some version of a fast, cheap, high-volume model tier, and MiniMax’s M3.1 Flash enters a field that already includes DeepSeek’s Flash line, Google’s lighter Gemini tiers, and Anthropic’s and OpenAI’s own budget models. What differentiates MiniMax’s pitch is the combination of a 1 million token context window at sub-$1-per-million-token pricing, a combination few Western labs currently match at the same price point. DeepSeek remains the closest direct comparison given the two companies’ overlapping Chinese origin and similar strategy of using aggressive pricing to win developer mindshare, and FourWeekMBA’s analysis specifically framed MiniMax’s pricing as tracking DeepSeek’s discounted rate, just applied continuously rather than during off-peak hours only.
Where MiniMax currently lags its rivals is transparency. DeepSeek’s MIT-licensed releases and Reflection AI’s Apache 2.0 licensed Beam model both give developers full visibility into model weights and the ability to self-host without relying on MiniMax’s API uptime or pricing changes. Until MiniMax commits to and delivers an open-weight release of M3.1 Flash, that gap in transparency and self-hosting flexibility remains MiniMax’s clearest competitive disadvantage against the open side of the market, even as its pricing stays competitive against hosted-API rivals.
Predictions: Where This Goes Next
Several outcomes look likely in the weeks following this launch. First, expect independent benchmark results for M3.1 Flash to surface within the next two to four weeks as third-party testers gain full API access and run it through standard coding and agentic suites; until then, MiniMax’s performance claims will remain unverified by outside parties. Second, expect at least one rival Chinese lab, most plausibly DeepSeek, to respond with either a further price cut or a new Flash-tier release of its own within the current quarter, continuing the pattern of rapid one-upmanship that has defined this market segment throughout 2026.
Third, the retirement of M2.5 and M2.7 on November 18, 2026, will force a wave of integration testing among MiniMax’s existing customer base, and any rough edges in that migration could generate visible developer complaints on forums and social media in the following weeks. Fourth, if MiniMax does eventually release open weights for M3.1 Flash, expect that announcement to be timed defensively, likely in response to a competitor’s open release rather than on MiniMax’s own independent schedule, consistent with the reactive pattern this market has shown all year. Fifth, expect continued downward pressure on per-token pricing across the entire Flash-tier segment, with sub-$0.10-per-million-input-token pricing becoming plausible among at least one major lab before the end of 2026 if the current rate of undercutting continues.
The Bigger Picture for the AI Model Market
MiniMax’s M3.1 Flash launch is a data point in a much larger story: the commoditization of mid-tier coding and agentic AI models. As more labs ship models that can handle routine development tasks competently, the differentiating factors are shifting away from raw capability and toward price, latency, context length, and licensing flexibility. MiniMax’s strategy, undercutting on price while keeping its weights closed for now, represents one bet in that landscape. DeepSeek’s and Reflection AI’s fully open approach represents another. Which strategy wins more developer mindshare over the next several quarters will likely depend less on any single benchmark score and more on which labs can sustain their pricing and feature claims under real-world, independently verified testing.
For now, MiniMax has staked its claim with bold language and aggressive pricing, backed by weeks of quiet production testing under the Space Bunny alias. Whether the “generational leap” the company promises holds up against independent scrutiny is the question the next month of benchmark reporting should begin to answer.
Frequently Asked Questions
What is MiniMax M3.1 Flash?
MiniMax M3.1 Flash is an updated AI model from Chinese AI company MiniMax, launched on October 19, 2026. The company says it improves agentic tasks, coding, Computer Use, and cybersecurity compared to its prior generation, while making inference nearly three times faster and reducing usage costs by 70%.
How much does MiniMax M3.1 Flash cost?
MiniMax has listed pricing at $0.15 per million input tokens, $0.60 per million output tokens, and $0.015 per million cached-input tokens.
What is the context window for M3.1 Flash?
Reports indicate the model supports up to a 1 million token context window, large enough for extensive codebases or long conversation histories.
Where can developers access MiniMax M3.1 Flash?
MiniMax says the model is available through the MiniMax API, the M Plan subscription, and MiniMax Code, the company’s coding agent tool.
Will MiniMax release the model weights for M3.1 Flash?
MiniMax has stated that model weights will be released at a later date, but no confirmed release date has been established in the company’s official announcement as of this report.
What was Space Bunny Alpha, and how is it connected to M3.1 Flash?
Space Bunny Alpha was an anonymous model listed on OpenRouter that community researchers linked to MiniMax before the company publicly confirmed it was an earlier version of M3.1 Flash.
What happens to MiniMax’s older M2.5 and M2.7 models?
Reports indicate MiniMax is routing M Plan and Token Plan requests previously handled by M2.5 and M2.7 to M3.1 Flash, with the older models scheduled for retirement on November 18, 2026.
How does M3.1 Flash compare to DeepSeek’s latest model?
Analysts have noted that MiniMax’s pricing for M3.1 Flash tracks closely with DeepSeek’s discounted off-peak rate, though available continuously rather than only during off-peak hours. Independent benchmark comparisons between the two models were not available at the time of this report.
