What MiniMax Actually Announced on October 19

MiniMax rolled out an updated version of its M3.1 Flash artificial intelligence model on Monday, October 19, 2026, telling customers the refreshed system delivers a “generational” jump in agentic tasks, programming, Computer Use and cybersecurity work. The Beijing based startup said online inference speed has nearly tripled while the cost of running the model has dropped 70 percent, a combination that lands MiniMax squarely in the middle of an increasingly crowded price war among Chinese and US foundation model makers.

The announcement was first reported by Reuters and picked up by outlets including Fidelity’s newswire feed and WHTC, both of which carried the Reuters dispatch describing MiniMax’s statement in full. The model, MiniMax M3.1 Flash, is now available through the MiniMax API, the company’s M Plan subscription tier, and MiniMax Code, its coding focused product line. For developers tracking the broader shift toward cheaper, faster inference, the release fits a pattern this site has already covered in stories like DeepSeek’s V4.1-Flash-1017 API opening and the Claude Haiku 5.5 price cut that sparked an AI pricing war earlier this year.

What makes this launch notable is not just the speed and cost claims themselves, but the timing. MiniMax is pushing M3.1 Flash into the market at the same moment Gemini 4 Argon, GPT 6.1 Sol and Claude Opus 5.5 are all fighting for enterprise coding and agent workloads, and at a moment when Chinese model makers are explicitly positioning themselves as the low cost alternative for high volume agentic use cases. The sections below cover down what MiniMax actually announced, what remains unverified, how the pricing stacks up against rivals, and what the move could mean for the next few months of the AI infrastructure race.

·

Google · Preferred Sources

Don't miss new tech stories on Google

Add TrendinTech once in the Google app and our stories appear in your news suggestions.

Add Now

What MiniMax Actually Announced on October 19

According to Reuters, MiniMax said in a statement that the updated M3.1 Flash model delivers generational improvements in four specific areas: agentic tasks, programming, Computer Use, and cybersecurity. Those four categories matter because they overlap almost exactly with the workloads enterprises are currently trying to automate with AI agents, from writing and reviewing code to controlling a desktop environment on a user’s behalf.

The company’s own blog post, cited in reporting gathered by this outlet, framed the release as “redefining agent performance and cost.” MiniMax said online inference is now nearly three times faster than the prior version, and that usage costs for customers have fallen by 70 percent. Neither Reuters nor MiniMax’s public statements included a specific benchmark score, parameter count, or 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 also confirmed that model weights for M3.1 Flash will be released at a later date, without specifying when. That detail matters to the open source AI community, which has been watching Chinese labs closely since DeepSeek’s earlier releases reshaped expectations around what a freely downloadable frontier-adjacent model could do. Until weights ship, independent researchers and smaller developers will not be able to run M3.1 Flash locally or audit its behavior outside MiniMax’s own hosted API.

The Pricing: $0.15 In, $0.60 Out, $0.015 Cached

The most concrete numbers in Monday’s announcement are the API prices. MiniMax is charging $0.15 per 1 million input tokens, $0.60 per 1 million output tokens, and $0.015 per 1 million cached input tokens for M3.1 Flash. The model ships with a 1 million token context window, matching the long context ceiling that has become table stakes among flagship models over the past year.

For comparison, that pricing sits well below what US frontier labs charge for their flagship tiers, though it lands closer to the “flash” or “mini” tier pricing that OpenAI, Google and Anthropic already offer for lighter weight, high volume workloads. The distinction MiniMax is drawing is less about beating the absolute cheapest option on the market and more about offering frontier-level agentic and coding capability at a price point that makes high volume, latency sensitive deployments economically viable. That is the same pitch Mistral made with Large 4, which this site covered in Mistral Large 4 vs GPT-6 Astra: 12x Cheaper Output, and it echoes OpenAI’s own push toward volume pricing with its Decisions API public beta launch.

Cached input pricing at $0.015 per million tokens is worth separate attention. Caching discounts matter most for agentic workflows where a model repeatedly reasons over the same long context, such as a codebase, a document set, or a running conversation history. A 10x discount on cached tokens relative to fresh input tokens suggests MiniMax is explicitly optimizing for the kind of multi-turn, tool-calling agent loops that dominate current enterprise AI deployments, rather than one-shot question answering.

Model	Input price (per 1M tokens)	Output price (per 1M tokens)	Context window	Primary focus
MiniMax M3.1 Flash	$0.15	$0.60	1M tokens	Agentic tasks, coding, Computer Use, cybersecurity
DeepSeek V4.1-Flash-1017	Not disclosed in this report	Not disclosed in this report	Not disclosed in this report	Search, coding (DeepSWE)
Mistral Large 4	Lower than GPT-6 Astra per prior reporting	Up to 12x cheaper output than GPT-6 Astra	Not disclosed in this report	General + coding benchmark
Claude Haiku 5.5	Cut roughly 90% in recent price change	Cut roughly 90% in recent price change	Not disclosed in this report	Low cost, high volume tasks

That table should be read as directional rather than exhaustive. MiniMax’s figures come directly from its own pricing documentation and Monday’s statement. The comparison rows for DeepSeek, Mistral and Claude Haiku reflect what this site has previously reported in separate coverage and are included here only to illustrate the competitive pattern, not as a precise apples to apples benchmark, since each vendor measures and discloses pricing and context limits differently.

Why Agentic Tasks and Computer Use Are the Battleground

MiniMax’s choice to highlight Computer Use as a named capability is a signal about where the company believes the next wave of AI spending will go. Computer Use, the ability for a model to interpret a screen, move a cursor, click buttons and fill in forms, has moved from research demo to commercial feature over the past 18 months. Anthropic, OpenAI and Google have all shipped some version of this capability, and enterprise buyers increasingly ask vendors to prove their models can operate existing software rather than requiring a custom integration layer.

Pairing Computer Use with cybersecurity in the same announcement is also telling. Security teams have been both the biggest beneficiaries and the biggest worriers about agentic AI this year. A model that can autonomously navigate a desktop and reason about code can speed up vulnerability triage and patch testing, but the same capability in the wrong hands raises the risk of automated attacks. This site covered that tension directly in AI Agent Hacker Breaches 100+ Firms at $50 a Hit, where attackers used agentic tooling to compromise dozens of companies cheaply and at scale. MiniMax framing its own cybersecurity gains as a selling point suggests the company sees defensive tooling, not just offensive risk, as part of the pitch to enterprise customers evaluating M3.1 Flash.

Programming is the third leg of MiniMax’s stated improvements, and it is the most contested. Coding benchmarks have become the preferred battleground for AI labs precisely because they are relatively easy to measure and resonate strongly with developer audiences, the same group most likely to adopt and evangelize a new model. MiniMax’s own M3.1 Flash preview was already live on its Token Plan subscription before this broader relaunch, based on the company’s own social media posts, suggesting the company has been iterating on this model family for weeks or months before Monday’s wider rollout.

Historical Context: MiniMax’s Path From Chatbot Startup to Flash Model Maker

MiniMax built its early reputation on consumer facing products, including chatbot and roleplay applications that gained traction in parts of Asia before the company pivoted harder into enterprise grade foundation models. The shift mirrors a path other Chinese AI companies have taken as venture funding and government backed compute access made it more attractive to compete directly with OpenAI, Anthropic and Google on developer facing APIs rather than purely consumer apps.

The “Flash” naming convention itself is borrowed from an industry wide pattern. Google popularized the term with Gemini Flash as a lighter, faster, cheaper sibling to its flagship Gemini models. DeepSeek, Mistral and now MiniMax have all adopted similar naming to signal the same tradeoff: slightly reduced raw capability in exchange for dramatically lower cost and higher throughput. MiniMax’s own product changelog describes M3.1 Flash as cutting p95 and p99 first token latency by more than 90 percent in its coding assistant product, a separate and more specific claim than the general “nearly tripled” inference speed figure cited in Monday’s company wide statement.

That history matters for readers trying to judge how much weight to put on Monday’s claims. MiniMax has shipped incremental Flash updates before, including a preview version that went live on its Token Plan in the weeks leading up to this broader announcement. Monday’s release appears to be a formal, fuller rollout of a model line the company had already been testing with a subset of customers, rather than an entirely new model appearing without warning.

How M3.1 Flash Compares to Gemini 4 Argon and Claude Opus 5.5

Direct benchmark comparisons between M3.1 Flash and the current crop of US flagship models are not possible from Monday’s disclosures, since MiniMax did not publish a specific benchmark score alongside its speed and cost claims. That is a meaningful gap. When this site compared Gemini 4 Argon against Claude Opus 5.5 on coding tasks, both companies published specific benchmark percentages that let developers judge real world tradeoffs. MiniMax’s choice to lead with cost and speed multipliers rather than a benchmark score suggests the company is betting that price sensitive developers will be more swayed by a 70 percent cost reduction than by a marginal benchmark gain.

That strategy has precedent. Open and semi-open Chinese models have repeatedly used aggressive pricing to win developer mindshare even when benchmark leadership stayed with US labs. The approach pairs with broader market dynamics this site has tracked, including Nvidia’s shrinking share of China’s AI chip market as reported in Nvidia’s China Share Falls to 34%, Huawei Hits 28%. A domestic hardware ecosystem that increasingly favors homegrown chips changes the cost structure for Chinese AI labs training and serving their own models, which can translate directly into the kind of aggressive API pricing MiniMax unveiled this week.

Capability area named by MiniMax	Company’s stated claim	Independent verification status
Agentic tasks	“Generational” improvement	Not independently benchmarked in available reporting
Programming	“Generational” improvement	Not independently benchmarked in available reporting
Computer Use	“Generational” improvement	Not independently benchmarked in available reporting
Cybersecurity	“Generational” improvement	Not independently benchmarked in available reporting
Online inference speed	Nearly tripled (roughly 3x)	Company stated figure, no third party test cited
Usage cost	Reduced by 70%	Consistent with published API pricing structure

What MiniMax Has Not Confirmed

Several details that would typically accompany a major model launch are absent from Monday’s disclosures. MiniMax has not published a parameter count or active parameter count for M3.1 Flash, has not released specific benchmark scores against third party test suites, and has not disclosed a tokens per second throughput figure to substantiate the “nearly tripled” inference speed claim. The company also has not set a date for releasing the model’s weights publicly, saying only that a release will come later.

No named MiniMax executive or spokesperson was quoted directly in the Reuters report or in the company’s own blog post as reviewed for this story, beyond the company’s general statement. That is worth flagging because many AI launch stories attribute specific color commentary to named founders or research leads, and no such individual attribution appears in the documentation available for this particular release.

These gaps do not necessarily mean MiniMax’s claims are inflated. Companies frequently hold back granular benchmark detail for competitive reasons, especially when a model is still being rolled out gradually across API, subscription and coding product tiers. But the absence of third party verification means enterprise buyers evaluating M3.1 Flash for production workloads will need to run their own comparative tests before committing budget, rather than relying solely on MiniMax’s self reported multipliers.

Market Impact: What This Means for the AI Pricing War

MiniMax’s announcement lands in a market where price compression has become the dominant competitive dynamic among mid tier and flash class models. When Claude Haiku’s price fell roughly 90 percent earlier this year, as covered in this site’s report on the move, it triggered visible responses across the industry as competitors raced to avoid losing high volume customers to the cheapest viable option. A 70 percent cost cut from MiniMax, even without a hard benchmark to back the capability claims, adds more downward pressure to that same trend.

For startups building agentic products, that pressure is good news in the near term. Cheaper flash tier models mean lower infrastructure costs for companies running high volume, multi-step agent workflows, the kind of use case MiniMax is explicitly targeting with its Computer Use and agentic task framing. It also means the gap between what a well funded lab can afford to deploy and what a smaller team can afford continues to narrow, at least on raw API cost, assuming capability claims hold up under real world testing.

For incumbent US labs, the pressure cuts differently. OpenAI, Anthropic and Google have all built pricing tiers designed to capture high volume customers at the low end while preserving premium pricing for flagship capability at the high end. MiniMax undercutting that structure, even at the flash tier, forces a choice: match the pricing and compress margins, or hold the line and risk losing price sensitive developers to Chinese alternatives that increasingly offer comparable stated capability on paper.

Competitive Field: Chinese Labs Racing on Cost, Not Just Capability

MiniMax is not operating alone in this strategy. DeepSeek’s V4.1-Flash-1017 opened its API with a stated benchmark score of 77.0 on relevant coding evaluations, a figure this site detailed in its earlier coverage. The pattern across Chinese labs this year has consistently paired aggressive pricing with claims of narrowing or closing the capability gap against US flagship models, rather than conceding on either dimension alone.

That dual pressure, cheaper and claimed to be better, is harder for Western labs to dismiss than pure price competition would be. If MiniMax, DeepSeek and other Chinese labs were simply cheaper but clearly less capable, enterprise buyers with strict accuracy requirements could ignore the pricing pressure. But when companies claim both cost and capability gains simultaneously, even without full third party verification, it forces more serious evaluation cycles at companies that might otherwise have defaulted to an established US vendor.

Open source dynamics add another layer. MiniMax’s promise to eventually release M3.1 Flash’s weights, even without a firm date, keeps the company in the conversation with researchers and smaller developers who prioritize the ability to self host and audit a model over raw hosted API performance. That mirrors the open weight strategy that has helped Chinese labs build developer goodwill even when their hosted API products trail US competitors on headline benchmark leaderboards.

What Developers Should Watch Before Adopting M3.1 Flash

Teams evaluating M3.1 Flash for production use should treat Monday’s announcement as a starting point for testing, not a final verdict. The 1 million token context window and the cached input pricing make the model worth testing for long document and multi-turn agent workloads specifically, since that is where the pricing structure offers the clearest advantage over models that charge flat rates regardless of cache hits.

Latency sensitive applications are the other obvious test case, given MiniMax’s own claim about first token latency improvements in its coding assistant product. Developers running real time or near real time agent loops, where every additional second of model response time compounds across multiple tool calls, stand to benefit most if MiniMax’s throughput claims hold up under independent load testing.

Security and compliance teams evaluating the model for Computer Use or cybersecurity related tasks should proceed carefully given the lack of independent benchmark verification. A model that can control a desktop environment carries real operational risk if its actual error rate or security behavior differs meaningfully from MiniMax’s stated improvements. Running a controlled pilot before granting production level permissions remains the prudent path, regardless of how aggressive the pricing looks.

Five Predictions for the Next Six Months

• MiniMax will publish specific benchmark scores for M3.1 Flash within the next few months, following the pattern set by DeepSeek and other Chinese labs that initially led with cost claims before releasing detailed benchmark data.
• At least one major US flash tier model will see another price reduction in response to MiniMax’s 70 percent cost cut, continuing the pattern already visible after Claude Haiku’s price drop earlier this year.
• MiniMax will release M3.1 Flash’s model weights before the end of the first quarter of 2027, based on the company’s pattern of following preview rollouts with fuller public releases within a few months.
• Enterprise adoption of M3.1 Flash for Computer Use specific tasks will lag adoption for pure coding and agentic text tasks, as security teams demand more verification before granting desktop control permissions to a newly launched model.
• Competitive pressure from MiniMax, DeepSeek and Mistral will push at least one more US lab to introduce a cached token pricing discount similar to MiniMax’s $0.015 per million token rate, if one does not already exist by the time this is read.

The Bigger Picture for Enterprise AI Buyers

MiniMax’s M3.1 Flash launch is ultimately a data point in a much larger story about where AI infrastructure spending is heading. Enterprises are no longer simply asking which model scores highest on a leaderboard. They are asking which model delivers acceptable capability at a cost structure that survives scaling to millions of daily requests. That shift favors exactly the kind of flash tier, cost optimized releases MiniMax, DeepSeek, Mistral and the US labs have all been racing to ship this year.

The unanswered questions around M3.1 Flash, particularly the lack of published benchmark scores and the open timeline for weights release, mean the real test of Monday’s claims will play out over the coming weeks as independent developers run their own comparisons. If MiniMax’s 70 percent cost reduction and near tripled inference speed hold up under outside scrutiny, the model could become a serious option for cost conscious teams building agentic products. If the gap between marketing claims and real world performance proves wider than advertised, M3.1 Flash will join a long list of AI releases that generated a news cycle but failed to shift developer behavior at scale.

Frequently Asked Questions

What is MiniMax M3.1 Flash?
MiniMax M3.1 Flash is an updated AI model from the Chinese AI startup MiniMax, launched on October 19, 2026. The company says it delivers major improvements in agentic tasks, programming, Computer Use and cybersecurity, with a 1 million token context window.

How much does MiniMax M3.1 Flash cost to use?
MiniMax prices M3.1 Flash at $0.15 per 1 million input tokens, $0.60 per 1 million output tokens, and $0.015 per 1 million cached input tokens through its API.

Where can developers access M3.1 Flash?
The model is available through the MiniMax API, the company’s M Plan subscription, and MiniMax Code, its dedicated coding product.

Is MiniMax M3.1 Flash faster than the previous version?
MiniMax says online inference speed has nearly tripled compared to the prior version, though this figure comes from the company’s own statement and has not been independently benchmarked in available reporting.

Will MiniMax release the M3.1 Flash model weights?
MiniMax has said weights will be released later but has not confirmed a specific date.

How does M3.1 Flash compare to DeepSeek or Mistral’s latest models?
Direct benchmark comparisons are not possible from Monday’s disclosures since MiniMax did not publish specific benchmark scores. The company is instead emphasizing cost and speed multipliers, a strategy similar to pricing moves from DeepSeek’s V4.1-Flash-1017 and Mistral’s Large 4.

Is MiniMax M3.1 Flash suitable for production Computer Use tasks?
MiniMax names Computer Use as an improved capability, but given the lack of independent verification, security and compliance teams should run controlled pilots before granting the model production level desktop control permissions.

What does this launch mean for AI pricing overall?
MiniMax’s 70 percent cost reduction adds to a broader pattern of price compression among flash tier AI models in 2026, following similar moves from Claude Haiku and other competitors, and is likely to pressure additional price cuts across the industry.

Related Coverage

• OpenCode Console Breach Exposes OpenAI, Anthropic Keys [2026]
• Meta Blames Claude AI Tool for No Kings IG Ban [2026]
• Gemini 4 Argon Agents Beat Minecraft in 27 Hours [2026]
• Anthropic Pledges 5% Stake, Up to 20% Over Time [2026]
• Gemini API Key Setup: Build an App in 13 Steps [2026]