MiniMax’s newest speed-tier model is cheaper and faster than its predecessor by a wide margin, but a wave of independent testing published alongside the October 19, 2026 release is raising a different question: does MiniMax M3.1 Flash trade accuracy for that speed? The Shanghai-based AI lab says the model delivers a “generational leap” in agentic tasks, coding, Computer Use, and cybersecurity, cutting usage costs by 70% and nearly tripling online inference speed over the prior M3 generation, according to the official MiniMax announcement. Within 24 hours, however, VentureBeat and The Outpost had each published head-to-head testing suggesting the model’s error rate climbs noticeably even as its benchmark scores and price tag both look favorable.
That split verdict, cost efficiency on one hand, accuracy concerns on the other, is the real story emerging from MiniMax M3.1 Flash’s first days on the market. It also says something broader about where China’s AI price war is headed in late 2026: speed and price cuts are now table stakes, and the next competitive axis is reliability under real workloads.
What MiniMax Actually Shipped on October 19
MiniMax rolled out M3.1 Flash on October 19, 2026, making it available through the MiniMax API and MiniMax Code, the company’s developer-facing coding agent tool. The model supports a 1-million-token context window, roughly equivalent to 1,500 pages of text, and MiniMax priced it at $0.15 per 1 million input tokens, $0.60 per 1 million output tokens, and $0.015 per 1 million cached input tokens. The company says typical output speed exceeds 200 tokens per second during everyday peak usage, with a median output speed above 170 tokens per second.
MiniMax framed the release around four capability areas: agentic tasks, coding, Computer Use, and cybersecurity. The company’s own benchmark testing put M3.1 Flash at 61.7 on OSWorld V2, a widely used Computer Use benchmark, and 76.2 on BankerToolBench, an agentic finance-tool benchmark. MiniMax also named Automation-Bench, DeepSWE, and Agents’ Last Exam (ALE) as benchmarks used in its evaluation, though it did not publish every specific score for those three in its release material. Separately, the company said its weighted-average task cost for agentic work fell to $0.22 per task on M3.1 Flash, down from $0.51 per task on the M3 predecessor, which SiliconANGLE also reported as the company’s headline cost-reduction figure.
Readers looking for the full rundown of MiniMax’s launch-day claims, including the speed and pricing figures, can find more detail in tech-insider’s coverage of the M3.1 Flash launch itself. This article focuses on what happened next: the scramble by outside testers to check whether the cost cuts hold up once accuracy enters the picture.
Breaking Down MiniMax’s Benchmark Claims
Benchmark comparisons across AI labs are notoriously hard to standardize, since each company tends to publish the scores that flatter its own model. Still, the numbers MiniMax did disclose give a reasonably clear picture of where M3.1 Flash sits relative to recent releases from DeepSeek and other agentic-focused competitors. The table below collects the figures that have been independently reported by named outlets, alongside the comparable score for DeepSeek V4.1-Flash-1017, a model tech-insider covered in depth after it hit 77.0 on DeepSWE and opened its own search API in October.
| Benchmark | MiniMax M3.1 Flash | DeepSeek V4.1-Flash-1017 | What it measures |
|---|---|---|---|
| OSWorld V2 | 61.7 | Not disclosed in MiniMax comparison | Computer Use / desktop automation |
| BankerToolBench | 76.2 | Not disclosed in MiniMax comparison | Agentic finance tool-use tasks |
| DeepSWE v1.1 | Reported in the 70s range | 77.0 | Agentic software-engineering tasks |
| Weighted avg. task cost | $0.22 | Open-weight, self-hosted | Blended agentic-task cost |
| Context window | 1,000,000 tokens | Large context (model-specific) | Maximum input length |
The pattern that emerges is one of rough parity on raw capability, with MiniMax M3.1 Flash competitive on agentic and tool-use benchmarks while undercutting most proprietary rivals on price. That is a familiar shape for 2026’s Chinese AI releases: DeepSeek, Zhipu’s GLM line, Alibaba’s Qwen team, and Moonshot AI’s Kimi have all pursued the same strategy of shipping frontier-adjacent capability at a fraction of the cost charged by US labs. MiniMax’s pitch with M3.1 Flash is that it can match or beat that competitive set on agentic workloads specifically, the kind of long-running, tool-calling tasks that increasingly define how enterprises evaluate models rather than single-turn chat quality.
The Hallucination Question Independent Testers Are Asking
The headline that cuts against MiniMax’s cost-cutting narrative came from VentureBeat, whose testing concluded that M3.1 Flash beats both DeepSeek and Anthropic’s Opus 5 on a key business benchmark at a lower price point, but does so with a markedly higher hallucination rate than either rival, according to VentureBeat’s reporting. The outlet’s framing is notable because it does not dispute MiniMax’s performance or pricing claims outright, it simply adds a dimension that MiniMax’s own release material did not emphasize: how often the model confidently states something false.
The Outpost reached a similar conclusion from a different testing angle, reporting that MiniMax M3.1 Flash cuts AI agent costs by 70% as advertised, but that independent tests reveal accuracy trade-offs serious enough to flag for anyone deploying the model in production agentic workflows, per The Outpost’s analysis. Neither outlet published a single standardized hallucination percentage that can be compared cleanly across models, which is itself a reminder that hallucination measurement remains one of the least standardized corners of AI evaluation. What both reports agree on is directional: M3.1 Flash is faster and cheaper, and it appears to hallucinate more often than the higher-priced models it is being compared against.
That trade-off matters differently depending on the use case. A coding agent that hallucinates an API signature will usually fail fast, a compiler or test suite catches the error before it ships. A cybersecurity agent or a finance tool-use agent that hallucinates a false positive, or worse, a false negative, carries a much higher cost of error. MiniMax explicitly markets M3.1 Flash for cybersecurity and agentic finance tasks (the BankerToolBench score is built for exactly this), which puts extra weight on the accuracy question that VentureBeat and The Outpost are both raising.
Why Flash-Tier Models Tend to Hallucinate More
There’s a structural reason “Flash” or speed-tier models across the industry, not just MiniMax’s, tend to show higher error rates than their flagship siblings. These models are generally optimized for lower latency and lower compute cost per token, often through more aggressive distillation or smaller active parameter counts at inference time. That optimization target is in direct tension with the kind of careful, multi-step verification that reduces hallucination. MiniMax has not published the exact parameter count for M3.1 Flash; a widely circulated figure of 428 billion total parameters with 23 billion active has appeared in some secondary coverage, but that figure is not confirmed by MiniMax’s own released results, so it should be treated as unverified rather than fact.
Pricing Math: How the 70% Cost Cut Actually Breaks Down
MiniMax’s cost claim is specific enough to check: $0.22 per task on M3.1 Flash versus $0.51 per task on the M3 predecessor is a 57% reduction on that particular weighted-average metric, not quite the 70% figure MiniMax cites for overall agentic-task usage cost, which suggests the 70% number reflects a broader basket of workloads than the single per-task average. Both figures point the same direction, but reporters and readers comparing the two should be careful not to treat “70% cheaper” and “$0.22 versus $0.51” as the same claim measured the same way.
| Model | Input price (per 1M tokens) | Output price (per 1M tokens) | Weighted avg. cost per task | Access |
|---|---|---|---|---|
| MiniMax M3.1 Flash | $0.15 | $0.60 | $0.22 | MiniMax API, MiniMax Code |
| MiniMax M3 (predecessor) | Higher than M3.1 Flash | Higher than M3.1 Flash | $0.51 | MiniMax API |
| DeepSeek V4.1-Flash-1017 | Free (open-weight) | Free (open-weight) | Self-hosting cost only | MIT-licensed, downloadable |
That comparison also explains why MiniMax is emphasizing the “no long-context surcharge” detail in its pricing: at a 1-million-token context window, surcharges on large prompts can dominate a bill. Pricing a flash-tier model aggressively on both the token rate and the context-handling fee is MiniMax’s clearest lever for winning developer mindshare away from DeepSeek’s free, open-weight approach, since not every team wants to manage its own GPU infrastructure for self-hosting, even when the model weights are free.
How M3.1 Flash Stacks Up Against DeepSeek, Claude, and Gemini
MiniMax M3.1 Flash enters a market where the agentic-AI price war has already compressed margins for every lab with a speed-tier offering. DeepSeek’s V4.1-Flash-1017 update, covered in tech-insider’s report on its MIT license change, took the opposite pricing approach entirely: give the weights away for free and let the ecosystem build around them, monetizing through a search API instead of per-token fees. MiniMax, by contrast, is still running a metered API model, betting that aggressive per-token pricing combined with the no-surcharge context policy will be competitive enough without going fully open-weight.
Against the premium tier, the comparison is less direct. MiniMax’s VentureBeat-reported benchmark edge over Anthropic’s Opus 5 on a specific business benchmark is a genuine data point, but it is a single benchmark, not a general capability claim, and VentureBeat’s own framing paired that win with the hallucination caveat. Readers who want a broader three-way framing of where the premium chat-assistant tier stands can see tech-insider’s existing ChatGPT vs Claude vs Gemini comparison, which covers general-purpose assistant performance rather than the agentic/coding-specific niche MiniMax is targeting with M3.1 Flash.
It’s also worth placing M3.1 Flash next to MiniMax’s own prior comparison coverage. Before this release, tech-insider examined how the previous M3.1 generation stacked up in a three-way price comparison against GPT-6 Luna and DeepSeek V4.1 Flash, which found a roughly 3x price gap between the cheapest and most expensive options in that set. M3.1 Flash’s new pricing narrows that gap further on MiniMax’s side, reinforcing that the agentic-AI segment is converging toward a narrow band of per-token prices even as accuracy, not price, becomes the more interesting point of differentiation.
The Space Bunny Connection, Still Unconfirmed
One thread running through coverage of this release is a possible link to “Space Bunny,” a stealth model MiniMax quietly tested before unveiling it, which tech-insider covered separately in its report on the Space Bunny reveal after roughly 200 trillion tokens of testing. Unite.AI’s coverage of the M3.1 Flash launch raised the possibility that Space Bunny served as an early testing identity for what eventually shipped as M3.1 Flash, according to Unite.AI’s reporting. MiniMax has not confirmed that connection in its own release material, so it remains third-party speculation rather than an established fact, and readers should treat it accordingly until MiniMax addresses it directly.
Market Impact: China’s AI Price War Enters a New Phase
The M3.1 Flash release lands in the middle of an already crowded year for Chinese AI labs chasing the same developer audience: engineers building coding agents and automation tools who weigh latency and token price as heavily as raw benchmark scores. What’s shifting, based on this week’s coverage, is that the competitive story is no longer purely about who can cut prices fastest. MiniMax, DeepSeek, GLM, Qwen, and Kimi have all converged on similar price bands for comparable capability tiers. The next differentiator, if VentureBeat’s and The Outpost’s testing is representative of a broader pattern, is going to be measured reliability: which lab can hold its hallucination rate down while still competing on price and speed.
That shift has real implications for enterprise buyers. A model that’s 70% cheaper but requires an extra verification layer, additional human review, or a fallback to a pricier model for high-stakes outputs may not actually be cheaper once total cost of ownership is factored in. Procurement teams evaluating MiniMax M3.1 Flash against DeepSeek’s free weights or a premium alternative from Anthropic or Google will need to weigh accuracy risk explicitly rather than defaulting to the lowest per-token price, a dynamic that ground.news also highlighted in its summary of MiniMax’s claimed improvements, noting the gap between the company’s own benchmark framing and the scrutiny that followed.
Historical Context: MiniMax’s Rapid-Fire Release Cadence
MiniMax’s release cadence has been unusually fast even by 2026’s standards. The jump from M3 to M3.1 Flash came with a nearly 3x inference-speed improvement and a 70% usage-cost cut in a single generational step, according to MiniMax’s own characterization of the update as a “generational leap.” That pace mirrors the broader trend among Chinese AI labs of shipping incremental-sounding version bumps (M3 to M3.1, V4 to V4.1) that carry outsized practical improvements, rather than waiting for a full-number release to ship major gains. It also means developers evaluating these models face a moving target: a benchmark comparison published in September may already be stale by the time a procurement decision gets made in November.
Developer Access and the Cost-Math Behind M3.1 Flash
For teams trying to estimate real-world spend on M3.1 Flash, the published rate card makes the arithmetic straightforward. The snippet below illustrates the basic cost calculation using MiniMax’s published per-token pricing, useful for budgeting an agentic workload before committing to a migration.
# MiniMax M3.1 Flash cost estimate (published rates)
INPUT_PRICE_PER_M = 0.15 # USD per 1M input tokens
OUTPUT_PRICE_PER_M = 0.60 # USD per 1M output tokens
CACHED_INPUT_PRICE_PER_M = 0.015 # USD per 1M cached input tokens
def estimate_cost(input_tokens, output_tokens, cached_tokens=0):
fresh_input = input_tokens - cached_tokens
cost = (
(fresh_input / 1_000_000) * INPUT_PRICE_PER_M +
(cached_tokens / 1_000_000) * CACHED_INPUT_PRICE_PER_M +
(output_tokens / 1_000_000) * OUTPUT_PRICE_PER_M
)
return round(cost, 4)
# Example: a 50k-token agentic task with 10k cached context
print(estimate_cost(input_tokens=50_000, output_tokens=5_000, cached_tokens=10_000))
Running that math across a large batch of agentic tasks is roughly how MiniMax arrived at its $0.22 weighted-average figure, though actual task composition, how much of the context is cached, how long outputs run, will shift the real number for any given deployment.
What the Named Outlets Are Saying
MiniMax’s own framing of the release has been consistent across its official channels. As the company put it in its release announcement, “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%,” a statement that appears on MiniMax’s investor relations news page, cited earlier in this article.
The New Stack ran its own factual-accuracy test and reported that M3.1 Flash answered 56% of factual questions correctly, ahead of DeepSeek V4.1 Flash’s 43% on the same test set, a result that complicates the simple “MiniMax hallucinates more” framing by showing the model actually outperforming at least one rival on plain factual recall even as other testers flagged elevated hallucination on agentic tasks specifically. Marktechpost, covering the release from the API-availability angle, noted plainly that “MiniMax has just released MiniMax M3.1 Flash on its API, priced at $0.15 per 1M input tokens,” underscoring that for many developers the headline news is simply that the model is live and affordable, accuracy debates aside.
5 Predictions for MiniMax M3.1 Flash’s Next Few Months
- Hallucination benchmarks get standardized pressure. Expect more outlets to publish side-by-side hallucination testing methodology as the VentureBeat and The Outpost reports get cited repeatedly in developer forums.
- MiniMax publishes a rebuttal or clarification. Companies facing public accuracy criticism after a cost-focused launch typically respond within weeks with either a methodology dispute or a quiet model patch.
- Enterprise buyers segment their usage. Expect teams to route low-stakes, high-volume tasks to M3.1 Flash while keeping cybersecurity and finance agentic work on pricier, more thoroughly vetted models.
- DeepSeek and GLM respond with their own flash-tier updates. The pattern across 2026 has been rapid tit-for-tat releases whenever one Chinese lab claims a cost or speed lead.
- The Space Bunny question gets an official answer eventually. MiniMax has a track record of eventually confirming or denying stealth-model speculation once it becomes a recurring media talking point.
Frequently Asked Questions
What is MiniMax M3.1 Flash?
MiniMax M3.1 Flash is a speed-tier AI model released by MiniMax on October 19, 2026, built for agentic tasks, coding, Computer Use, and cybersecurity, with a 1-million-token context window.
How much does MiniMax M3.1 Flash cost?
MiniMax prices the model at $0.15 per 1 million input tokens, $0.60 per 1 million output tokens, and $0.015 per 1 million cached input tokens, with a weighted-average task cost of $0.22 according to the company’s own agentic-task comparison.
Does MiniMax M3.1 Flash hallucinate more than other models?
Independent testing from VentureBeat and The Outpost both reported elevated hallucination or accuracy trade-offs when comparing M3.1 Flash to models like DeepSeek and Anthropic’s Opus 5, even though the model performed competitively on cost and certain benchmarks. The New Stack’s separate factual-accuracy test, however, found M3.1 Flash outperforming DeepSeek V4.1 Flash on plain factual questions, so results vary by test type.
Is MiniMax M3.1 Flash faster than the previous M3 model?
Yes. MiniMax says M3.1 Flash delivers nearly 3x faster online inference than its predecessor, with typical output speeds above 200 tokens per second during peak usage.
How does MiniMax M3.1 Flash compare to DeepSeek V4.1-Flash-1017?
DeepSeek V4.1-Flash-1017 scored 77.0 on DeepSWE and is distributed as a free, MIT-licensed open-weight model, while MiniMax M3.1 Flash is a metered API product priced at $0.15/$0.60 per million input/output tokens. The two represent different monetization strategies within the same competitive agentic-AI segment.
Is Space Bunny the same model as MiniMax M3.1 Flash?
That connection has been raised by third-party outlets like Unite.AI but has not been confirmed by MiniMax, so it remains unverified speculation rather than an established fact.
Where can developers access MiniMax M3.1 Flash?
The model is available through the MiniMax API and through MiniMax Code, the company’s developer-facing coding agent tool.
What benchmarks did MiniMax use to evaluate M3.1 Flash?
MiniMax named Automation-Bench, DeepSWE, and Agents’ Last Exam (ALE) as evaluation benchmarks, and disclosed specific scores of 61.7 on OSWorld V2 and 76.2 on BankerToolBench.