![MiniMax M3.1 Flash Hits 93% Hallucination Rate [2026]](https://shattered.io/wp-content/uploads/2026/10/minimax-m3-1-flash-93-percent-hallucination-rate-2026-gen.webp)
MiniMax rolled out M3.1 Flash on October 19, 2026, pitching the model as a cheaper, faster option for agentic work, coding and computer-use tasks. Independent benchmarking from Artificial Analysis, reported by VentureBeat, backs up part of that pitch: the model’s Intelligence Index score jumped to 42 from 29 for its predecessor, while the average cost of completing a benchmark task dropped from roughly $0.51 to about $0.22. The same testing, though, surfaced a problem MiniMax did not mention in its own announcement: a 93% hallucination rate on a benchmark that measures how often a model gives a confidently wrong answer instead of admitting it does not know.
That split result, strong task performance paired with a reliability red flag, is the real story behind MiniMax M3.1 Flash. It also lands at a moment when the market for cheap, fast “Flash”-tier AI models has turned into one of the most competitive corners of the industry, with DeepSeek’s V4.1 Flash, Google’s Gemini 3.8 Flash, Z.AI’s GLM-5.3 Flash and Meta’s Muse Spark all fighting for the same low-cost, high-volume workloads. This piece breaks down what MiniMax actually shipped, what independent testing found, how the pricing stacks up against rivals, and what the hallucination finding means for anyone planning to put the model into production.
Don't miss new tech stories on Google
Add Shattered once in the Google app and our stories show up in your news suggestions.
What MiniMax Actually Announced on October 19 #
MiniMax’s own release note was short and direct: “Today, we officially roll out MiniMax M3.1 Flash.” The company describes M3.1 Flash as a next-generation Flash model that delivers what it calls a generational leap in agentic tasks, coding, Computer Use and cybersecurity, while making online inference nearly three times faster and cutting usage costs by 70% compared with the prior M3 model. MiniMax says the model is now available through its first-party API, its M Plan subscription tier, and MiniMax Code, its desktop coding tool.
The company also confirmed something industry watchers had suspected for weeks: M3.1 Flash is the production version of “Space Bunny,” an anonymous model that had been quietly processing requests in community coding arenas and had logged more than 200 trillion tokens before MiniMax put a name on it, according to Unite.AI’s reporting. MiniMax never put out a formal press release for the preview; instead, as Startup Fortune reported, the company slipped an “M3.1-Flash-Preview” build directly into MiniMax Code on September 27, available to anyone already on a Token Plan subscription, weeks before the formal announcement. That stealth-launch pattern, test in the wild under a pseudonym, then confirm the identity once the model has a track record, has become a common move among Chinese AI labs looking to generate buzz without committing to a public benchmark fight on day one.
Starting October 19, MiniMax began automatically routing requests that previously went to its older M2.5 and M2.7 models over to M3.1 Flash for subscribers on its Token Plan, according to MiniMax’s own API documentation. The model ships with a 1-million-token context window and what MiniMax describes as tunable thinking depth, letting developers dial reasoning effort up or down depending on the task.
The Independent Benchmark Numbers Tell a More Complicated Story #
Vendor-reported benchmarks are one thing. Third-party verification is another, and that is where the M3.1 Flash story gets interesting. According to Artificial Analysis’s evaluation, as detailed in VentureBeat’s reporting, M3.1 Flash scores 42 on the Intelligence Index at its maximum reasoning setting, up from 29 for M3. The average cost of completing an Intelligence Index task fell from about $0.51 to roughly $0.22, a reduction of nearly 57%, Artificial Analysis found.
Against its closest rivals in the same evaluation, the picture is tight. GLM-5.3 Flash scores 42 at about $0.28 per task. DeepSeek V4.1 Flash-1017 scores 45 at roughly $0.26 per task. Gemini 3.8 Flash scores 41 but costs about $1.24 per task, far more than the other three. MiniMax’s model is not the top scorer in that group, but it is competitive on intelligence and, outside of DeepSeek, the cheapest to run per task.
On Humanity’s Last Exam, Artificial Analysis recorded M3.1 Flash at 48%, up from M3’s 39%, ahead of GLM-5.3 Flash’s 40% and DeepSeek V4.1 Flash-1017’s 43%, and roughly matching Gemini 3.8 Flash. On CritPt, a physics reasoning benchmark, MiniMax jumped from 4% to 22%, ahead of GLM-5.3 Flash’s 15% and Gemini 3.8 Flash’s 18%, though still behind DeepSeek’s 24%. Artificial Analysis also noted that M3.1 Flash used fewer output tokens during the Intelligence Index evaluation, about 189 million, compared with roughly 324 million for GLM-5.3 Flash and 334 million for Gemini 3.8 Flash. Since developers pay per token generated, including the intermediate reasoning steps a model produces before answering, a model that reaches similar results with fewer tokens can be cheaper in practice than its advertised per-token rate suggests.
MiniMax’s own benchmark claims, which the company has not framed as independently verified, show the model scoring 73.7% on DeepSWE v1.1, 61.7 on OSWorld V2 for computer-use tasks, and 29.6 on Agents’ Last Exam. VentureBeat’s reporting places those figures alongside rivals: Claude Opus 5 scores 74% on DeepSWE, GPT-5.6 Sol scores 73%, and DeepSeek V4.1 Flash-1017 scores 77%, meaning MiniMax lands in the middle of the pack on coding rather than leading it. On Automation-Bench v1.0.6, which tests AI agents on business workflows spanning sales, marketing, finance and HR tools, MiniMax reports a score of 62.2, narrowly ahead of DeepSeek V4.1 Flash-1017’s 60.5.
The 93% Hallucination Rate, Explained #
The headline-grabbing number from Artificial Analysis’s testing is not a benchmark score at all, it is a reliability metric. On the AA-Omniscience benchmark, M3.1 Flash correctly answered 56% of questions, ahead of DeepSeek V4.1 Flash-1017’s 43% and GLM-5.3 Flash’s 28%. That alone sounds like a win for MiniMax. But Artificial Analysis also found that M3.1 Flash attempted nearly every question it was asked, rather than declining to answer when it was unsure, producing what the benchmark classifies as a 93% hallucination rate.
That figure needs context, because it is easy to misread. It does not mean 93% of the model’s answers were wrong. The metric measures the share of non-correct responses that were confidently incorrect, as opposed to partial answers or outright refusals. In practice, when M3.1 Flash did not know the right answer, it chose to guess wrong far more often than it chose to say “I don’t know.” That combination, strong factual recall paired with an unwillingness to hedge, produced an overall AA-Omniscience score of just 16 despite the relatively high raw accuracy rate.
For engineering teams, the distinction matters more than the headline number suggests. A model that is wrong 93% of the time it fails to nail the exact answer is still a serious liability for any workflow where a confident, wrong output can cause downstream damage, financial reconciliation, compliance checks, customer-facing agents making claims about policy or pricing. Artificial Analysis’s framing suggests M3.1 Flash is well suited to tasks with clear success criteria and automated verification loops, like code execution or structured agent benchmarks, and less suited to open-ended knowledge retrieval without a human or a retrieval-augmented system double-checking its output.
Pricing: Where M3.1 Flash Actually Sits in the Market #
MiniMax priced M3.1 Flash at $0.15 per million input tokens and $0.60 per million output tokens, with cached input discounted to $0.015 per million tokens. At those rates, a workload that consumes one million uncached input tokens and produces one million output tokens costs $0.75 in raw model usage, before accounting for any other infrastructure costs. That is a notably aggressive rate card, but it is not the cheapest on the market, and it is not alone at that price point either. The table below, compiled from pricing data that VentureBeat sourced from provider rate cards and Artificial Analysis, shows where M3.1 Flash lands relative to a cross-section of current Flash-tier and frontier models.
| Model | Input ($/1M tokens) | Output ($/1M tokens) | Combined ($/1M+1M) | Provider |
|---|---|---|---|---|
| Muse Spark 1.2/1.3 Contributor | $0.10 | $0.20 | $0.30 | Meta | | MiMo-V2.6-Flash | $0.14 | $0.28 | $0.42 | Xiaomi | | GPT-6 Luna | $0.10 | $0.50 | $0.60 | OpenAI | | DeepSeek V4.1 Flash-1017 (off-peak) | $0.15 | $0.60 | $0.75 | DeepSeek | | MiniMax M3.1 Flash | $0.15 | $0.60 | $0.75 | MiniMax | | MiniMax M3 (predecessor) | $0.30 | $1.20 | $1.50 | MiniMax | | Gemini 3.8 Flash (through Dec. 31, 2026) | $0.75 | $3.75 | $4.50 | | | GLM-5.3 | $1.40 | $4.40 | $5.80 | Z.AI | | Claude Sonnet 5.5 | $2.00 | $10.00 | $12.00 | Anthropic | | Claude Opus 5.5 | $4.00 | $20.00 | $24.00 | Anthropic |
Two things stand out. First, MiniMax M3.1 Flash lands at exactly the same combined rate as DeepSeek V4.1 Flash-1017’s off-peak pricing, $0.75 per million tokens split 1:4 between input and output. The two companies are, in effect, pricing their flagship low-cost models identically, which suggests the rate has become something close to a market floor for a Flash-tier model with real agentic capability, rather than a MiniMax-specific discount. Second, M3.1 Flash costs roughly one-sixth of Gemini 3.8 Flash and one-sixteenth of Claude Sonnet 5.5 on a combined basis, which is the kind of gap that pushes cost-sensitive teams, particularly those running high-volume, low-margin workloads like customer support triage or data extraction, to at least evaluate switching.
Benchmark Scoreboard: M3.1 Flash vs. the Flash-Tier Field #
Pricing only tells half the story. The table below pulls together the Artificial Analysis-verified Intelligence Index results alongside MiniMax’s own reported coding and agent scores, to give a side-by-side view of where M3.1 Flash is genuinely competitive and where it trails.
| Model | AA Intelligence Index | Cost per Index Task | Humanity’s Last Exam | AA-Omniscience Score |
|---|---|---|---|---|
| MiniMax M3.1 Flash | 42 | $0.22 | 48% | 16 |
| MiniMax M3 (predecessor) | 29 | $0.51 | 39% | – |
| DeepSeek V4.1 Flash-1017 | 45 | $0.26 | 43% | – |
| GLM-5.3 Flash | 42 | $0.28 | 40% | – |
| Gemini 3.8 Flash | 41 | $1.24 | ~48% | – |
Read together, DeepSeek V4.1 Flash-1017 edges out M3.1 Flash on raw intelligence score, 45 to 42, at a slightly lower cost per task. GLM-5.3 Flash ties MiniMax on the index score but costs a bit more per task. Gemini 3.8 Flash is roughly on par in capability but multiple times more expensive per task. None of the four is a clear-cut winner across every column, which is precisely why the Flash-tier segment has turned into a genuine price-and-capability fight rather than a market any single lab has locked down. MiniMax’s pitch, as outlined in its original M3.1 Flash launch announcement, leaned heavily on the 70% cost reduction and 3x speed claim relative to its own predecessor; the Artificial Analysis numbers confirm the model improved substantially over M3, even if they don’t hand MiniMax outright leadership over its closest external rivals.
Open Weights Are Coming October 23 #
MiniMax is not keeping M3.1 Flash locked behind its API. According to VentureBeat’s reporting, the company plans to release the model’s weights on October 23, 2026, four days after the API and subscription rollout. That mirrors a pattern MiniMax and other Chinese AI labs, including DeepSeek, have followed throughout 2026: ship the hosted, commercial version first, then follow with an open-weight release that developers can self-host, fine-tune or run through third-party inference providers. The strategy serves two purposes at once. It captures early commercial usage and feedback through the hosted API, while the open-weight release builds developer mindshare and keeps the model inside benchmark leaderboards and research comparisons long after the initial news cycle fades.
An open-weight release also changes the competitive calculus. Once M3.1 Flash’s weights are public, independent researchers and smaller inference providers can test, and potentially dispute or confirm, the Artificial Analysis hallucination findings on their own infrastructure, rather than relying solely on MiniMax’s hosted endpoint. That kind of outside verification has become a standard part of the open-weight AI cycle this year, following similar scrutiny applied to releases from Mistral, Z.AI and DeepSeek.
Why the Flash-Tier Market Matters More Than the Frontier Race #
Most of the AI industry’s public attention goes to frontier models, the GPT-6, Claude Opus and Gemini 4 tier that chase the top of every leaderboard. But the Flash-tier segment, cheaper, faster, slightly less capable models built for high-volume production use, is where a much larger share of actual AI spending happens. Enterprises running millions of customer support tickets, document extractions or code-review passes a day do not need frontier-level reasoning on every call; they need a model that is fast, cheap and reliable enough for the specific, narrower task at hand. That is the market MiniMax, DeepSeek, Google and Z.AI are all fighting over with their respective Flash models, and it is a market where a few cents per million tokens, multiplied across billions of calls, determines who wins enterprise contracts.
Pricing in this tier has compressed sharply over the past year. As the pricing table above shows, Flash-tier combined rates now span from $0.30 to $0.75 per million tokens among the most aggressive players, down from the $1.50 to $5.50 range that was common for similarly capable models in late 2025. MiniMax’s own M3 model, launched at $0.30/$1.20, cost twice as much as M3.1 Flash on a combined basis. That kind of rapid price compression, visible across the broader market in pieces like shattered.io’s recent breakdown of the 20x pricing gap between Claude, ChatGPT and Gemini, is reshaping how enterprises architect AI systems: instead of picking one model for everything, many are now routing different task types to different models based on cost and required accuracy, a pattern sometimes called model cascading or LLM routing.
MiniMax’s Competitive Position Against Chinese Rivals #
MiniMax is one of several Chinese AI labs, alongside DeepSeek, Z.AI, Xiaomi’s MiMo team and Moonshot AI, competing for both domestic and international enterprise customers with aggressively priced models. The company’s M3.1 Flash launch comes just over a month after DeepSeek shipped its own V4.1 Flash-1017 update, which posted a Humanity’s Last Exam score of 42.6 and became one of the most widely adopted low-cost models among developers building agentic tools. The near-identical combined pricing between MiniMax M3.1 Flash and DeepSeek V4.1 Flash-1017’s off-peak rate, both landing at $0.75 per million tokens, is not a coincidence; it reflects how tightly the Chinese Flash-tier market has converged on a small handful of viable price points.
MiniMax has also supplied the underlying technology for region-specific deployments elsewhere. Saudi Arabia’s Humain previously built a 428-billion-parameter Arabic-language model on MiniMax’s architecture, a deal detailed in shattered.io’s earlier coverage of that partnership, underscoring that MiniMax’s commercial reach already extends beyond pure API sales into licensing its model architecture for sovereign and regional AI projects. That kind of B2B and government-facing business line gives MiniMax a revenue base that is somewhat insulated from the API price war, even as that price war intensifies among developers comparing raw per-token costs.
How M3.1 Flash Compares to Western Flash-Tier Models #
Against Western alternatives, the comparison is less about raw capability and more about cost and deployment flexibility. Gemini 3.8 Flash, as the pricing table shows, costs roughly six times more than M3.1 Flash on a combined basis while posting a similar Intelligence Index score, though Google’s pricing is set to rise further starting January 1, 2027, according to the rate data cited by VentureBeat. OpenAI’s GPT-6 Luna, at $0.10/$0.50 combined $0.60, undercuts MiniMax slightly on raw price but comes from a company whose flagship frontier models, like GPT-6 Astra, benchmarked against Claude Opus 5.5 on Terminal-Bench, sit at a completely different price tier.
Meta’s open Muse Spark models occupy an unusual middle position: the “Contributor” tier of Muse Spark 1.2/1.3 is the single cheapest option in the entire comparison at $0.30 combined, while the standard, non-contributor pricing for the same model family jumps to $5.50, a 17x spread based entirely on licensing terms rather than model capability. That split pricing structure, rewarding developers who contribute data or usage feedback back to Meta with sharply discounted rates, is a different monetization approach than the flat per-token pricing MiniMax, DeepSeek and Google all use.
Historical Context: MiniMax’s Model Release Cadence #
M3.1 Flash is not MiniMax’s first attempt at a low-cost, high-throughput model, and the company’s release pattern over the past year shows a consistent strategy of rapid iteration. TechFlow Post reported that MiniMax has described M3.1 Flash as costing roughly one-third of M3 to run, with inference speed increased threefold, figures consistent with MiniMax’s own 70% cost-reduction claim. MiniMax M3, the model M3.1 Flash replaces, launched at $0.30/$1.20 pricing and scored 29 on Artificial Analysis’s Intelligence Index. The jump to 42 on the same benchmark in roughly the time between M3 and M3.1 Flash represents one of the larger single-generation capability gains MiniMax has posted, according to the Artificial Analysis data cited by VentureBeat, even accounting for the hallucination-rate caveat.
The company’s decision to test the model under the “Space Bunny” pseudonym before naming it, rather than announcing it as a preview or beta under MiniMax’s own brand, also fits a broader industry trend. Several labs have used anonymous or codenamed deployments in community model arenas to gather real-world performance data and community sentiment before committing to a public launch, reducing the reputational risk if an early version underperforms. By the time MiniMax confirmed the Space Bunny connection on October 19, the model had already processed more than 200 trillion tokens in the wild, giving the company a much larger real-world testing base than a traditional internal-only beta would have provided.
What MiniMax Said, in Its Own Words #
MiniMax’s own launch post frames the release in characteristically confident terms: “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%.” The company’s investor relations announcement opens with the same direct line used across its marketing: “Today, we officially roll out MiniMax M3.1 Flash.” Both statements emphasize speed and cost, and neither mentions the hallucination-rate finding that Artificial Analysis later surfaced, which is not unusual, vendors rarely volunteer the metric that makes their own model look worst, but it does mean enterprises evaluating the model need to look past MiniMax’s own marketing copy to get the full reliability picture.
MiniMax also confirmed, in a post on its official account, that “M3.1 Flash is now available through the MiniMax API, M Plan, and MiniMax Code,” and separately described the earlier preview rollout: “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.” A later post tied the preview explicitly to the Space Bunny identity, stating: “MiniMax M3.1 Flash, aka Space Bunny, is now live and FREE in Command Code for limited time!”
Market Impact: What Enterprises Should Watch #
For enterprise buyers, the practical takeaway from the M3.1 Flash launch splits into two tracks. On cost and raw task throughput, the model is genuinely competitive, matching or beating several rivals on price while posting a meaningful generational improvement over its own predecessor. On reliability for open-ended factual tasks, the AA-Omniscience result is a legitimate caution flag that should push any team considering the model toward pairing it with retrieval-augmented generation, structured output validation, or a human-in-the-loop review step for anything resembling unsupervised knowledge work. The broader market impact is that MiniMax has reinforced, rather than broken, the current equilibrium in the Flash-tier segment. No single model in this comparison dominates every axis: DeepSeek wins on raw intelligence-per-dollar, MiniMax wins on self-reported agent benchmarks in a few categories, Gemini trails on price but ties on capability, and Meta’s Muse Spark undercuts everyone on price for contributors willing to share usage data. That fragmentation is good news for buyers, who can increasingly shop by specific task type rather than committing to a single vendor, and it puts continued pressure on every lab in this tier to keep cutting prices or improving reliability metrics like the one Artificial Analysis just exposed in MiniMax’s model.
Predictions: Where This Goes Next #
- MiniMax will likely face pressure to publish its own AA-Omniscience-style reliability metrics in future model cards, following the pattern where labs start disclosing the benchmarks that previously only showed up in third-party testing after public scrutiny.
- The October 23 open-weight release will almost certainly trigger a wave of independent re-testing of the hallucination finding, which could either confirm Artificial Analysis’s number or narrow it depending on how different inference setups handle the model’s confidence calibration.
- Expect DeepSeek, Z.AI and Google to continue cutting Flash-tier prices through early 2027, keeping the combined per-million-token rate for competitive Flash models in the $0.30 to $1.00 range rather than letting any single lab establish a durable price advantage.
- Enterprises building agentic workflows will increasingly adopt multi-model routing, sending verification-heavy tasks to higher-cost, lower-hallucination models and routing high-volume, verifiable tasks like code execution to cheaper Flash models such as M3.1 Flash.
- MiniMax’s B2B licensing business, exemplified by the Humain partnership in Saudi Arabia, is likely to grow as a share of revenue relative to direct API sales, since licensing deals are less exposed to the kind of per-token price war playing out in the table above.
Competitive Comparison Summary #
Pulling the threads together: MiniMax M3.1 Flash is a legitimate, independently verified improvement over its own predecessor, with a near-tripled Intelligence Index score relative to cost. It is not a clear leader over DeepSeek V4.1 Flash-1017, which edges it out on the same Artificial Analysis Intelligence Index at a slightly lower per-task cost. It is dramatically cheaper than Gemini 3.8 Flash and Claude’s Sonnet and Opus tiers, which puts it in direct competition for cost-sensitive, high-volume enterprise workloads. And it carries a reliability caveat, the 93% hallucination rate on AA-Omniscience, that MiniMax’s own marketing does not mention and that buyers should factor into any deployment decision involving unsupervised factual output.
Frequently Asked Questions #
What is MiniMax M3.1 Flash?
It is a Flash-tier AI model from Chinese AI company MiniMax, officially launched on October 19, 2026, designed for agentic tasks, coding, computer-use automation and cybersecurity workflows at a lower cost and higher speed than MiniMax’s prior M3 model.
How much does MiniMax M3.1 Flash cost?
MiniMax prices the model at $0.15 per million input tokens and $0.60 per million output tokens, with cached input priced at $0.015 per million tokens. A workload using one million uncached input tokens and producing one million output tokens costs $0.75 in model usage.
Is MiniMax M3.1 Flash the same model as Space Bunny?
Yes. MiniMax confirmed in its launch announcement that Space Bunny, an anonymous model that had processed over 200 trillion tokens in community testing arenas, was an early version of M3.1 Flash.
What is the 93% hallucination rate about?
According to Artificial Analysis’s testing, reported by VentureBeat, M3.1 Flash attempted nearly every question on the AA-Omniscience benchmark rather than declining to answer when uncertain. Of the answers it got wrong, 93% were confidently incorrect rather than hedged or refused, resulting in an overall AA-Omniscience score of just 16 despite a 56% raw correct-answer rate.
How does M3.1 Flash compare to DeepSeek V4.1 Flash-1017?
DeepSeek V4.1 Flash-1017 scores slightly higher on Artificial Analysis’s Intelligence Index, 45 versus 42, at a marginally lower cost per task, about $0.26 versus $0.22. The two models share identical combined pricing of $0.75 per million tokens at DeepSeek’s off-peak rate.
Will MiniMax release M3.1 Flash’s model weights?
Yes. According to VentureBeat, MiniMax plans to release the model’s open weights on October 23, 2026, four days after the API and subscription rollout.
What is MiniMax Code?
MiniMax Code is the company’s desktop coding application, where M3.1 Flash is available alongside the API and M Plan subscription tier. MiniMax tested an early preview build of the model inside MiniMax Code before the full public launch.
Is MiniMax M3.1 Flash good for enterprise use?
It depends on the task. For verifiable, structured workloads like code execution or automated agent tasks, independent benchmarks show solid performance at a low cost. For open-ended factual or knowledge-retrieval tasks without human review, the high hallucination rate on AA-Omniscience suggests additional guardrails, such as retrieval-augmented generation or verification layers, are advisable.