MiniMax M3.1 Flash: Cheap Model That Beats Frontier on Benchmarks, But Hallucinates 93%

MiniMax releases M3.1 Flash at 0.15 dollars per million tokens with benchmark scores matching DeepSeek. But a 93% hallucination rate makes enterprises hesitate. Cheap doesn't mean safe.

MiniMax releases "Space Bunny" which turns out to be M3.1 Flash

In early October 2026, an anonymous model called "Space Bunny" appeared on OpenRouter and OpenCode. The model was free and immediately became a community favorite. Many speculated about who made it. On October 19, MiniMax confirmed: Space Bunny was an early version of M3.1 Flash.

M3.1 Flash is a mixture-of-experts (MoE) model with 428 billion total parameters, but only about 23 billion are active per inference. This means the model is "large" but runs like a small one. Its context window is 1 million tokens, and output speed can reach 200+ tokens per second.

Pricing: 0.15 dollars per million input tokens, 0.60 dollars per million output tokens. For comparison, GPT-6 Sol costs 2.00 dollars for input and 10.00 dollars for output per million tokens. MiniMax is 13x cheaper for input, 16x cheaper for output.

MiniMax claims agent costs drop 70% compared to the previous model. Artificial Analysis confirmed: cost per task dropped from 0.51 dollars to 0.22 dollars, nearly 57% cheaper. That's a significant number for companies running millions of requests per day.

Benchmark: sharp rise, but not frontier

On the Artificial Analysis Intelligence Index, M3.1 Flash scored 42, up 13 points from M3 (score 29). That's a big jump in one generation. But its position is still behind several frontier models.

DeepSeek V4.1 Flash-1017 scored 45, GLM-5.3 Flash also 42 (tied with M3.1 Flash), Gemini 3.8 Flash at 41, and GPT-6 Sol significantly higher. So M3.1 Flash is not the smartest model, but it's one of the cheapest in its class.

On HLE (Humanity's Last Exam), M3.1 Flash reached 48%, up from 39%. That beats GLM-5.3 Flash (40%) and DeepSeek V4.1 Flash (43%), but still trails Gemini 3.8 Flash.

On CritPt (physics reasoning), M3.1 Flash rose from 4% to 22%. That beats GLM-5.3 Flash (15%) and Gemini 3.8 Flash (18%), but narrowly loses to DeepSeek (24%).

On coding, via DeepSWE v1.1, M3.1 Flash reached 73.7%. That's nearly identical to Claude Opus 5.5 (74%) and GPT-5.6 Sol (73%). On OSWorld 2.0 (computer use), it scored 61.7, below Claude Opus 5.5 (70.6) and GPT-5.6 Sol (65.7).

In short: M3.1 Flash is not the smartest model on the market, but the cheapest model that is "smart enough" for many practical tasks. Think of it like a compact car: not the fastest, but the most fuel-efficient and good enough for daily driving.

The big problem: 93% hallucination rate

This is what makes enterprises hesitate. On the AA-Omniscience benchmark, M3.1 Flash answered 56% of factual questions correctly. That's decent, above average. But when the model didn't know the answer, 93% of the time it still gave a wrong answer rather than admitting uncertainty.

In other words: M3.1 Flash doesn't "know that it doesn't know." It prefers to fabricate a wrong answer rather than say "I'm not sure." This is different from models like Claude or GPT, which tend to say "I don't have enough information" when uncertain.

For enterprises using AI agents to update records, analyze financial documents, or research customers, this is a serious problem. A smart model that doesn't know its own limits can be more dangerous than a less smart but honest model. Imagine an agent fabricating numbers in a financial report because it's "too confident."

VentureBeat calls this a "high hallucination rate" and recommends additional verification, retrieval systems, guardrails, or human review. MiniMax itself acknowledges that M3.1 Flash "overwhelmingly chose to give an incorrect one rather than acknowledge that it did not know."

Trade-off: cheap vs reliable

M3.1 Flash represents a new trend in AI: cheap models that are "good enough" for practical tasks. At 13x cheaper than GPT-6 Sol for input and 16x cheaper for output, the savings can be millions of dollars per month for workloads with millions of requests per day.

But cheap isn't everything. Enterprises that need high accuracy (legal, financial, medical) will still choose more expensive but more reliable models. M3.1 Flash fits tasks with high error tolerance: drafting, brainstorming, simple coding, basic customer service, or exploratory data analysis.

MiniMax also plans to release a more capable Pro model, but no date has been announced. M3.1 Flash weights will be released October 23, but licensing terms haven't been announced. Whether it's open-source or proprietary is still unclear.

Perbandingan langsung dengan kompetitor terdekat

Untuk memahami posisi M3.1 Flash, mari bandingkan dengan model yang paling dekat harganya.

M3.1 Flash dan DeepSeek V4.1 Flash harganya hampir sama (0.15-0.30 dolar input, 0.60-1.20 dolar output), tapi DeepSeek skornya 3 poin lebih tinggi di Intelligence Index. Artinya, untuk harga yang sama, DeepSeek sedikit lebih pintar.

Tapi M3.1 Flash punya keunggulan di kecepatan: 200+ TPS dibanding DeepSeek yang lebih lambat. Untuk workload yang butuh respons cepat (chatbot, customer service), kecepatan bisa lebih penting daripada 3 poin benchmark.

GLM-5.3 Flash harganya 9x lebih mahal dari M3.1 Flash tapi skornya sama (42). Gemini 3.8 Flash 5-10x lebih mahal dengan skor lebih rendah (41). Jadi dalam kelas "murah tapi cukup pintar," M3.1 Flash dan DeepSeek V4.1 Flash adalah dua pilihan teratas.

Mekanisme: bagaimana MoE membuat model besar jadi murah

Mixture-of-Experts (MoE) adalah arsitektur yang membuat model besar bisa berjalan seperti model kecil. Cara kerjanya: model punya banyak "expert" (sub-network), tapi setiap token hanya diarahkan ke beberapa expert saja, bukan semua.

M3.1 Flash punya 428 miliar parameter total, tapi per inference hanya sekitar 23 miliar yang aktif. Itu artinya sebagian besar parameter "tidur" saat inference. Hasilnya: model besar dengan biaya inference model kecil.

MiniMax juga menggunakan "sparse attention" yang mengurangi komputasi yang diperlukan untuk generate respons. Ditambah dengan optimasi inference yang membuat output speed mencapai 200+ TPS.

Ini bukan teknologi baru. GPT-4 juga MoE. Tapi MiniMax berhasil membuatnya jauh lebih murah dengan sparsity yang lebih agresif. Trade-off-nya: sparsity lebih tinggi berarti akurasi bisa turun, terutama untuk tugas yang butuh penalaran panjang.

Dampak bagi developer dan enterprise

Bagi developer individual, M3.1 Flash adalah kabar baik. Harga 0.15 dolar per juta input token berarti prototyping dan eksperimen jauh lebih murah. Developer bisa mencoba berbagai pendekatan tanpa khawatir biaya.

Bagi enterprise, situasinya lebih nuanced. Harga murah memungkinkan scaling: aplikasi yang sebelumnya terlalu mahal untuk di-deploy (karena biaya per request) sekarang bisa jalan. Tapi hallucination rate 93 persen berarti enterprise harus menambah layer verifikasi.

MiniMax juga menawarkan M Plan (subscription) dan MiniMax Code (desktop app), selain API. Ini berarti mereka menargetkan bukan hanya developer yang pakai API, tapi juga end-user yang pakai aplikasi langsung.

Konteks geopolitik: MiniMax dan DeepSeek

MiniMax adalah perusahaan China, sama seperti DeepSeek. Ini punya implikasi: data yang dikirim ke API MiniMax bisa disimpan di server China (sama seperti kekhawatiran terhadap DeepSeek), pemerintah AS bisa membatasi penggunaan model China di agensi federal, dan enterprise yang sensitif terhadap data mungkin enggan pakai model China.

Di sisi lain, open-weight model dari China (DeepSeek, Qwen, MiniMax) mendorong kompetisi harga. Tanpa mereka, OpenAI dan Anthropic bisa menaikkan harga seenaknya. Jadi meski ada kekhawatiran geopolitik, kehadiran model China secara keseluruhan menurunkan biaya AI untuk semua orang.

Analysis: what this means for the AI industry

M3.1 Flash is not just another cheap model. It is a signal that the AI industry is shifting from who is the smartest to who is cheap enough and smart enough. That is a big shift.

For the past two years, AI competition was dominated by benchmarks: who scored highest on HLE, who won on coding evals, who was best at reasoning. But now, with M3.1 Flash and DeepSeek V4.1 Flash, there are models that are smart enough at 10-15x cheaper than frontier.

The implication: applications that were previously economically unfeasible can now run. Customer service chatbots with millions of requests per day. Agents analyzing thousands of documents. Coding assistants affordable for small startups.

But there is a hidden cost: 93% hallucination rate means every output needs verification. For applications with high error tolerance (drafting, brainstorming), this is not a problem. For applications requiring accuracy (finance, medical, legal), it is a big problem.

The unanswered question: will cheap models get smarter without losing honesty? Or is the cheap but hallucinates trade-off a fundamental limit of MoE architecture with high sparsity? If the latter, then enterprise will always need expensive models for critical tasks, and cheap models only for non-critical ones.

What is clear: the AI industry is shifting from one model for everything to the right model for the right task. M3.1 Flash is proof that cheap models have their own place, as long as they are used correctly.

Kesimpulan: murah bukan berarti aman

M3.1 Flash adalah contoh sempurna dari trade-off yang dihadapi industri AI: harga turun, capability naik, tapi reliability belum ikut naik. Model ini murah, cepat, dan cukup pintar untuk banyak tugas. Tapi hallucination rate 93 persen bikin ia tidak cocok untuk tugas yang butuh kepercayaan tinggi.

Pertanyaan bagi enterprise bukan "model ini cukup pintar?" tapi "model ini cukup jujur?" Untuk M3.1 Flash, jawabannya: belum. Tapi dengan guardrail dan verifikasi yang tepat, ia bisa jadi pilihan yang sangat cost-effective untuk tugas-tugas yang toleransi error-nya tinggi.

Apa yang perlu diperhatikan

• M3.1 Flash cocok untuk agentic tasks yang toleransi error tinggi
• Untuk tugas yang butuh akurasi (keuangan, hukum, medis), tambahkan verifikasi
• Harga 0.15 dolar per juta input sangat murah, tapi kecepatan 170+ TPS median berarti 3 menit per pertanyaan kompleks
• Weights belum dirilis, lisensi belum jelas
• MiniMax adalah perusahaan China, ada pertimbangan geopolitik (sama seperti DeepSeek)
• Model mendukung text dan image input, text output, context window 1 juta token
• Output maksimal 524.288 token per request