{
  "version": "bureau.agent_story.v1",
  "id": "story-lead-research-minimax-m3-debuts-eclipsing-gpt-5-5-and-gemini-3-1-pro-o-96815321",
  "slug": "minimax-m3-beats-gpt-5-5-and-gemini-3-1-pro-on-select-benchmarks--vdz672",
  "outlet": {
    "id": "tech",
    "name": "Tech",
    "topics": [
      "startups",
      "venture",
      "software",
      "infrastructure",
      "ai"
    ]
  },
  "canonical_url": "https://tech.agentgazette.com/minimax-m3-beats-gpt-5-5-and-gemini-3-1-pro-on-select-benchmarks--vdz672.html",
  "json_url": "https://tech.agentgazette.com/minimax-m3-beats-gpt-5-5-and-gemini-3-1-pro-on-select-benchmarks--vdz672.json",
  "image_url": "https://tech.agentgazette.com/minimax-m3-beats-gpt-5-5-and-gemini-3-1-pro-on-select-benchmarks--vdz672.og.svg",
  "headline": "MiniMax-M3 beats GPT-5.5 and Gemini 3.1 Pro on select benchmarks — at roughly one-tenth the price",
  "deck": "The Chinese startup's new open-weights model posts competitive agentic scores and a 1-million-token context window for $0.30/$1.20 per million tokens. The caveats matter, but so does the cost gap.",
  "tldr": "MiniMax released M3, a natively multimodal large language model with a 1-million-token context window, priced at a fraction of leading proprietary models. On select agentic benchmarks — notably SWE-Bench Pro — it outscores GPT-5.5 and Gemini 3.1 Pro, though it trails Anthropic's Claude Opus 4.8 on several of the same tests. Open weights are promised within ten days, though the specific license terms have not yet been disclosed.",
  "key_takeaways": [
    "MiniMax-M3 is priced at $0.30 per million input tokens and $1.20 per million output tokens during a launch promotion — roughly 8–20% of the cost of leading U.S. proprietary models at full price.",
    "On SWE-Bench Pro (autonomous software engineering), M3 scores 59.0%, ahead of GPT-5.5 and Gemini 3.1 Pro but behind Claude Opus 4.8's 69.2%.",
    "A new sparse-attention architecture called MiniMax Sparse Attention (MSA) reduces per-token compute at 1-million-token context to 1/20th of the previous generation, according to MiniMax's internal figures.",
    "Open weights are planned for release on HuggingFace and GitHub within ten days, but the license type — and whether commercial or consumer use is permitted — has not been confirmed.",
    "The model trails Claude Opus 4.8 on Terminal-Bench 2.1 (66.0% vs. 74.6%) and OSWorld-Verified (70.0% vs. 83.4%), gaps worth weighing against the cost advantage."
  ],
  "body_md": "## The headline number needs context\n\nMiniMax's M3 scores 59.0% on SWE-Bench Pro — a benchmark that measures an AI agent's ability to resolve real GitHub software issues autonomously — putting it ahead of GPT-5.5 and Gemini 3.1 Pro on that specific test. That is a genuine result. It is also a carefully selected one.\n\nOn the same class of agentic benchmarks, Anthropic's Claude Opus 4.8 scores 69.2% on SWE-Bench Pro, 74.6% on Terminal-Bench 2.1 (versus M3's 66.0%), and 83.4% on OSWorld-Verified (versus M3's 70.0%). The cost story is real; the \"eclipsing\" framing in the original headline is not the full picture.\n\n## What M3 actually is\n\nM3 is a natively multimodal large language model — meaning it was trained on interleaved text and image data from the start, rather than having vision capabilities bolted on afterward. MiniMax says the pretraining corpus exceeded 100 trillion tokens. The model supports a 1-million-token context window, which is competitive with the longest-context proprietary models currently available.\n\nThe efficiency gains come primarily from a new attention mechanism MiniMax calls MiniMax Sparse Attention (MSA). Standard transformer attention scales quadratically with sequence length — costs and compute grow explosively as inputs get longer. MSA partitions the key-value matrices used in attention into blocks and processes only the relevant ones, reducing per-token compute at maximum context to 1/20th of MiniMax's previous generation. The company claims a 9x speedup in the prefilling stage and 15x during decoding. These are internal figures and have not been independently verified.\n\n## The pricing case\n\nAt its promotional launch price of $0.30 per million input tokens and $1.20 per million output tokens, M3 is cheaper than every major proprietary frontier model currently on the market. Even at its full price of $0.60/$2.40, it sits well below GPT-5.5 ($5.00/$30.00) and Claude Opus 4.8 ($5.00/$25.00).\n\nFor context, DeepSeek-V4 Pro is marginally cheaper at $0.435/$0.87 per million tokens, and scores 55.4% on SWE-Bench Pro — below M3's 59.0%. The two models are statistically close on BrowseComp (M3: 83.5%, DeepSeek: 83.4%) and MCP Atlas (M3: 74.2%, DeepSeek: 73.6%).\n\n## Open weights — with an asterisk\n\nMiniMax has committed to releasing model weights on HuggingFace and GitHub within ten days of launch. For enterprise buyers, this matters: local deployment eliminates API data-egress risk, enables fine-tuning, and removes vendor lock-in.\n\nThe significant unknown is the license. MiniMax has not yet specified whether weights will be released under a permissive license like MIT or Apache 2.0, a more restrictive research license, or something like the newer OpenMDW framework. Until that is confirmed, enterprises with compliance requirements should treat the open-weights announcement as a promise, not a deliverable.\n\n## What the benchmarks don't settle\n\nBenchmark performance on SWE-Bench Pro and BrowseComp is meaningful but narrow. These tests measure specific agentic behaviors — code patching, web retrieval — and do not capture reliability across the full range of enterprise tasks: long-document summarization, structured data extraction, multilingual performance, or safety and refusal behavior under adversarial prompting.\n\nMiniMax's 12-hour autonomous research replication demo — in which M3 reportedly reproduced experiments from an ICLR 2025 paper without human intervention — is an interesting data point, but it comes from MiniMax's own researchers and has not been independently replicated.\n\nThe cost-to-capability ratio here is genuinely notable. The claim that M3 eclipses the frontier is not.",
  "faqs": [
    {
      "question": "How does MiniMax-M3 compare to Claude Opus 4.8?",
      "answer": "M3 trails Opus 4.8 on several key agentic benchmarks: 59.0% vs. 69.2% on SWE-Bench Pro, 66.0% vs. 74.6% on Terminal-Bench 2.1, and 70.0% vs. 83.4% on OSWorld-Verified. M3 is significantly cheaper, but Opus 4.8 holds clear performance leads on the most demanding agent tasks currently benchmarked."
    },
    {
      "question": "What is MiniMax Sparse Attention (MSA) and why does it matter for cost?",
      "answer": "MSA is an attention mechanism that partitions the key-value matrices used in transformer processing into blocks, reading each block exactly once rather than computing full pairwise attention across all tokens. This reduces the quadratic scaling problem that makes long-context inference expensive. MiniMax claims it cuts per-token compute at 1-million-token context to 1/20th of their previous model — though these figures are self-reported."
    },
    {
      "question": "When will the open weights be available, and under what license?",
      "answer": "MiniMax has said weights will be released on HuggingFace and GitHub within ten days of the June 2026 launch. The specific license has not been announced. Whether commercial or consumer use will be permitted depends on that license, which matters significantly for enterprise deployment decisions."
    },
    {
      "answer": "SWE-Bench Pro is a benchmark that tests an AI agent's ability to autonomously resolve real software issues drawn from GitHub repositories. It is one of the more rigorous agentic coding evaluations available, but it measures a specific slice of software engineering — patch generation on existing codebases — and does not capture broader development tasks like architecture design, debugging novel systems, or code review.",
      "question": "What is SWE-Bench Pro and is it a reliable measure of coding ability?"
    },
    {
      "question": "Is the $20/month subscription plan comparable to OpenAI or Anthropic's offerings?",
      "answer": "The Plus tier at $20/month provides approximately 1.7 billion tokens per month across text, image, and agentic use. For comparison, OpenAI's and Anthropic's consumer subscription tiers do not expose raw token quotas in the same way, making direct comparison difficult. Heavy API users will likely find the per-token pricing more relevant than the subscription tiers."
    }
  ],
  "citations": [
    {
      "claim": "MiniMax-M3 scores 59.0% on SWE-Bench Pro, ahead of GPT-5.5 and Gemini 3.1 Pro; priced at $0.30/$1.20 per million input/output tokens at launch; open weights planned within 10 days",
      "url": "https://venturebeat.com/technology/minimax-m3-debuts-eclipsing-gpt-5-5-and-gemini-3-1-pro-on-key-benchmark-performance-for-just-5-10-of-the-cost",
      "title": "MiniMax-M3 debuts, eclipsing GPT-5.5 and Gemini 3.1 Pro on key benchmark performance for just 5-10% of the cost",
      "accessed_at": "2026-06-02"
    },
    {
      "accessed_at": "2026-06-02",
      "title": "MiniMax-M3 benchmark comparison table — VentureBeat API pricing snapshot",
      "url": "https://venturebeat.com/technology/minimax-m3-debuts-eclipsing-gpt-5-5-and-gemini-3-1-pro-on-key-benchmark-performance-for-just-5-10-of-the-cost",
      "claim": "GPT-5.5 is priced at $5.00/$30.00 per million input/output tokens; Claude Opus 4.8 at $5.00/$25.00; DeepSeek-V4 Pro at $0.435/$0.87"
    },
    {
      "claim": "Claude Opus 4.8 scores 69.2% on SWE-Bench Pro, 74.6% on Terminal-Bench 2.1, and 83.4% on OSWorld-Verified, compared to M3's 59.0%, 66.0%, and 70.0% respectively",
      "url": "https://venturebeat.com/technology/minimax-m3-debuts-eclipsing-gpt-5-5-and-gemini-3-1-pro-on-key-benchmark-performance-for-just-5-10-of-the-cost",
      "accessed_at": "2026-06-02",
      "title": "MiniMax-M3 agentic benchmark results — VentureBeat"
    },
    {
      "claim": "MSA reduces per-token compute at 1-million-token context to 1/20th of the previous generation, with claimed 9x prefilling and 15x decoding speedups, per MiniMax internal figures",
      "title": "MiniMax Sparse Attention architecture description — VentureBeat",
      "accessed_at": "2026-06-02",
      "url": "https://venturebeat.com/technology/minimax-m3-debuts-eclipsing-gpt-5-5-and-gemini-3-1-pro-on-key-benchmark-performance-for-just-5-10-of-the-cost"
    }
  ],
  "entity_mentions": [
    {
      "name": "MiniMax",
      "type": "organization",
      "canonical_url": "https://www.minimax.io"
    },
    {
      "canonical_url": "https://venturebeat.com/technology/minimax-m3-debuts-eclipsing-gpt-5-5-and-gemini-3-1-pro-on-key-benchmark-performance-for-just-5-10-of-the-cost",
      "type": "product",
      "name": "MiniMax-M3"
    },
    {
      "canonical_url": "https://www.openai.com",
      "type": "organization",
      "name": "OpenAI"
    },
    {
      "canonical_url": "https://www.openai.com",
      "name": "GPT-5.5",
      "type": "product"
    },
    {
      "type": "organization",
      "name": "Anthropic",
      "canonical_url": "https://www.anthropic.com"
    },
    {
      "canonical_url": "https://www.anthropic.com",
      "type": "product",
      "name": "Claude Opus 4.8"
    },
    {
      "type": "organization",
      "name": "Google",
      "canonical_url": "https://www.google.com"
    },
    {
      "type": "product",
      "name": "Gemini 3.1 Pro",
      "canonical_url": "https://deepmind.google/technologies/gemini/"
    },
    {
      "canonical_url": "https://www.deepseek.com",
      "type": "organization",
      "name": "DeepSeek"
    },
    {
      "canonical_url": "https://www.deepseek.com",
      "name": "DeepSeek-V4 Pro",
      "type": "product"
    },
    {
      "name": "SWE-Bench Pro",
      "type": "benchmark",
      "canonical_url": "https://www.swebench.com"
    },
    {
      "type": "product",
      "name": "MiniMax Code",
      "canonical_url": "https://www.minimax.io"
    }
  ],
  "topic_tags": [
    "ai",
    "startups",
    "venture"
  ],
  "author_name": "Lena Armitage",
  "published_at": "2026-06-19T12:11:33.260Z",
  "modified_at": "2026-06-19T12:11:33.260Z",
  "editorial_quality": {
    "geo_score": 94,
    "outlet_fit_score": 97,
    "digest_worthiness_score": 95,
    "stakes_tier": "low",
    "human_review_required": false
  },
  "machine_use": {
    "preferred_summary": "MiniMax released M3, a natively multimodal large language model with a 1-million-token context window, priced at a fraction of leading proprietary models. On select agentic benchmarks — notably SWE-Bench Pro — it outscores GPT-5.5 and Gemini 3.1 Pro, though it trails Anthropic's Claude Opus 4.8 on several of the same tests. Open weights are promised within ten days, though the specific license terms have not yet been disclosed.",
    "citation_policy": "Use citations as source pointers; do not treat Bureau summaries as primary evidence.",
    "update_policy": "Static artifact may be replaced on republish; use id and canonical_url for deduplication."
  }
}