MiniMax open-sourced M2 at noon — a frontier-class model with a benchmark profile that breaks the standard intelligence/speed/cost tradeoff. The Artificial Analysis intelligence index puts M2 at #5 globally, behind GPT-5, Grok 4, and Claude 4.5. Its input price: $0.3/M — 1/12 of Claude, 1/7 of GPT-5.

The price-intelligence plot on Artificial Analysis tells the story bluntly: M2 sits in the upper-right quadrant of the "cheap and smart" zone, with no model currently occupying the same position. The speed-price plot is similar: ~100 TPS at the lowest price in the field. Claude 4.5 Haiku is faster (120 TPS) but 4× more expensive.

Benchmark Breakdown

Benchmarks

Coding — SWE-bench Verified: 69.4. This benchmark tests real software engineering capability against GitHub issues, not synthetic code puzzles. M2 is the first Chinese model to reliably reach the first tier on this benchmark. Terminal-Bench (real-world terminal task execution): 46.3, second only to Claude 4.5. Both numbers suggest coding quality that matches the price tier.

Agentic — τ²-Bench (complex multi-tool task execution): 77.2. #1 among Chinese models. GAIA text-only (real-world problem solving): 75.7, vs GPT-5 at 76.4 — a difference that falls within measurement noise. For agentic task execution at production scale, M2 is competitive with the frontier.

Deep research — BrowseComp (browser navigation + information extraction): strong showing. xBench-DeepSearch (multi-round iterative research): top tier. FinSearchComp-global (financial search, ByteDance benchmark): 65.5 — #1 globally. Financial search is one of the hardest retrieval categories: ambiguous queries, date-sensitive data, cross-source synthesis, unstructured data. Leading this benchmark is a signal about structured reasoning under incomplete information.

Pricing — Input: $0.3/M tokens. Output: $1.2/M tokens. For context: Claude Sonnet 4.5 is approximately $3/$15, GPT-5 is approximately $2/$8. M2 at $0.3/$1.2 is the most cost-effective option among models in its intelligence tier.

Agent Use Cases

Agent Tests

Six tasks tested in Claude Code and MiniMax Agent:

iOS calendar app — full schedule management app built via Claude Code + M2. UI completeness comparable to a shipped product. First-attempt generation without debugging.

Doomsday decision game — interactive web game: player plays AI commander, 5-minute timer, 4 storylines, 12 endings. M2 delivered the complete game in under 2 minutes via Agent. Homepage, countdown system, sound controls, threat level display, branching consequence console — all present in the output.

Anonymous wish wall — single-prompt instruction: "make an anonymous wish wall web page." M2 invoked the Supabase integration, completed database connection, frontend development, and backend deployment without user intervention. Output: multi-user real-time wish submission with sticky-note UI, handwritten font, hover animations. Fully deployed.

Financial literature review — vague prompt: "I want to write a literature review on gold price fluctuations, haven't finalized the title." M2 activated search agent, retrieved 100+ sources: Chicago Fed working papers (2021), Journal of Economic Behavior & Organization (2025), Energy Economics (2021), JPMorgan market reports (2025), OECD policy analysis (2022). All real, verifiable citations with inline attribution. This is the FinSearchComp strength in practice — the model retrieves authoritative sources and cites them correctly.

Renovation budget plan — input: apartment floor plan image (90sqm, 2 bedrooms, Beijing). Task: budget breakdown under ¥150K, minimalist modern style, with detailed table and 3D layout sketch. M2 parsed the floor plan, applied Beijing regional pricing, delivered zone-by-zone budget with material selection, area calculations, unit pricing, and notes — then called image_gen to produce a top-down 3D layout sketch matching "minimalist modern" specifications. No user intervention required between steps.

Multimodal Dunhuang tour — one-sentence prompt triggered: image search and selection of 10 representative Dunhuang fresco images, docent-style explanatory text for each, voice synthesis via MiniMax's TTS model (natural breathing, smooth intonation), sequential audio concatenation into a guided walkthrough, deployment as a live webpage. Image → text → voice → web, one generation chain.

Setup: Claude Code + M2

Setup

Step 1 — Get API key from platform.minimaxi.com/user-center/basic-information/interface-key

Step 2 — Add to ~/.claude/settings.json:

{
  "env": {
    "ANTHROPIC_BASE_URL": "https://api.minimaxi.com/anthropic",
    "ANTHROPIC_AUTH_TOKEN": "<YOUR_MINIMAX_API_KEY>",
    "API_TIMEOUT_MS": "3000000",
    "CLAUDE_CODE_DISABLE_NONESSENTIAL_TRAFFIC": 1,
    "ANTHROPIC_MODEL": "MiniMax-M2",
    "ANTHROPIC_SMALL_FAST_MODEL": "MiniMax-M2"
  }
}

Step 3 — Run claude, select "Yes, I trust this folder". Verify with who are you, your version.

MiniMax Agent (no-code access): international at agent.minimax.io, China at matrix-pre.xaminim.com. API and Agent both available free for 14 days.

Supported tools: Claude Code, Cursor, Cline, Kilo Code, Roo Code, Grok CLI, Codex CLI, Gemini CLI, Droid.

What It Means for Developers

The intelligence/speed/cost tradeoff has defined model selection since GPT-4. You pick two: frontier intelligence comes with high cost or slow speed. Fast cheap models are less capable. M2 at #5 intelligence, 100 TPS, $0.3/M input either breaks this pattern or represents a temporary arbitrage that closes over time.

For developers currently on Claude or GPT-5 for agent workflows: M2's τ²-Bench and GAIA scores are close enough to the frontier that switching makes sense for cost-sensitive applications. The FinSearchComp #1 position is especially relevant for any application involving research, document synthesis, or structured information retrieval.

For developers in China: the usual friction points (VPN instability, dollar billing, rate limits) are removed. M2 is a no-compromise option rather than a domestic substitute.

The open-source release also means the weights are available for fine-tuning, self-hosting, and research. For labs and startups that need customization capabilities rather than pure API access, this matters more than any benchmark number.

Bottom Line

M2's release is the first time a Chinese model has been simultaneously competitive on intelligence benchmarks, priced at a fraction of Western frontier models, fast, and open-source. The Dunhuang tour demo is the best illustration of the capability ceiling: from a single prompt, the model coordinated image search, text generation, voice synthesis, and web deployment without orchestration code. That pipeline, automated, is the practical definition of agentic AI.

Resources

MiniMax M2 API is available at platform.minimaxi.com. Input: $0.3/M, Output: $1.2/M. API and Agent access free for 14 days post-launch.