MiniMax Code Hands-On: A Personality-Driven Agent That Beats DeepSeek on Cost

Cost Analysis of MiniMax Code Plans

The Go plan costs 49 yuan/month (~$7), Explore 119 yuan (3× Go quota), Build 469 yuan (12×). All tiers include M3.1 Flash, image/audio models, web search, multimodal understanding, and cross-platform access (web, desktop, TUI, iOS, Android) plus API key integration. Video model H3 is exclusive to Explore and Build; Build adds peak-hour priority. Quotas are managed via a 5-hour rolling window and a 7-day weekly window; both must have headroom for text models. Unused quota does not roll over. Exceeding the 5-hour window switches to credit-based billing until the window refreshes.

Official reference: one 5-hour window supports >40M tokens; using one window daily for 30 days yields ~1.2B tokens/month (theoretical). In two test runs the author consumed a significant portion of the 5-hour window; a fourth test triggered the limit. Recommendations: occasional users should rely on daily sign-in credits (400 points/day, 1000 on days 4 and 7, 4000/week, valid 30 days); daily users with another primary agent can use Go as a cost-effective sidekick; heavy primary users or those needing H3 video should upgrade to Explore (3× quota).

API pricing for M3.1 Flash: input 1 yuan, output 4 yuan, cache read 0.1 yuan per million tokens. Agent tasks are ~98% cache reads. The Go plan's 1.2B-token equivalent would cost ~159 yuan via API, while DeepSeek off-peak ~57 yuan and peak ~115 yuan. However, the 49-yuan Go subscription undercuts both DeepSeek rates for users who saturate the 5-hour window daily.

MiniMax Code Features and Security Hardening

MiniMax Code is based on Pi v0.79.1, inheriting its loop, message scheduling, state transitions, terminal toolchain, and multi-vendor API abstraction. Key improvements over upstream Pi:

Large file reading : Pi's read tool hard-coded a 2000-line/50KB truncation; MiniMax lifts this limit for code-execution scenarios.

Streaming fault tolerance : Pi treated mid-stream truncation as normal completion; MiniMax detects truncation and retries.

Long-number precision : Pi rounded 18-digit IDs (e.g., Chinese ID numbers, snowflake IDs) when models emitted them as numbers; MiniMax patched this locally.

Process guarding : Bash stops send SIGTERM first, then force-kill; parent exit guards child process tree; Windows taskkill hardened.

Security boundaries are tightened compared to Pi's known risks (no per-tool approval, no workspace sandbox, /trust bypass, plugin privilege escalation, session export leakage):

Sandbox & permission tiers : Integrated independent sandbox-runtime; read-only Plan mode; multi-level permissions; high-risk or external operations trigger human approval; tool subprocesses use short-lived tokens.

Plugin & privacy controls : Plugins restricted to official marketplace and controlled local directories; arbitrary third-party imports blocked; local config files written with 0600 permissions; usage and diagnostic telemetry disabled by default.

Agent Personality: Proactive, Stubborn, Transparent

The agent exhibits a distinct "personality": proactive and slightly stubborn. When asked for M3.1 Flash technical blogs, it not only retrieved official blog, integration guide, model card, pricing, and vLLM day-0 inference benchmarks, but also auto-generated a Chinese deep-read visualization page. When the user interrupted requesting only URLs, the agent hesitated between obeying and completing the visualization, ultimately finishing the page — its reasoning chain is fully exposed.

Multi-Device Collaboration

Desktop client merges local and cloud session lists. A half-finished local task can be handed off to the cloud with summaries, key dialogues, and files; upload is confirmed per file. Tasks continue running after the laptop lid is closed — valuable for long-running jobs.

M3.1 Flash Model Capabilities

The model (early version known as "Space Bunny" on OpenRouter, 10-day anonymous leaderboard streak, 56.7T tokens processed) posts strong benchmarks: DeepSWE v1.1 73.7%; OSWorld V2 61.7%; Agents' Last Exam 29.6% (both domestic first); CyberGym 82%. Inference speed reaches 200+ TPS (median 170+). High activation-to-weight ratio and dense effective compute per token make it cost-effective for private deployment.

Demo Comparison: 3D Teardown (/teardown Skill)

Task: create a 20-second teardown video explaining why claw machines rarely grab prizes. Compared MiniMax Code + M3.1 Flash against DSH standard mode + DeepSeek V4.1 Flash (same "tool + model" pairing as prior article). MiniMax result slightly better than DeepSeek, though both had imperfections.

Price Comparison: Subscription vs. Pay-As-You-Go

Pure API : DeepSeek cheaper due to cache pricing (off-peak 0.02, peak 0.04 yuan/M tokens vs. MiniMax 0.10). Claw-machine task: DSH off-peak ~2.2 yuan, peak ~4.5 yuan; MiniMax ~5.3 yuan.

With subscription : MiniMax Go's 1.2B-token equivalent costs 49 yuan vs. DeepSeek off-peak ~57 yuan, peak ~115 yuan, MiniMax API ~159 yuan. Go wins for daily users who fill the 5-hour window; occasional users stay flexible with pay-as-you-go.

Additional Demos

Article-to-narrated video : Pure code pipeline calling multiple third-party services (long dependency chain; any parameter/path/timeline mismatch breaks the flow). M3.1 Flash completed it reliably — a task the author previously reserved for GPT.

Three.js procedural yarn farm : Recreated a Happy Farm scene in yarn-world style using zero textures or model files; houses, sheep, trees generated as geometry with simulated yarn texture and fuzz.

Postscript

Author plans to test H3 video model (open-sourced, ~6.4M Hugging Face downloads, hundreds of derivatives) but Go plan lacks H3 access; may upgrade to Explore or Build later.