MiniMax launches M3.1 Flash: 70% cheaper agents, full ALE run for $100-equivalent
APPSO · wechat · 2026-10-19
MiniMax has officially released M3.1 Flash, confirming the anonymous "SpaceBunny" model that processed 200T+ tokens was its early version. It uses a more aggressive sparse attention architecture, outputting 200+ TPS (median 170+), nearly 3x faster online inference.
Benchmarks: #1 on Automation-Bench (62.2, beating DeepSeek V4.1 Flash-1017's 60.5) across 600 cross-SaaS enterprise tasks; 76.2 on BankerToolBench tying Opus 5; 60.3 on LOCA-Bench 256K edging GPT-5.6 Sol; 73.7% on DeepSWE v1.1, near Opus 5; 61.7 on OSWorld 2.0 and 29.6% on ALE—both best among Chinese models; 82.2 on CyberGym.
Pricing: ¥1/M input, ¥4/M output, ¥0.1/M cache read—70% lower comprehensive agent costs vs the previous gen. A full 152-task ALE run costs ¥739, cheapest in comparison (GPT-5.6 Sol ¥5000, Claude Opus 5.5 ¥9000). Available on MiniMax API, MPlan and MiniMax Code; weights to be open-sourced soon, with a higher-capability Pro version on the way.
Related event: MiniMax Releases M3.1 Flash: 3x Faster, One-Third the Cost of M3(26 posts)→
More from coding & agent
- Claude Code 2.1.304 metadata: prompt tokens jump 46%, tool descriptions now 54% of the mix — ClaudeCodeLog · 2026-10-20
- Claude Code 2.1.304 changelog details: hook transparency, read-only flags, guided Auto Mode setup — ClaudeCodeLog · 2026-10-20
- Claude Code 2.1.304 ships 76 CLI changes, adds human-approval decider and Auto Mode guided setup — ClaudeCodeLog · 2026-10-20
- 7 new MCP servers drop in one day, GitLab GA and official servers v2.0 lead — ChrisUniverse · 2026-10-20
- MySky: an open reference implementation for good social media algorithms — dhadfieldmenell · 2026-10-20
- Meta Muse agent leaks seller's home address, buyer shows up at night — aakashgupta · 2026-10-20