MiniMax doesn't name the previous model, and says the downloadable weights come later. Three more stories inside.
MiniMax, an AI lab, launched M3.1 Flash on Monday and says agent tasks cost 70% less than on its previous model, which it doesn't name. It scores 73.7% on the DeepSWE v1.1 coding benchmark by MiniMax's count, and it's the most-upvoted release on Hacker News since our last issue. Also today: DHS orders civilian agencies to remove DeepSeek and anything built on it; GitHub retires six Copilot models and adds an admin API for model access; and Inception says its Mercury 3 runs 4–8x faster than Claude Haiku 5.5 and GPT-6 Luna, at a launch price half its list.
In this issue
•
MiniMax launches M3.1 Flash, saying agent tasks cost 70% less than on its previous model
•
DHS orders civilian agencies to remove DeepSeek, and anything built on it
•
GitHub retires six Copilot models, adds an API for admins to manage replacements
•
Inception says Mercury 3 runs 4–8x faster than Claude Haiku 5.5 and GPT-6 Luna
What changed this week #
MiniMax launches M3.1 Flash, saying agent tasks cost 70% less than on its previous model
MiniMax, an AI lab, rolled out M3.1 Flash on Monday, calling it a "generational leap" in agentic tasks, coding, Computer Use and cybersecurity. MiniMax says agent tasks cost 70% less to run than on its predecessor, which it doesn't name, and that inference runs nearly three times faster. It charges a flat $0.15 per million input tokens and $0.60 per million output tokens. On DeepSWE v1.1, a benchmark for independently understanding code, building features and fixing bugs, M3.1 Flash scores 73.7%, which MiniMax says puts it "alongside top models" without naming which ones. DeepSeek's own model card gave its V4.1 Flash update 77.0 on the same test Saturday. Both are the vendors' own numbers. One detail from the post: "Space Bunny," a stealth model that quietly processed over 200 trillion tokens online, was an early build of this one. MiniMax says the weights, the downloadable model files, come later.
DHS orders civilian agencies to remove DeepSeek, and anything built on it
The Department of Homeland Security has ordered every federal civilian agency to stop buying DeepSeek and remove its products, plus "third-party products and services developed using DeepSeek implementations," from its IT systems, citing China's legal reach over DeepSeek's own servers. That clause reaches past DeepSeek's app into anything built on its models. Signed in July and made public this week, per the AI-news site madrobot, the order excludes the military and intelligence agencies, with 60 days to request an exception.
GitHub retires six Copilot models, adds an API for admins to manage replacements
GitHub deprecated six Copilot models Monday across Copilot Chat, inline edits, agent mode and code completions: Gemini 3.7 Flash, GPT-5.5, GPT-5.4, GPT-5.4 mini, GPT-5 mini and Grok 4.5, each pointed to a replacement. "No action is required to remove the deprecated models," GitHub says. The same day, GitHub shipped a REST API letting admins manage Copilot's model access directly, not just by hand in settings. A separate wave hits Claude Opus 4.8, Opus 5 and GPT-5.6 Terra on November 16.
Inception says Mercury 3 runs 4–8x faster than Claude Haiku 5.5 and GPT-6 Luna
AI lab Inception says Mercury 3 runs 4–8x faster than Claude Haiku 5.5 and GPT-6 Luna. Inception's table puts Mercury ahead of Haiku on the SWE-Bench Atlas coding test, 51.08 to 32.26, and behind on DeepSWE v1, 46.31 to 51.62. Don't set either score against MiniMax's 73.7%, from v1.1. Output costs $0.35 per million tokens at half-off launch pricing, beating Haiku's $0.50 output price for prompts up to 100K. At the $0.70 list price, it doesn't.
Also worth your time #
- [DeepSeek open-sources DeepInfer, the inference framework behind its own production services](https://github.com/deepseek-ai/DeepInfer?ref=thenewway.ai)
- [Launch HN: Laminar, an RL-trained model that finds errors in agent traces](https://laminar.sh/blog/flow-1?ref=thenewway.ai)
Know someone who'd want this in their inbox? Forward it — that's how this grows. And if we got something wrong, or you think we buried the real story today, hit reply. A person reads every one.
The New Way is written with AI. It gathers the day's stories, checks them against their sources and drafts every summary. A person decides what runs and reviews every issue before we hit send.