1 MiniMax ships M3.1 Flash at $0.15 per million input tokens
MiniMax says its new Flash model runs agent tasks nearly 3x faster and about 70% cheaper than its predecessor, with a 1M-token context. Pricing is flat: $0.15 per million input tokens and $0.60 per million output. Weights are promised later.
Why builders care: If you tried the free :free variant on OpenRouter, it ends today at 12:00 UTC (7 AM CT). Switch to the base slug minimax/minimax-m3.1-flash so nothing breaks.
2 Anthropic previews Claude Duet, an agent that comes to you first
Duet is a separate app (iOS, Android, claude.ai) that checks your connected apps every few hours, sends a morning brief, and keeps working on its own cloud computer after you close it. It asks for approval before it sends, pays, signs, or deletes anything. It's waitlist-only for US Pro and Max users.
Why builders care: Its background checks count against your plan's usage limits, which is a useful model for how to budget an always-on agent of your own.
3 Recursive's Flint: a coding model that runs on your laptop
Flint is a 40B mixture-of-experts model that uses about 2.5B parameters per token. Recursive reports 74.2% on SWE-bench Verified and more than 125 tokens/sec on an M4 Pro MacBook Pro. It's Apache 2.0 on Hugging Face and works with Ollama, LM Studio, llama.cpp and vLLM.
Why builders care: A private, offline coding agent with no usage bill. Compare it with the hosted options in our coding agent guide.
4 More than a third of Steam Next Fest games now use generative AI
TheGamer's count of SteamDB data found 2,595 of the 7,145 games in October's Next Fest (over 36%) carry Steam's AI tag, up from just over 19% in June. Players are pushing back, and many of the top-charting demos are AI-free.
Why builders care: Disclosed AI use is now normal on Steam, and players are watching. Being clear about how you used it can help you stand out.
5 Repo of the day: OPAL, an optimizer an agent rewrites mid-training
From MIT researchers: wrap your PyTorch optimizer in one line, and a coding agent (Claude Code by default) watches the run and rewrites the update rule, schedules, and loss-spike guards as it trains. MIT-licensed, with a cost cap you set.
Why builders care: For a new task with no tuned recipe, the authors report it matches or beats the best off-the-shelf recipe across their image and language tests.