4 new AI tools to try this week: Cursor Computer Use, Kimi K3.1 Pro, MiniMax M3.1 Flash and Theorem
Published· By Mishal Brands Editorial · ai-tools, ai-agents, coding
Cursor's agents can now drive desktop apps, Moonshot's Kimi K3.1 Pro brings a 1M-token context, MiniMax shipped open multimodal weights, and Theorem opened free proofs for open-source code.
Four launches from the past few days are worth your time if you build with AI. Here's what each one does, what the makers report, and how to try it.
1. Cursor Computer Use: agents that test the apps they build
Cursor says its agents can now "click and type across any app on your computer, not just the browser." Its demo shows an agent driving the iOS Simulator to test a feature it just built (Cursor announcement on X, Oct 23, 2026).
- How to use it: Cursor says the agent turns Computer Use on by itself when a task needs it, or you can @-mention "Computer Use" in your prompt. Update Cursor to try it.
- Self-hosted workers: Cursor's docs say you start a macOS or Linux worker with
--computer-use. On a Mac, its helper app needs Accessibility and Screen Recording permissions (Cursor docs). - Tip: close unrelated apps first. An agent that can see your screen can see whatever is open.
More on plans and pricing: Cursor tool page.
2. Kimi K3.1 Pro: a 1M-token model with open weights on the way
Moonshot AI introduced Kimi K3.1 Pro, with a 1-million-token context window, and a Highspeed version it says runs 2-3x faster. Both are available on Kimi.ai, Kimi Work, Kimi Code and the Kimi API (kimi-k3.1-pro). Moonshot says the full weights will be released by November 4 (Kimi blog, Oct 23, 2026).
Moonshot reports 65.7% on Terminal-Bench 4.0, narrowly behind Claude Opus 5.5 at 66.4%. These are vendor-reported numbers and haven't been independently verified.
3. MiniMax M3.1 Flash: open-weight multimodal model
MiniMax M3.1 Flash is a mixture-of-experts model for coding, agent workflows and computer use. MiniMax's model card lists about 428B total parameters (23B active per token), text, image and video inputs, and up to 1M tokens of context. The weights are on Hugging Face, with serving recipes for SGLang and vLLM (Hugging Face model card). MiniMax reports 73.7% on DeepSWE v1.1. Check the license before any commercial use.
4. Theorem: free formal proofs for open-source code, plus a Claude skill
Working with Theorem, Anthropic reports that Claude proved there are no crashes on any input up to 64 MB for 239 fuzz-test entry points across 123 open-source libraries. Anthropic says Theorem is offering to run the same pipeline for other open-source projects and is publishing a Claude skill to help developers get started (Anthropic, Oct 22, 2026). That skill, add-project, is in Theorem's oss-verify repository.
Anthropic is clear about the limits: the proofs cover single-threaded, static RISC-V binaries, and "formally verified" does not mean "bug-free."
Also this week
OpenAI says it now trains its frontier models with no live internet access. Read our breakdown: OpenAI now trains its frontier models without live internet access.
The bottom line
Agents are moving from writing code to testing it. Cursor lets them click through your apps, and Theorem uses them to prove code can't crash. If you only try one thing this week, update Cursor and ask an agent to test the last feature you shipped.
Tools in this post
- Cursor: Write, fix and review code faster in an AI code editor.
- Kimi K3.1 Pro: Moonshot AI’s flagship model with a 1M-token context, for coding, 3D and finance work.
- MiniMax M3.1 Flash: An open-weight multimodal model from MiniMax for coding, agents and computer use.
- Theorem (OSS Verify): Machine-checked proofs for open-source code, plus a Claude skill to get your project started.