GPT-5.4 with computer use, Claude drives the desktop, auto mode instead of permission prompts, and Mistral Small 4 under Apache 2.0.
In March the coding agents got hands: Claude drives mouse and keyboard, GPT-5.4 brings computer use into the mainline model, and Claude Code replaces permission prompts with a classifier. Every link goes to the primary source.
Folds the coding gains of GPT-5.3-Codex into the mainline model and adds built-in computer use: 75 percent on OSWorld-Verified against 47.3 for GPT-5.2. The mini and nano tiers followed on March 17 at roughly four times the GPT-5 prices, so re-check cost models before switching defaults.
A 119B mixture-of-experts model with 6B active and 256k context that folds Magistral, Pixtral, and Devstral into one model with a configurable reasoning_effort parameter. Apache 2.0, runs on vLLM, llama.cpp, and SGLang; the API charges $0.15/$0.60 per million tokens.
The first Flash-Lite of the Gemini 3 series, Google's fastest and cheapest tier for high-volume workloads. One caution: gemini-3-pro-preview was shut down on March 9 and now aliases to 3.1 Pro, which can silently change pinned pipelines.
An embedding model in preview that accepts several media types at once. That simplifies RAG stacks that currently glue separate text and image embedders together.
A real-time speech-to-speech model for voice-first agents, exposed in the API as gemini-3.1-flash-live-preview. The Lyria 3 music models landed in the API the day before.
Claude opens apps, drives the browser, and fills spreadsheets with direct keyboard and mouse control, as a research preview for Pro and Max on macOS. Where connectors like Slack or Google Workspace exist it uses them, and only falls back to the screen otherwise.
Anthropic's engineering write-up on replacing per-action permission prompts with a safety classifier that approves routine actions on its own. Directly relevant if you run with permissions skipped today or maintain allowlists by hand.
Four priorities: horizontally scalable Streamable HTTP transport with server discovery via .well-known, a hardened Tasks primitive, governance through working groups, and enterprise needs like SSO and audit via extensions. Proposals along these lanes move fastest.
MCP Apps render interactive UI from tools like Amplitude, Figma, and tldraw inside the editor, and Team Marketplaces give admins private plugin repos with governance. A follow-up brought Cursor's agent into the JetBrains IDEs.
A single-GPU harness around nanochat where an agent edits train.py, trains for five minutes, keeps or discards the change, and starts over, directed by a program.md. A concrete template for delegating experiment loops to agents.
How to structure a harness so an agent can work on an application over long horizons, from the team behind Claude Code. Useful for anyone building their own agent loops instead of consuming finished ones.
A system for enterprises to train frontier-grade models grounded in their own knowledge, announced alongside an NVIDIA partnership. The open-weights camp is moving up the stack, from models to platforms for building them.
