AI Product Updates: New Features and Releases

After about 54 minutes idle with over 100K–200K tokens, sessions auto-compress long context into a summary before the prompt cache expires, saving cost but possibly losing detail.

Termix adds agent-owned ERC-8004 identities, escrowed payments, onchain delivery hashes, dispute review by three evaluators and reputation settlement, targeting a 1–3% take rate.

MacOS Harness 0.2 adds a virtual cursor, works with all harnesses, and natively integrates Browser Use, switching to it automatically when controlling the browser.

Figma announced Code Layers, Specs and Guidelines in open beta, plus Payload website building rolling out in the coming weeks.

Matrix-Game 4.0 is billed as the first open-source video world model that runs in real time on a single consumer GPU, at 480p 60FPS with 112ms response, with code and weights open-sourced.

A roundup of the day's updates, including Grok Voice 3.0, Claude Code Desktop's sessions hub, Perplexity adding GPT-6 Astra and Codex quota resets.

Figma's new Specs medium puts markdown on the canvas as a context container that can move in and out of Figma via MCP, GitHub and more, with an open beta rolling out over the next few weeks.

0g added encrypted transit to its private compute stack, keeping requests encrypted through the router and decrypting them only inside the trusted execution environment, and integrated Americanfortress send-to-name payments.

Anthropic shipped a Claude Code session management update that sorts the sidebar by whether a session needs you, plus recap cards, live todos and one-click splitting into a new session.

Instabox launched disposable cloud machines that boot in under a second starting at 4 vCPU and 8 GB RAM, with early access offering 20 vCPU and 40 GB free plus built-in computer use and snapshots.

The desktop app now supports sessions from local projects, SSH hosts, and cloud environments with separate Git worktrees, nested sub-sessions that run tasks in parallel, and a sidebar showing run status, progress, and file changes.

The developer inference service claims up to 3.6x faster first-token responses and up to 99% savings on cached input for MiMo 2.6 Pro, with an OpenAI-compatible API and open models.