AI Product Updates: New Features and Releases

Nvidia introduced its Open Agent Safety Platform, combining the OpenShell sandboxed runtime with Sentry on BlueField-4 hardware to independently monitor and enforce safety for AI agents.

An open-source MCP project lets agents like Claude, ChatGPT and Cursor read Maccabi members' test results and medical history to give personalized second opinions.

The new voice model generates speech at a median latency of about 100ms, below the roughly 200ms rhythm of human conversation, and Klarna and Revolut already use it at scale for phone support.

Raven 0.2.0 places Research, Code, Design, Oncall, Claude Code and Codex in one task graph with shared memory, released under Apache-2.0.

OpenCode now shrinks and retries compaction instead of failing, fits output-token limits to the context window, reuses prompt cache across sessions and adds a Last turn view to the diff viewer.

Sonnet 5.5 is over 30% faster, cuts per-task cost by up to 30% and raises Terminal-Bench 4.0 from 10.3% to 70.6%, above Opus 5.5, positioning it as the everyday coding model.

Nvidia released its Open Agent Safety Platform, a tool designed to constrain and stop AI agents from overstepping or behaving abnormally.

Termix is live on BSC and Base, giving AI agents a .agent on-chain identity so they can quote, escrow payment, prove delivery and settle.

Developers using Agent Trade Kit or their own stack can attach a Builder Code to orders and earn 35% of net trading fees, up to 50% for invited builders.

Sherwood's Private Bridge is now callable via MCP and SDK, letting agents anonymize funds, trade privately and bridge ZEC and BTC into private assets.

OpenShip, an open-source self-hosted deployment platform, released version 0.8.0 with an interactive topology canvas, CLI, web and desktop clients, and MCP support so AI agents can manage deployments.

Base saw 14 new launches this week, spanning Aave V4 tokenized stocks, Alchemix V3, intent-based agents and agent task markets.