Moonshot AI's 2.8-trillion-parameter open-weight model reached general availability in GitHub Copilot on August 6, hosted on Fireworks AI at $3 per million input tokens, per GitHub's changelog. At 2.8T parameters it is the largest open-weight model deployed in a mainstream coding tool to date. Read our take.

Cursor's per-request classifier now routes Auto Intelligence at 68% cost reduction, up from 60% two weeks ago, per Cursor's changelog. In early-access data, Balance mode ran $4.63 per commit versus $12.69 for Fable 5, per Cursor's data. Routing per turn beats standardizing on one model. See the cost data.

The macOS agentic browser shut down August 9; agentic browsing folds into ChatGPT Work plus a Chrome extension. OpenAI is consolidating on ChatGPT as one front door instead of a browser per surface. The cloud browser has no published DPA yet, so hold on routing sensitive tabs through it. Read our take.

A beta launched August 6 scans every third-party skill and plugin at upload and on each edit, returning a pass, warn, or fail verdict in one to two minutes before it reaches an employee session. It is off by default and set per org, and it skips MCP connectors and pre-existing plugins, so review your allow-list. Read our take.

Our verdict: Cursor stays the top AI editor for solo devs and small teams, but hold off on an annual Teams commitment until the SpaceX deal closes this quarter. Read the Cursor review or try Cursor.

GPAI obligations went enforceable August 2. This maps which duties fall on you versus your model provider, and the three controls that close most of the gap. Read the EU AI Act guide.

Every way you feed data to an agent creates a distinct exfiltration surface. This maps all three through a DLP lens so your security team knows where to place controls. Read the data boundaries guide.