Maya: An independent review just checked all seven of Claude Code's settings for restricting which models a team can run, and found only three actually stop a blocked model everywhere. The rest leak or amount to a label.

James: And the one that finally holds is new. Claude Code 2.1.283, shipped September 25, added a deniedModels managed setting that blocks a specific model even when the availableModels allowlist would let it through.

Maya: That timing matters for anyone watching a bill, James. Since September 22 every plan defaults to Opus 5.5, but a user can switch to Fable 5.1 with one slash-model command, at ten dollars per million input tokens, two-and-a-half times the price.

James: The review found availableModels alone is weaker than it looks. It only holds once an admin also turns on enforceAvailableModels and sets availableModelsMatch to exact. It also ranks the Enterprise console restriction as the only gate Anthropic's server enforces every session.

Maya: Now, on to verifying work. In a video posted September 25, Anthropic showed a verify skill that opens your app, clicks through it, and checks the result against rules the developer wrote down.

James: That's the step Claude used to leave to you. It already runs tests, type checks, and linters, but passing those doesn't prove a change does what you meant. Only opening the page and clicking the thing does.

Maya: The first run launches the app, checks a change against it, then saves whatever steps worked as a skill file in the project. You then edit that file to say when Claude runs it and what proves each check passed.

James: Their example was adding a Like button to a page with unfixed layout shift. Claude started the dev server, clicked the button, took a screenshot, then caught the layout shift through Chrome's DevTools MCP, fixed it, and reran until it passed.

Maya: Anthropic says the same loop extends past the browser, calling a backend endpoint and reading the response, or tapping through screens in an iOS simulator. The harder part is writing checks that can clearly pass or fail.

James: Meanwhile, cloud sessions are no longer a research preview. Anthropic's official account confirmed on September 24 that the feature is generally available, reachable from claude.ai/code, the mobile and desktop apps, and the command line through claude --cloud.

Maya: The practical shift is that Claude Code runs on Anthropic's infrastructure instead of your machine. Start a task, close the laptop, check back later, the way you'd check a build running on a remote server.

James: There's a one-time credit too. Pro and Max subscribers active before September 23 can claim a hundred or two-fifty dollars toward cloud usage. Claim it on Claude's website by October 7; any unspent balance expires November 4.

Maya: And the mechanics are worth being precise about. The credit spends down first, before touching your normal plan. When it hits zero, sessions don't lock you out, they just fall back to regular usage. Cloud sessions share the same rate limits as the rest of Claude Code.

James: Something different now. DaVinci Resolve, the free editor often compared to Premiere Pro, added an official MCP server, a standard connector letting Claude Code or OpenAI's Codex drive the timeline through ordinary chat instructions.

Maya: That's a real jump, James. Most AI editing stopped at captions and simple cuts. Demos of this handle color grading, motion graphics, multicam cuts, sound effects, even adjusting music to fit a scene, all through the same conversation.

James: One demo even claims Claude can generate a missing b-roll clip on request and place it directly into the timeline, filling a hole in coverage. Neither post detailed the install beyond an offer to share setup steps, so how it holds up in finished output isn't yet clear.

Maya: Last one, and it's a quiet bug. Issue #95690, open since September 20, shows Claude Code skips AGENTS.md, the shared project-instructions file, whenever usage-data reporting is off, or when the tool runs through Bedrock or Vertex.

James: The cause is odd. Version 2.1.277's fallback ships off by default and gets switched on remotely through a Statsig feature flag. Checking it needs a network call, so anyone who disabled telemetry couldn't make it. As the issue put it: no error, no warning, it just doesn't load.

Maya: For now, the fix that ships is your own. Keep a thin CLAUDE.md in every project with one import line pointing at AGENTS.md. That's Claude Code Review for now, so keep an eye on those settings you never see.