The --restricted flag (or CLAUDE_CODE_RESTRICTED=1 env var) removes code and command execution tools, removes WebFetch unless explicitly named in --tools, keeps file tools sandboxed inside the working directory, refuses bypassPermissions, and ignores local settings files. Forty-nine CLI changes total in the release.

Restricted mode is not a permissions prompt. It is a tool-surface reduction. The model still reads and writes files, but cannot execute commands, run code, or reach the web. For CI pipelines and shared runners, that is the containment primitive security teams have been asking for since Claude Code left beta. The secondary item in the release: experimental.cacheTtl in agent frontmatter, set to "5m" or "1h", overrides the prompt-cache TTL per agent. On long-running agentic workflows, per-agent cache TTL is the difference between burning tokens on every context reload and keeping the cache alive across a multi-hour run.

The pattern: restricted mode is the third major enterprise-readiness feature in Claude Code this month. Cross-session messaging (v2.1.224, Aug 7), self-hosted runners in public beta, and now restricted mode. Anthropic is building Claude Code into something a security team can approve.

Builder's move: Add --restricted or set CLAUDE_CODE_RESTRICTED=1 in any non-developer environment. Review the updated --tools allowlist before deploying.

A Chromium-based browser opens in Cowork's side panel when Claude detects a task requiring the web. It is separate from your personal browser. Claude cannot see your tabs, bookmarks, or saved passwords unless you manually import credentials site by site. Live on Enterprise plans as of August 26, rolling out this week to Pro, Max, and Team on macOS, Windows, and Linux.

The prior Chrome extension was a workaround. This is a first-party integration. The model navigates, reads, clicks, and fills forms in its own sandboxed session, with no extension to install or maintain. Enterprise admins who blocked the extension face a different security conversation now: a native browser in a managed desktop app is not the same risk surface as a browser extension with broad permissions.

The contrast with OpenAI's browsing story matters here. OpenAI's web access runs primarily through API tool calls and ChatGPT's built-in search, not a dedicated client-side browser the model controls directly. Anthropic is shipping the full stack. The sequence is coherent: Cowork as the workspace, the browser as the navigation layer, Claude Code as the terminal layer, cross-session messaging as the coordination layer. At some point these pieces fuse into something that looks less like a chatbot and more like a managed agent runtime.

Builder's move: Enterprise users can test the browser in Cowork today. Pro, Max, and Team users should see the update within the week.

Scene extension up to 40 seconds in 10-second increments, first/last frame keyframe control, a 360p draft mode that generates previews 60 percent faster at one-third the cost of standard output, and 4K upscaling. Available in Google AI Studio, the Gemini API, Flow, and Gemini Enterprise Agent Platform.

The mechanism: the model analyzes up to 10 seconds of prior context to extend footage seamlessly, removes the ceiling that made early Gemini video feel clipped, and the 360p-then-upscale workflow is a real cost-optimization pattern for iterative production pipelines. Prototype at 360p, upscale the keepers. The economics of video generation just changed for anyone building on the API.

The contrast: OpenAI's Sora is the competing video model. Sora has not shipped an API control surface matching this level of developer configurability. Google just pushed video generation API capabilities meaningfully past what any other platform delivers to developers today.

Builder's move: Test Gemini Omni 1.1 Flash in Google AI Studio before committing to another video pipeline. The 360p draft mode alone justifies a half-hour evaluation.

Per Ramp data on 70,000 U.S. businesses: Fable 5 holds 11.4 percent of Anthropic dollar spend and just 6 percent of tokens two months after launch. Opus 5, launched in late July at roughly half the price, has already surpassed Fable 5 in enterprise spending. Anthropic's annualized revenue run rate hit $65 billion in July, up from $47 billion in May. The overall picture is growth. The distribution is telling.

Fable 5 runs around $10 per million tokens, double GPT-5.6 Sol at its current promotional rate. Anthropic's 30-day data retention requirement was explicitly flagged by Ramp economist Ara Kharazian as a compounding drag. Enterprise procurement teams can approve a capable model at half the price. They have a harder time explaining a data retention clause to legal.

The read: the enterprise model pricing war is not a race to the top. It is a race to the acceptable floor. Opus 5 crosses that floor. Fable 5, at current pricing and terms, does not, at least not for workloads where price and data handling are the decision variables. Anthropic leads OpenAI on adoption count (43.5 percent of tracked businesses vs. 39.7 percent), but the mix inside its own portfolio is not what the launch roadmap implied.

Builder's move: If you default to Fable 5 on the assumption it is best for every task, run your evals. On code and knowledge work, Opus 5 may deliver comparable results at half the cost. The retention policy is a variable, not a given.