Starting with a webcam photo of two bare walls, an Anthropic team member used Opus 5.5 to iteratively build a 2D and 3D room model at real scale. The full prompt chain is described, covering everything from initial layout to drag-and-drop scene editing and art sizing from purchase links.
Opus 5.5 is 20% cheaper per input and output token than Opus 5, and 60% cheaper on cache reads. The ClaudeDevs team published a cost breakdown and a calculator accessible from /usage to help users estimate real task costs.
An Anthropic engineer says Claude Tag in Slack writes more than 50% of their PRs every day, handles nearly all data analysis, and resolves most product bugs. Powered by Opus 5.5, Tag is proactive, programmable, and has memory and connector access. Example prompts include auto-triaging bugs, running end-to-end reproductions, brainstorming data hypotheses with 10M tokens of analysis, and generating interactive code explainers.
A new portal makes it easier to submit plugins, track review status, and monitor usage. MCP usage across Claude products has grown 110x this year, and plugins, which bundle MCP and skills, are becoming the primary way to build for Claude.
When hitting the 5-hour usage limit mid-task, Claude Code now tries to find a clean stopping point instead of cutting off abruptly. It uses a small, fixed allowance drawn from the weekly limit to finish what it can.
The new graceful stopping behavior at the 5-hour limit is available once per week on Pro plans, and each time a session limit is hit on Max and Team Premium plans. Users can continue past the wrap-up with extra usage if needed.
Developers can submit plugins either as a single MCP connector pointing to a remote server, or as a plugin bundle combining MCP servers and skills in a GitHub repo. Submissions happen directly inside Claude.
Once a plugin is live, the portal surfaces usage analytics including installs broken down by product and version, listing views, and which search queries led users to it.
A unified plugin discovery experience will roll out across Claude and Claude Code over the coming weeks. Plugins are the primary extension mechanism for third-party developers.
An Anthropic team member is actively collecting real-world examples of Claude Computer/Browser Use failures to fix them, citing a personal case where paying a barber via PayPal in a secondary Chrome profile broke down.
When asked to pay via PayPal logged into a second Chrome profile, Claude used its built-in browser with the wrong profile, couldn't import cookies, and never attempted to hand off to the user for manual payment entry.
A user responding to a call for browser-use failure examples wishes Claude had said 'I'll open the payment page; you enter the details and hit submit' rather than refusing or failing silently. The missing handoff pattern is the core frustration.
Browser automation stalls on authenticated flows like PayPal, airline sites, and Google Docs because Claude opens a clean browser session with no cookies and never prompts the user to import credentials. trq212 asked whether cookie import has been tried.
Claude often fails to detect prefilled usernames and passwords in browser automation, claiming it cannot see the fields, which requires users to manually convince it to proceed. This happens with custom browser control skills too.
A detailed investigation into effort levels, including evals and original tests, found some surprising results about when changing effort actually changes outcomes.
Low effort is useful when you want to stay in the loop and review Claude's steps; max effort is reserved for fully autonomous runs or security vulnerability hunting.
Extended thinking improved outcomes when failures came from edge cases, but made no difference when the root cause was a flawed approach or bad design. Separately, Claude rarely takes different reasoning paths based on effort level unless the prompt is extremely vague.
Interactive explainers for benchmarks and demos from the effort-level deep-dive are available on the new developer site.
