Anthropic faces lawsuit over frontier AI development
Claude Fable 5.2 is now the default model with a 1M token context window, while Code Mode lets Claude invoke its own tools through JavaScript to reduce round trips. Anthropic is extending Claude into workplace documents and meetings, and faces a lawsuit seeking to halt frontier model development alongside scrutiny over its compute spending plans.

Anthropic releases first frontier models since Amodei's call for slowdown as IPO buzz builds — CNBC
On this page
Highlight
Claude Fable 5.2 becomes the default Fable model
Fable 5.2 ships with a 1M token context window at $8/$40 per Mtok and $0.20/Mtok cache reads, and it is now the default. Saved Fable 5.1 picks move to fable automatically on Pro and Max, and on Team and Enterprise once your organization allows it; run /model claude-fable-5-1 if you want to stay on the older version.
Worth Knowing
Code Mode lets Claude call tools through JavaScript
Code Mode is an experimental opt-in where Claude writes JavaScript to invoke its own tools rather than issuing individual tool calls, which can cut round trips on multi-step work. See the code mode docs for how to turn it on, or how to disable it across an organization.
Sandboxed Bash commands now ask for permission by default
If you run with the Bash sandbox on, sandboxed commands no longer skip the permission prompt. Set sandbox.autoAllowBashIfSandboxed: true to restore the previous behavior, and if your sandbox is turned on by managed settings, that key needs to go in your highest-ranking managed source.
Unknown model IDs now assume a 1M context window
If you point Claude Code at a gateway or third-party model it does not recognize, it now assumes 1M tokens instead of 200K. On models with a smaller real window this can leave a session failing every turn once full, so set CLAUDE_CODE_DISABLE_1M_CONTEXT=1 before that happens, or use /autocompact to compact earlier.
VS Code gains prompt suggestions and live command output
The VS Code extension now shows the same prompt suggestions as the terminal, with Tab to accept and promptSuggestionEnabled: false to turn them off, and streams the live output of running Bash or PowerShell commands under the command card. You will also see a trust prompt before your first session in a folder you have not trusted; declining leaves that folder's settings and MCP servers unused.
Fast mode is available on Vertex
If you run Claude Code through Vertex in the global, us, or eu regions, /fast now works there and falls back to standard speed when Vertex declines the request.
Sessions can now message each other 100,000 times
The cap on messages your sessions on the same machine can send one another without your intervention rises from about 20 to 100,000, so multi-session and Desktop Code tab workflows no longer stall partway through.
Under the Hood
Security fix in zstd decoding plus Bash permission gaps closed
A security issue in how zstd-compressed responses are decoded is patched, alongside a batch of fixes where compound commands, sudo/env wrappers, wildcard allow rules, and cp/mv targets could skip a permission prompt; restarting Claude Code activates the already-downloaded patch.
Lower memory with subagents and faster startup
Finished subagent transcripts are freed sooner instead of lingering while later subagents run, a memory spike after Claude edited a very large file is fixed, and startup and first-command time improve by resolving PATH lookups without spawning a helper process.
In the News
If you rely on the Bash sandbox or a third-party gateway, the permission-flow and context-window defaults are the two changes most likely to alter how your sessions behave this week.



