The `claude -p` flag now has its own included budget ($20-$200/month depending on plan) rather than sharing limits with interactive Claude Code sessions. No extra cost; it just stops competing with your interactive usage.
Starting June 15, paid plans include a claimable monthly credit that covers the Agent SDK, the `-p` flag, GitHub Actions, and third-party Agent SDK apps. It is a separate budget from subscription limits rather than a cap on top of them.
Programmatic usage via the SDK and the `-p` flag was previously sharing rate limits with interactive Claude Code and chat sessions. Starting June 15, it gets its own separate budget, leaving subscription limits reserved exclusively for interactive use.
The UK AI Safety Institute tested Mythos Preview and found it's the first model to solve both of their cyber ranges end-to-end, including the previously unsolved 'Cooling Tower' range. Anthropic says it's working to get the model to defenders quickly and has more to share on its Glasswing security work.
A new /goal command keeps Claude working until a user-defined completion condition is met, part of a broader set of tools designed for long-running, unattended workflows.
The /goal mechanic is a built-in implementation of the Ralph loop: every time Claude tries to stop, it checks the stated condition against the transcript. If the condition isn't met, it keeps going; if it is, you get a "Goal achieved" summary.
You give /goal a natural-language condition, such as all tests in a specific directory passing and the lint step being clean, and Claude keeps working until that condition is satisfied.
Auto mode removes the need for Claude to pause and wait for user input during long-running tasks. Enable it with shift + tab in the CLI or via the mode selector on desktop.
The /schedule slash command lets you trigger Claude on a recurring schedule. Nightly test runs, morning triage, weekly cleanup tasks, that sort of thing.
A stop hook lets you programmatically control when Claude can call a task complete. You can wire it to a test suite, a CI endpoint, or any custom condition before Claude wraps up.
Weekly usage limits jumped 50% for all Pro, Max, Team, and seat-based Enterprise users, effective immediately and running through July 13.
The 50% weekly limit increase applies across CLI, IDE extensions, desktop, and web. It runs through July 13 at 6PM PDT, requires no opt-in, and stacks with last week's separate 2x expansion to 5-hour limits.
A team member flagged a circulating link as fake and warned users not to click it.
On June 15, Agent SDK and 'claude -p' usage will draw from a new dedicated credit bucket instead of your normal interactive limits, which stay unchanged. Every plan gets between $20 and $200 in monthly SDK credits, and those credits also work with third-party tools like T3 and OpenCode.
Developers building on Claude via the Agent SDK will draw from a new dedicated monthly credit pool instead of needing a personal API key, separate from existing subscription limits.
Claude subscription plans can now be used with the Agent SDK, opening up a new usage bucket for subscribers building with or using third-party apps.
Apps built on the Agent SDK, including tools like Conductor and OpenClaw, now work with a paid Claude plan but draw from the new dedicated programmatic credit rather than subscription limits.
No action is needed before June 8, when users will receive an email to claim their new programmatic usage credits. The new billing structure officially takes effect June 15.
Claim the monthly credit once and all programmatic usage draws from it automatically. Once exhausted, usage either continues via pay-as-you-go API credits (if enabled) or pauses until the credit resets the following month.
