🧭 Cache Diagnostics Leaves Beta — Drop the Header, Add the Object

Cache diagnostics is now generally available on the Claude API. If you've been relying on the cache-diagnosis-2026-04-07 beta header to get cache miss reasons in API responses, you can — and should — migrate to the new stable interface: include a diagnostics object in your Messages request body instead. The beta header still works for now, but it will eventually be retired.

What changed

• Opt-in mechanism: Add the diagnostics object to any POST /v1/messages request to enable diagnostics for that call.
• Always-present response field: The response from POST /v1/messages now always includes a top-level diagnostics field — it is null when the request did not include the diagnostics object, and populated when it did.
• Cache miss reasons: The populated diagnostics object contains cache_miss_reason with concrete variant values so you can distinguish cold misses from expiry from key mismatches.

# Before (beta header approach — works but will be retired)
headers = {
    "anthropic-beta": "cache-diagnosis-2026-04-07"
}

# After (GA approach — use the diagnostics object)
body = {
    "model": "claude-opus-5-5",
    "messages": [...],
    "diagnostics": {}      # empty object is enough to opt in
}
# Response will include: { "diagnostics": { "cache_miss_reason": "..." } }

🧭 Pre-Output Refusals Now Billed in Three New Categories — Audit Your Cost Assumptions

Anthropic has expanded refusal billing: pre-output refusals (those that arrive before any model text is generated) are now billed when they fall into one of three categories: bio, frontier_llm, and reasoning_extraction. The stated reason is that false-positive rates for these classifiers are low enough that billing is now warranted. Mid-stream refusals — where the model begins generating and then stops — were already billed at full rates and remain unchanged.

What is and isn't affected

• Now billed (pre-output): bio (biosafety category), frontier_llm (attempts to elicit Claude to replicate other frontier models), reasoning_extraction (attempts to extract the model's internal reasoning traces).
• Still free (pre-output): All other refusal categories retain the existing behaviour — no charge before output.
• Billing rate: When billed, refusals are charged at the standard rates of the model that processed the request — the same as a successful completion.
• Fallback credit policy: Unchanged — the credit mechanism for valid requests that are incorrectly refused still applies.
• Surfaces: Applies across the API, Claude.ai, and Enterprise deployments.

You can inspect the refusal category in the API response via stop_details.category — this field has been present since the refusal-details rollout and now doubles as the billing signal for these three categories.

🧭 Claude Tag in Slack Now Pulls from Your Personal Connectors — Not Just Shared Channel Resources

Claude Tag in Slack has gained a significant new capability: when you @Claude in a channel, it can now access connectors tied to your individual account — Google Drive, calendars, CRM accounts — not just the shared connectors an admin has attached at the channel or workspace level. This distinction matters more than it might seem: it enables individual team members to contribute personalised, permission-gated data to collaborative channel conversations without exposing that data to the whole channel by default.

Two output modes to understand

• Review mode (default): Claude composes a response and shows it to you privately first. You approve before it posts to the channel. Useful when the response might contain data you want to redact or context-check before colleagues see it.
• Auto mode: Claude screens the response for sensitive content before posting automatically. Enterprise admins can enforce review mode org-wide, overriding individual user preferences where compliance requires it.

Audit trail and privacy

Activity from personal connectors logs to the individual user's account activity trail — not to the shared channel audit log. This preserves privacy between team members while still giving IT and compliance teams per-user accountability. From a compliance architecture standpoint, it means personal connector invocations do not appear as "channel events" and must be reviewed through per-user access logs in the Anthropic Console.

Availability: Rolling out now on Team plans. Enterprise to follow.

🧭 Claude Code v2.1.282: maxProseWidth, Critical Thinking-Block Fixes, and a Web Search Session-History Patch

Claude Code v2.1.282 is primarily a stability and correctness release, but it ships one new setting worth knowing — maxProseWidth — alongside fixes for two bugs that could cause silent, hard-to-diagnose session degradation.

New: maxProseWidth setting

The new maxProseWidth setting lets you cap the column width of Claude's prose output in terminal sessions while leaving tables and code blocks free to use full terminal width. Add it to your project's settings.json:

{
  "maxProseWidth": 100   // prose wraps at 100 chars; code/tables still expand to terminal width
}

This is useful when piping Claude's output to tools that expect fixed-width text, or when working in wide terminals where long prose lines become hard to scan.

Critical fixes to know about

• Web search results in session history: Conversations were failing with API errors when resuming sessions whose history contained undecryptable web search results. This manifested as unexplained errors on session resume with no obvious relationship to the current prompt — now fixed.
• Extended thinking dropped during slash commands: Running a slash command (e.g. /review, /test) could silently drop extended thinking configuration for the remainder of the session — now preserved correctly.
• Thinking blocks lost after model switches: Switching models mid-session caused thinking blocks to disappear from the context — a subtle regression that affected users pinning to specific models for different phases of a task.
• Auto mode prompt-cache reuse after session resume: Auto mode now correctly reuses the permission classifier's prompt cache when resuming a session rather than recomputing it from scratch — a meaningful latency and cost improvement for long-running sessions.
• Security: dangerous-rm check extended to shell variable patterns: The check that flags potentially dangerous rm operations now catches patterns like rm -rf $DIR/subpath where a shell variable is followed by a directory name — previously only bare paths were flagged.