v2.1.111 dropped yesterday alongside Claude Opus 4.7 and it's one of the biggest Claude Code releases to date. Here's what shipped.
The headline: Claude Code now runs on Opus 4.7 with a new xhigh effort level that sits between high and max. This is the new default for coding and agentic work. The model uses adaptive thinking — fixed budget_tokens are gone (see the Agent Architecture topic below for the full breakdown).
Auto mode goes general. No more --enable-auto-mode flag. Max subscribers get auto mode out of the box with Opus 4.7. Toggle with Shift+Tab. A classifier-based system auto-approves safe operations and pauses on risky ones — 94.3% synthetic exfiltration detection rate, 0.4% false positive rate per Anthropic's internal testing (n=10,000 real traffic actions).
New features:
/ultrareview— deep multi-agent code review in the cloud (see next topic)/less-permission-prompts— scans your session history for commonly-prompted bash/MCP commands that are actually safe, then proposes a prioritized allowlist for.claude/settings.json/effortwithout arguments opens an interactive slider with arrow-key navigation between all five levels- "Auto (match terminal)" theme in
/theme— matches your terminal's dark/light mode - PowerShell tool rolling out on Windows; opt in on Linux/macOS with
CLAUDE_CODE_USE_POWERSHELL_TOOL=1
Quality-of-life:
- Plan files now named after your prompt (
fix-auth-race-snug-otter.mdinstead of random words) - Read-only bash commands with globs (
ls *.ts) andcdprefixed commands skip permission prompts&& Ctrl+Uclears entire input buffer (not just to start of line);Ctrl+Yrestores itCtrl+Lnow forces full screen redraw plus clears prompt/skillsmenu supports sorting by token count (presst)- Subcommand typo suggestions (
claude udpate→ "Did you meanclaude update?") OTEL_LOG_RAW_API_BODIESenv var for full API request/response body logging
Bug fixes worth knowing: Fixed LSP diagnostics from before an edit appearing after it (caused unnecessary re-reads). Fixed tab-completing /resume resuming an arbitrary session instead of showing the picker. Fixed /clear dropping the session name set by /rename. Fixed plugin dependency handling across conflicting/invalid/complex version requirements.
v2.1.112 followed as a hotfix: fixed "claude-opus-4-7 is temporarily unavailable" in auto mode.
/ultrareview launches a fleet of reviewer agents in a remote sandbox to find bugs in your branch or PR. Unlike /review (single-pass, local, seconds), ultrareview runs multiple agents in parallel that independently reproduce and verify every finding before reporting it. Reviews take 5-10 minutes and run entirely in the cloud — your terminal stays free.
Two modes:
/ultrareview 1234 # reviews PR #1234 directly from GitHub
PR mode clones directly from GitHub rather than bundling your local tree. If your repo is too large to bundle, Claude prompts you to push and use PR mode instead.
Pricing: Pro and Max subscribers get 3 free runs (one-time, don't refresh). After that, $5-$20 per review as extra usage depending on change size. Your account must have extra usage enabled — check with /extra-usage. Team and Enterprise get no free runs.
When to use which:
/review— quick feedback while iterating, counts toward normal usage/ultrareview— pre-merge confidence on substantial changes, deeper multi-agent analysis
The review runs as a background task. Track it with /tasks, and findings appear as a notification when complete. Each finding includes file location and explanation — ask Claude to fix it directly from the findings.
Requires authentication with a Claude.ai account (not API key only). Not available on Bedrock, Vertex, or Foundry.
Opus 4.7 introduces a breaking API change: thinking: {"type": "enabled", "budget_tokens": N} returns a 400 error. Adaptive thinking is the only thinking mode on Opus 4.7.
What adaptive thinking does differently. Instead of you specifying a token budget, the model decides per-step whether to think and how deeply. At high effort (the API default), Claude almost always thinks. At low, it skips thinking on simple problems. The effort parameter is now your primary control surface for thinking depth:
| Effort | Thinking behavior |
|---|---|
max | Always thinks, no depth constraint |
xhigh | Always thinks deeply with extended exploration (Opus 4.7 only) |
high | Always thinks, deep reasoning on complex tasks |
medium | Moderate thinking, may skip on simple queries |
low | Minimizes thinking, prioritizes speed |
Three more breaking changes (API only, not Claude Code interactive):
- Sampling parameters removed. Setting
temperature,top_p, ortop_kto non-default values returns 400. Use prompting to guide behavior instead. - Thinking content omitted by default. Thinking blocks still appear in stream, but
thinkingfield is empty unless you setdisplay: "summarized". If your UI streams reasoning, this will look like a long pause — fix it with one line. - New tokenizer. 1.0-1.35x more tokens depending on content. Update your
max_tokensto give additional headroom, including compaction triggers.
The interleaved thinking win. Adaptive thinking automatically enables interleaved thinking — Claude can think between tool calls. This is the biggest improvement for multi-step agentic work. Previously on Opus 4.6, interleaved thinking required adaptive mode specifically; it wasn't available in manual mode.
Migration path for claude -p agents: If you're using the Messages API directly with budget_tokens, switch to thinking: {"type": "adaptive"} with an appropriate effort level. In Claude Code interactive and claude -p, this is handled automatically — use /effort to tune.
Opus 4.7 is not a drop-in replacement for 4.6. The official best practices blog and the model docs both flag behavioral changes that require workflow adjustments. Here's what to expect and how to adapt.
1. More literal instruction following. Opus 4.7 won't silently generalize. If you ask it to fix one function, it won't refactor the surrounding code. This is a win for precision but means your prompts need to be more complete. "Fix auth" won't cascade into related improvements the way 4.6 sometimes did.
2. Fewer tool calls, more reasoning. The model reasons more before reaching for tools. If your workflow depends on aggressive file reads or search calls, you may need to explicitly describe when and why tools should activate: "Read all files in src/auth/ before proposing changes."
3. Fewer subagents by default. Opus 4.7 is more judicious about spawning subagents. For parallel work, be explicit: "Spawn multiple subagents when fanning out across items or reading multiple files."
4. Response length calibrates to complexity. No more fixed verbosity — the model matches output length to perceived task difficulty. If you want detailed explanations, say so. If you want terse output, say that too.
5. Better file-system memory. The model is noticeably improved at writing and using scratchpad files, notes, and structured memory stores across turns. If your agents maintain persistent memory (MEMORY.md, ledger files), expect better recall and more accurate updates.
6. More direct tone. Less validation-forward phrasing, fewer emoji, more opinionated output. If you had scaffolding to force interim status messages during long runs, try removing it — Opus 4.7 provides more regular progress updates on its own.
The tactical adjustment: Lead with complete task context rather than progressive clarification. Batch your instructions up front. Treat Claude as a capable engineer you delegate to, not a pair programmer you guide line by line. Opus 4.7 excels on long-running, ambiguous tasks (code review, multi-file changes, debugging) when given enough context to work autonomously.
Opus 4.7 ships with a new beta feature: task budgets. A task budget gives the model a soft token allowance across an entire agentic loop — thinking, tool calls, tool results, and final output. The model sees a running countdown and uses it to prioritize work and finish gracefully as the budget depletes.
model="claude-opus-4-7",
max_tokens=128000,
output_config={
"effort": "high",
"task_budget": {"type": "tokens", "total": 128000},
},
messages=[{"role": "user", "content": "Review the codebase and propose a refactor plan."}],
betas=["task-budgets-2026-03-13"],
)
Key distinctions:
task_budgetis advisory — the model sees it and self-moderates. It's not a hard cap.max_tokensis a hard per-request ceiling the model never sees.- Minimum task budget: 20K tokens.
When to use it: Token-constrained deployments where you need the model to scope its work. CI pipelines, budget-capped review passes, or any agent where "do as much as you can in X tokens" is the right framing.
When NOT to use it: Open-ended agentic tasks where quality matters more than cost. A budget too restrictive for the task causes less thorough completion or outright refusal. Default to no task budget and let the model work freely.
This is Opus 4.7-only and requires the task-budgets-2026-03-13 beta header. Not yet available in Claude Code interactive sessions — API and SDK only.
What you'll do: Walk through the six configuration points Boris Cherny (Claude Code's creator) recommends for Opus 4.7, verifying each one works in your environment.
Prerequisites: Claude Code v2.1.111+ (claude --version to check; claude update if needed). Max subscription for auto mode.
Steps:
- Verify you're on Opus 4.7. Open Claude Code, check the model indicator in the status bar. If it shows
opus-4-6, run/modeland selectclaude-opus-4-7. Confirm with:
The model column should show claude-opus-4-7.
- Try the effort slider. Type
/effortwith no arguments. Use arrow keys to move between levels:low → medium → high → xhigh → max. Set it toxhigh(the recommended default for coding). Press Enter to confirm. Run a small task and compare the depth of reasoning to what you're used to.
- Enable auto mode. Press Shift+Tab to toggle into auto mode. The mode indicator in your prompt changes. Run a sequence of file reads and edits — notice that read-only commands execute without prompts. Try something that touches external state (like a
curlcommand) and observe the classifier intervene. Toggle back with Shift+Tab.
- Generate your permission allowlist. Run:
This scans your recent session transcripts and proposes bash/MCP commands that were safe but prompted you repeatedly. Review the proposed allowlist. Accept what makes sense — it writes to .claude/settings.json.
- Test recaps. Start a medium-complexity task (e.g., "refactor the error handling in this module"). Let it run for a few turns, then leave the session idle for a minute. When you return, check if Claude provides a recap of what happened and what's next. If not, verify
/recapworks manually.
- Verify your agent's work. Give Claude a task with a verification step: "Add input validation to this function. Write a test that proves it rejects invalid input. Run the test." Watch whether Opus 4.7 follows through on verification more reliably than your previous experience. If it skips verification, add explicit instructions to your CLAUDE.md: "Always verify changes by running tests before reporting completion."
Expected outcome: You should have xhigh as your default effort, auto mode accessible via Shift+Tab, a tailored permission allowlist reducing prompt fatigue, and confidence that Opus 4.7's behavioral changes work with your setup.
Verify: Run /effort — should show xhigh. Check .claude/settings.json for any new permission entries from step 4. Run /cost after a few interactions to confirm you're on Opus 4.7 and observe token usage patterns.