Claude Code Finally Lets You Cap AI Effort —

Somewhere between the release notes and your next invoice, Claude Code changed. The last release we covered was v2.1.235 back in August — a spellcheck and polish drop. Since then, Anthropic shipped 30 more versions in four weeks, and buried in them are two settings that matter more to your wallet than any feature list: a hard effort cap and a new default model that makes the 1M context window cheap.

This is the release-cycle story told in plain English — what changed, what it costs you, and the one setting worth turning on today.

The headline: you can now cap how hard Claude tries

Claude Code’s models work in “effort levels” — low, medium, high, max. More effort means more reasoning before it touches a file, which means better answers on hard problems and higher bills on easy ones. Until now, that knob was per-session and easy to forget.

v2.1.267 adds maxEffortLevel (changelog): a top-level setting (or per-model under modelSettings) that caps effort on every provider — including Bedrock, Vertex, and Foundry. You can always pick a lower level per session, but it can never go above your cap.

Why this matters for beginners: the most common “why is my bill $40?” story is a model silently running at max effort on tasks that need none — a one-line rename doesn’t need ten minutes of chain-of-thought. A cap makes spend predictable.

// .claude/settings.json
{
  "modelSettings": {
    "claude-sonnet-5": { "maxEffortLevel": "medium" }
  }
}

The other headline: Fable 5.1 is now the default — 1M context, cheaper

Back in v2.1.257, Anthropic made Claude Fable 5.1 the default Fable model (changelog). The numbers that matter:

1M token context window — a big codebase fits in one conversation without manual pruning.
$10/$50 per million tokens with $0.25/Mtok cache reads — the cheapest frontier-class pricing Anthropic has shipped.
Sessions on the 1M window now auto-compact at ~967K tokens instead of ~934K (v2.1.247) — you get more useful memory before summarization kicks in.

The [1m] tag you may have seen in /model is this window. If you’re on a Max plan or an API key, worth explicitly checking which model you’re pinned to — older Fable 5 sessions stayed on the smaller window unless you picked the new one.

The quiet theme: a month of token-efficiency fixes

The biggest cost leak in agent use is prompt-cache misses — re-sending the whole conversation because the tool list changed slightly. Claude Code shipped a long list of fixes for exactly this:

Parallel tool calls and hook output no longer dropped on resumed sessions over 5 MB (v2.1.267).
Mid-session MCP/plugin tools no longer rewrite the tool list and break prompt-cache reuse (v2.1.267).
Model switching no longer re-sends every tool definition (v2.1.267) — previously a guaranteed cache miss.
A likely-cause line for cache misses was added to /cost (v2.1.260) — it tells you why you missed, which is the first step to fixing it.
Per-session prompt-cache stats (hit ratio, tokens re-cached, warm/cold) landed in /cost (v2.1.251).
OAuth token refreshes no longer invalidate the cache (v2.1.248) — that used to nuke it roughly once an hour in long sessions.

For a beginner the takeaway is simple: update, then run /cost after a long session. If you see a low hit ratio, the cause is now printed right there instead of being a mystery.

Three more features worth knowing

/diff panel in fullscreen mode (v2.1.260) — a side panel that shows your uncommitted changes live as Claude edits. No more switching to git diff to see what the agent just did; it’s next to the conversation.

/skill-doctor (v2.1.261) — lists which loaded skills go unused and what they cost in context, so you can prune them. Every unused skill is tokens that could be cache hits.

--restricted mode (v2.1.248) — a lock-it-down flag that strips the tools that run commands or fetch web pages, keeps file tools inside the working directory, and refuses bypassPermissions. One flag for a genuinely safer sandbox.

What the release train says

Look at the changelog from v2.1.236 through v2.1.267 as one document and two things stand out.

First, cost control is the product now. Effort caps, cache-cause reporting, per-session cache stats, cheaper model pricing — Anthropic is competing on your bill, not just on benchmark scores. That is exactly the direction we argued in The Pricing Problem With AI Coding Tools: the tools that win are the ones that make spend predictable.

Second, the safety perimeter keeps tightening quietly. The [v2.1.260](https://github.com/anthropics/claude-code/blob/main/CHANGELOG.md) release reverted a change that made Read() deny rules apply to Bash arguments — it had started denying npm run build under innocent-looking rules. The bigger trust story is that these permission boundaries are being stress-tested in public, and patched in days.

What to actually do

Update — claude update (or restart the VS Code extension).
Set a default effort cap — pick medium and see if your tasks notice; most won’t.
Check your model — run /model and look for the [1m] Fable 5.1 or Opus options if your plan includes them.
Check /cost after a long session — read the prompt-cache line; it now tells you what broke the cache.
If you run agents unattended, try --restricted for anything that doesn’t need shell access.

Full changelog: v2.1.235...v2.1.267

Disclosure: some links on this site are affiliate links (aiFiesta).

Claude Code Finally Lets You Cap AI Effort — And Stop Wasting Money