Run checkable tasks at low effort first, and re-run only the failures at high
If tests can tell you whether a task worked, a fixed high-effort setting is the expensive option. A low-effort first pass plus a high-effort retry on the failures solved slightly more tasks at a little over half the cost in Anthropic's measurements.
The problem
Most teams pick one effort level per model and leave it there. Set it to high “to be safe” and you pay for deep thinking on every task, including the easy majority that low would have solved on the first try.
The tip
When an outcome is checkable (tests pass, a linter is clean, the row count matches), stop treating effort as a fixed setting. Make it a retry policy:
- Run every task at
low. - Check the result automatically.
- Re-run only the failures at
high.
@camperaigg’s thread on Anthropic’s new playbook flags this as one of the key tips, and Anthropic’s cost guide backs it with numbers. On its SWE-bench Pro subset with Claude Opus 5.5, about 13% of tasks failed at low. Re-running those at high solved about 97% for about $0.17 per task, versus 95.3% at $0.29 with everything on high.
Example
def solve(task):
result = run_agent(task, model="claude-opus-5-5", effort="low")
if run_tests(task, result): # your failure signal
return result
# fresh session, so the cache and effort change don't fight each other
return run_agent(task, model="claude-opus-5-5", effort="high")
Start the retry as a new session. Changing top-level effort mid-conversation invalidates the prompt cache.
When NOT to use it
Skip it when you have no reliable checker (a verifier that passes bad work lets failures through), or when latency matters more than cost, because every first-pass failure takes two runs’ worth of wall-clock time.