The effort dial isn't the higher, the better
Claude's "effort" setting — how deeply the model thinks before answering — has long been treated as a simple rule: crank it up for better results. But with Opus 5, Anthropic's own guidance shows that on coding tasks, pushing effort past high into xhigh or max can actually lower the score. If your team defaults to max "just to be safe," it's worth a second look.
The moment "higher is
always better" broke
Claude models expose an effort parameter that controls how deeply the model reasons before responding. Higher effort means more time and tokens spent thinking things through. Through the Opus 4.8 generation, the rule of thumb was simple: harder task, higher effort. Many teams went further and just defaulted everything to max "to be safe."
On July 24, 2026, Anthropic shipped Claude Opus 5 as the successor to Opus 4.8 — and upended that assumption. Anthropic's official Migration Guide states teams should "use high as the default for most tasks, including workloads that ran at xhigh on Claude Opus 4.8." Because the base model itself got stronger, a lower effort setting now reaches the same bar — or better.
There's a sharper point underneath that. On coding tasks specifically, pushing effort from high up to xhigh or max can actually lower the score, according to the same guide. The old Opus 4.8-era intuition that "more thinking is always better" simply doesn't hold for Opus 5.
| Opus 4.8-era habit | What to check with Opus 5 |
|---|---|
| Default to max effort when unsure | Start from high, then benchmark per task |
| Use xhigh for hard coding tasks | High often matches or beats it |
| Assumed more effort = more safety | xhigh/max risk overthinking and lower scores |
Five levels on the dial —
where should it sit?
The effort levels Opus 5 exposes, and the new baseline the official guide recommends.
The What's new in Claude Opus 5 doc spells out the same trade-off, and the migration guide is explicit about the top of the dial: "Test max effort where maximum capability matters more than token spend. It can deliver gains on the most demanding tasks but may show diminishing returns from increased token usage and can be prone to overthinking on simpler ones."
In other words, max has been redefined from "always the safest choice" to "an option you reach for selectively."
Engineers and PMs feel
this differently
A hardcoded effort setting bites in different places depending on your role.
Engineering teams
Check whether your coding agents or auto-review pipelines are hardcoded to max effort. On Opus 5 that setting may actually be lowering review pass rates and test success rates. Audit every effort flag in your CI pipeline.
PMs and product leads
Higher effort means higher latency and higher token spend. An "always max for accuracy" policy can drive up cost and slow response times — and, for some tasks, hurt quality at the same time. It's worth revisiting SLA targets and cost projections together.
Start with these three
Re-run your own effort sweep
Don't carry over settings tuned for Opus 4.8 — re-benchmark low through max on your own task set. The official guide itself recommends running "a fresh effort sweep on your own evals rather than carrying over a setting tuned for an earlier model."
Make high the default for coding
Unless you have a specific reason, stop hardcoding xhigh or max for coding tasks and start from high instead.
Log every max deployment
Wherever you still use max, track token usage, latency, and actual task success so you can measure cost-effectiveness over time. Max that isn't earning its keep is wasted spend.
Run a fresh effort sweep
on your own evals — don't carry over
a setting tuned for an earlier model.
This isn't "never use max"
This isn't a simple reversal into "lower is always better," either. Anthropic's guidance here is qualitative — there's no published benchmark table showing exactly how much scores drop above high, and how large (or even present) that drop is likely depends on the task type and prompt design. Teams need to confirm the actual effect on their own evals rather than assume it applies uniformly.
The guide itself notes that "max effort... can deliver gains on the most demanding tasks," so treating max as always wrong would be its own oversimplification. The right effort level is workload-specific. The real lesson here is to stop defaulting to max on autopilot and start verifying per task instead.