Claude Code Effort Showing 10/100? It Was a Display Bug, Not a Downgrade

If your Claude Code status line suddenly read effort: 10 while you had "high" selected, you weren't imagining a nerf. On August 22–23, 2026, a Hacker News thread — "Anthropic appears to be A/B testing reduced effort levels in Claude Code" — collected reports of the numeric effort value dropping to the number that used to mean low, even though users had explicitly picked high. Anthropic's answer, three hours into the thread: it's a display bug from an internal experiment, not a quiet downgrade.

This is the second time in five months Claude Code's effort setting has become a trust story. The first time, in March 2026, Anthropic really did ship a silent downgrade and apologized for it in a postmortem. This time, the company says the number was wrong but the behavior wasn't. Here's what happened, who confirmed what, and how to check your own sessions.

TL;DR

What users actually saw

Reports collected in the thread described a task run at "high" effort producing output that read like a rushed pass — then checking the session's numeric effort value and finding 10 out of 100, a figure with no obvious mapping to any of Claude Code's five labeled tiers (low | medium | high | xhigh | max). One X user tracking the change, @argofowl, reported it was server-side, not the app — "Anthropic enrols Claude Code 2.1.236+ sessions into an experiment that shrinks the effort scale, older versions and Opus 5 are left alone... probably an A/B test, so not everyone will see it."

That detail matters. A server-side toggle explains why some builders saw it and others on the same Claude Code build didn't — and why it looked, at a glance, indistinguishable from the intentional downgrade Anthropic pulled in March.

Anthropic's response: Thariq on Hacker News

Three hours into the thread, Thariq from the Claude Code team posted the clarification that settled most of the discussion:

"We sometimes test API serving configs in Claude Code before rolling them out, and one running now maps the numerical effort value differently... the effort you selected is the effort you're getting. We've run in-depth evals to confirm this doesn't affect model performance."

Read carefully, that's two separate claims. First: the number on screen was remapped by a serving-config test — a technical detail about how effort tiers get encoded and displayed, unrelated to model weights or reasoning budget. Second: Anthropic ran evals to confirm the actual effort tier your session received matched what you selected, regardless of what the display showed.

That second claim is the one worth being skeptical of by default — "trust our internal evals" is exactly what a company would say whether or not it were true — but it's consistent with how the bug presented: a display-layer mismatch, not a pattern of degraded output tied to specific request IDs. No one in the thread produced before/after outputs at matched prompts showing a quality drop that tracked the display change, which is the evidence that mattered in March.

Why this isn't the March 2026 incident

It's worth being precise here because the two stories rhyme but aren't the same:

• March 2026 (real downgrade): Anthropic switched Claude Code's default reasoning effort from high to medium to fix UI latency — the interface looked frozen during long thinking passes. This silently reduced effort for anyone who hadn't manually overridden the default. Anthropic called it "the wrong tradeoff," reverted it April 7, and published a postmortem covering it alongside two other quality bugs from the same window.
• August 2026 (this story): The effort tier itself wasn't changed — the number displayed next to it was, as a side effect of testing serving configs. explainx.ai has not found before/after output comparisons in the thread that contradict Anthropic's claim.

Confusing the two is understandable — both produced the same headline symptom ("my high effort doesn't feel like high effort anymore") and both trace back to the same underlying tension: Anthropic tests infrastructure changes against a UI number that users have learned to trust as a proxy for output quality. The Claude Code effort vs. model guide explains why that number exists in the first place — effort controls files read, tool calls made, and verification depth, not just "thinking time," which is exactly why a mismatch between the label and the number reads as alarming even when nothing downstream actually changed.

What to do about it

1. Update Claude Code. Anthropic corrected the display mapping shortly after the thread surfaced — claude update (or your IDE extension's auto-update) resolves it for most users.
2. Check with /effort. Run the command to confirm your current tier and compare it against whatever numeric readout your status line or session info shows.
3. Don't self-diagnose "nerfing" from the number alone. If output quality genuinely drops, that's a real signal worth reporting — but a mismatched number by itself, per this incident, wasn't evidence of reduced work. Compare actual outputs at the same prompt, same model, before assuming a downgrade.
4. Report mismatches with specifics. The HN thread got a same-day answer because reporters included version numbers and screenshots. Vague "Claude feels dumber" posts get slower, vaguer responses.

This is also a good moment to revisit whether your team is even using effort correctly day to day — see Claude's Effort Parameter: The Complete Guide for when to raise it versus when to just fix your prompt or CLAUDE.md, which fixes more perceived "quality drops" than any settings change does.

The trust cost of unlabeled experiments

Anthropic is within its rights to test serving configs — that's normal infrastructure work every model provider does. But testing against a number users have been trained to treat as a quality signal, without a changelog note, is what turned a cosmetic bug into a "did they nerf Claude again" thread within hours. That pattern — Opus 5 feeling worse to work with even when benchmarks say otherwise, Opus 5 over-engineering simple tasks, and now an effort readout that looked like a stealth cut — keeps recurring because Anthropic's own March 2026 admission means the community's prior isn't neutral anymore. The next unlabeled experiment that touches a user-visible number will get the same reaction, deserved or not.

Version numbers, thread details, and Anthropic's statement are accurate as of August 23, 2026 — Claude Code updates ship frequently, so verify your build's current effort mapping with /effort before assuming this post still describes your session.

Related reading

• Claude Code Model vs Effort: Knowing More vs Trying Harder
• Claude's Effort Parameter: The Complete Guide
• Why Does Claude Opus 5 Feel Worse to Work With? The HN Debate
• Opus 5 Over-Engineering: The Reddit Reaction
• Claude Code Auto Mode Default for Pro, Max, and Team
• Claude Code Settings.json Complete Reference
• Official: Anthropic — An update on recent Claude Code quality reports
• Source: Hacker News — Anthropic appears to be A/B testing reduced effort levels in Claude Code