The community is buzzing with accusations that Claude Opus 5.5 has been "nerfed," citing a sudden drop in performance for complex workflows. However, a deep dive into the Claude Code changelog and API documentation reveals that the model weights haven't changed. Instead, Anthropic shipped a series of silent default changes around November 3 that cause Claude Code to ignore user-defined effort levels for Opus 5.5, forcing it to run at a lower medium effort instead of the high effort many users expect.

The Silent Default Shift

The primary source of confusion is a behavioral change in how Claude Code handles the effortLevel parameter. While most Claude models default to high effort, Opus 5.5 and Haiku 5.5 default to medium. Critically, if users set a global effortLevel in ~/.claude/settings.json before Opus 5.5 shipped, Claude Code now ignores it for the new model. The settings reference explicitly states that top-level effort settings "keep applying where it applied before... [but] Opus 5.5 and models released after it ignore it and start at their own default." This means users who haven't updated their per-model settings have been running Opus 5.5 at medium effort for weeks, believing they were on high.

November 3 Changelog Changes

Anthropic did not announce a new checkpoint for Opus 5.5, but they did release several Claude Code updates between October 26 and November 4 that altered behavior. Version 2.1.317, released on November 3, changed how subagents inherit effort levels. Previously, subagents might have inherited the session's high effort; now, unless a level is explicitly saved for Opus 5.5, they revert to the model's lower default. Additionally, version 2.1.318 removed a fix that preserved earlier thinking patterns when resuming old conversations, which can make long-running sessions appear to degrade in quality due to dropped context rather than model failure.

Perception vs. Reality

Much of the perceived degradation is also visual. Opus 5.5 enforces adaptive thinking, which cannot be disabled, and defaults to omitting thinking blocks in the display. This makes the model appear less transparent, leading users to believe it is "fumbling" basic tool calls. While some users report faster usage limit consumption, this is often a side effect of the metering system or the increased token usage from thinking blocks, not necessarily a drop in reasoning quality. The Hacker News thread highlights a divide: some see a genuine regression, while others correctly identify the issue as a configuration mismatch.

How to Fix Your Effort Levels

To restore expected performance, users must stop relying on global aliases and start pinning specific models. First, use the full model ID claude-opus-5-5 instead of the opus alias, which updates silently. Second, move effort settings to the modelSettings block in ~/.claude/settings.json to ensure they apply to Opus 5.5 specifically. Finally, for subagents, explicitly define the effort level in the agent definition or set the CLAUDE_CODE_SUBAGENT_EFFORT environment variable. Running a small regression test with pinned parameters is the only way to distinguish between a true model regression and a misconfiguration.

Key Takeaways

  • Opus 5.5 defaults to medium effort, unlike previous high-effort defaults, and ignores global effortLevel settings in user config.
  • The November 3 update (v2.1.317) stopped subagents from inheriting session effort levels, forcing them to use Opus 5.5's lower default.
  • Perceived performance drops are often due to hidden thinking blocks and dropped context in resumed sessions, not model weight changes.
  • Users must pin the model ID claude-opus-5-5 and set effort per-model to ensure consistent high-effort behavior.

The Bottom Line

Anthropic is playing fast and loose with silent defaults, turning configuration drift into a perceived model regression. Until you pin your model IDs and verify your effort levels, you are flying blind.