Breaking

Anthropic's Opus 5.5 Performance Drop Was Never a Downgrade, Investigation Reveals

October 22, 2026 at 21:36 EDT

Foundation Models
AI Agents
Research & Papers

Weeks after launch, users reported that Anthropic's flagship Opus 5.5 model seemed degraded when run with identical prompts 20 days apart—but a pre-registered benchmark study and software analysis have traced the perceived decline to client bugs and transparent model routing, not deliberate downgrades.

Breaking News · October 2026

Opus 5.5 Degradation Traced to Client Bugs, Not Model Downgrade

Investigation reveals Anthropic's flagship model underwent no deliberate performance cut; instead, a Claude Code software regression, effort-level defaults, and routing logic explain user reports of weaker performance.

The Timeline

Sept 22

Opus 5.5 released

Sept 24

LiveNerf benchmark starts

Oct 5

Client v2.1.289 deployed; GitHub issue filed

40%

Cost reduction vs Opus 5

30%

Speed improvement

68%

Thinking tokens lost (buggy client)

Benchmark test questions

Root Causes Identified

Client Software Regression

Claude Code v2.1.289 reduced thinking tokens from 148–196 to 63 per turn

Effort Level Default Changed

From high to medium; cost: −4.2 accuracy points and −26% output tokens

Model Fallback Routing

Cybersecurity tasks automatically routed to older Opus 4.8 without user notification

Behavioral Breaks

Thinking blocks moved between tool calls; forced tool use removed; API changes

Verdict: LiveNerf benchmark found no statistical evidence of deliberate model downgrade. Perceived degradation attributed to client bugs, API defaults, and transparent routing rather than model regression.

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.