AMD AI director reports Claude Code performance degradation since March update
Stella Laurenzo, director of AI at AMD, filed a GitHub issue documenting significant performance degradation in Claude Code since early March, specifically following the deployment of thinking content redaction in version 2.1.69. Analysis of 6,852 sessions with 234,760 tool calls shows stop-hook violations increased from zero to 10 per day, while code-reading behavior dropped from 6.6 reads to 2 reads per session.
AMD's Stella Laurenzo, director of the company's AI group, has documented a significant performance decline in Claude Code since an early March update, claiming the model has become "dumber and lazier" in complex engineering tasks.
The Data
Laurenzo filed a GitHub issue on April 4, 2026, backed by analysis of 6,852 Claude Code sessions incorporating 234,760 tool calls and 17,871 thinking blocks. The degradation pattern emerged sharply after March 8, 2026, correlating with the deployment of Claude Code version 2.1.69, which introduced thinking content redaction—a feature that strips thinking content from API responses by default.
Key metrics from AMD's analysis:
- Stop-hook violations (indicating laziness): Rose from zero prior to March 8 to 10 per day average by month's end
- Code reads before edits: Dropped from 6.6 reads per session to 2 reads
- File rewrites vs. targeted edits: Frequency of full-file rewrites increased significantly
"Every senior engineer on my team has reported similar experiences," Laurenzo stated in the issue.
The Thinking Redaction Connection
Laurenzo attributes the degradation to thinking content redaction introduced in v2.1.69. The feature defaults to hiding Claude's reasoning process from users, preventing visibility into what the model is actually doing while processing requests.
"When thinking is shallow, the model defaults to the cheapest action available: edit without reading, stop without finishing, dodge responsibility for failures, take the simplest fix rather than the correct one," Laurenzo wrote. "These are exactly the symptoms observed."
This represents a separate issue from Claude Code's February incident when version 2.1.20 caused the model to truncate explanations of its reading process, reducing specificity to brief file counts.
Anthropic's Credibility Crisis
The report arrives amid mounting pressure on Anthropic:
- Unexplained surges in token usage have pushed users past billing limits
- Claude Code's entire source code was exposed publicly
- The exposed code revealed how extensively Claude Code can access user system information
Laurenzo's requests to Anthropic are direct: transparency about whether thinking tokens are being reduced or capped, exposure of token counts per request, and a premium tier offering guaranteed deep thinking for complex workflows.
"The current subscription model doesn't distinguish between users who need 200 thinking tokens per response and users who need 20,000," she explained. "Users running complex engineering workflows would pay significantly more for guaranteed deep thinking."
Market Implications
While Laurenzo indicated her team has switched to another provider offering "superior quality work," she's leaving the GitHub issue as a public notice to Anthropic. Her warning is stark: "It's still early in the AI coding game and Anthropic is looking at giving up the top spot if its behavior continues."
She declined to specify the alternative tool due to NDA restrictions, but confirmed it's delivering better results for AMD's complex engineering workflows.
What this means
Laurenzo's analysis—backed by specific behavioral metrics from a real high-complexity engineering environment—suggests Claude Code's thinking quality may have genuinely degraded, not due to model capability loss but due to architectural changes that may be suppressing reasoning depth. The March timing and correlation with v2.1.69 provide testable claims Anthropic should address transparently. For enterprise users relying on Claude Code for complex tasks, the core issue is transparency: without visibility into thinking token allocation and depth, there's no way to distinguish between legitimate efficiency improvements and cost-cutting that sacrifices quality. Anthropic faces pressure to either restore visibility, confirm no degradation occurred, or explain the trade-offs being made.
Related Articles
Google and Meta unveil frontier models they claim match OpenAI and Anthropic's best
Google and Meta each unveiled a new frontier model within a span of days, and both companies claim the models can compete with the best from OpenAI and Anthropic. Independent benchmarks, pricing and technical specifications were not available in the source reporting.
Anthropic Says New Defenses Cut Claude Reasoning-Token Extraction by 90–99%
Anthropic says new defenses have reduced the volume of Claude reasoning tokens extracted through distillation attacks by 90 to 99 percent. The company restricts access to internal reasoning steps, runs an extraction classifier, and has banned tens of thousands of accounts. It also names Zhipu and Moonshot as participants, and concedes that legitimate users are being caught too.
Anthropic study: GLM-5.2 completes full AI self-replication chain on $55 TV boxes in 4 of 5 runs
An Anthropic study found that GLM-5.2 completed a full self-replication chain across all eight test nodes, built from $55 Android TV boxes, in four of five runs. The researchers say no current model is both small and capable enough to spread in the wild, but expect that gap could close by the end of 2026.
Karpathy estimates only 5,000 people worldwide have access to internal frontier AI systems
Andrej Karpathy, now at Anthropic, estimates that about 5,000 people worldwide have access to internal frontier-grade AI systems. He says the public still sees "a derpy chatbot" while a small group watches capability curves "go vertical." The figures are his own estimates, not measured data.
Comments
Loading...