- AMD AI director says Claude Code lost performance in a February 2026 update
- Claude Code “cannot be trusted” on complex tasks per thousands of coding sessions
- Anthropic says it reduced effort to medium, but Teams and Enterprises may be able to get high
AMD AI director Stella Laurenzo has claimed Claude Code has become less effective since around February 2026, arguing that it “cannot be trusted to perform complex engineering tasks.”
Laurenzo’s criticism isn’t unfounded, based on the company’s analysis of over 6,800 coding sessions, nearly 235,000 tool calls and nearly 18,000 reasoning blocks.
“Every senior engineer on my team has reported similar experiences/anecdotes,” Laurenzo wrote, noting that stop-hook violations (where Claude gave up early, dodged responsibility or asked for unnecessary permissions) increased from zero in early March to around 10 per day afterwards.
Article continues below
Claude Code is getting worse, AMD head warns
In a GitHub post, user stellaraccident (otherwise known as Stellar Laurenzo) identified a strong correlation between the introduction of thinking redaction (redact-thinking-2026-02-12) and a decline in performance on complex tasks. The AMD director argues that extended reasoning can be “load-bearing” for advanced engineering work.
Laurenzo also observed a shift from research-first to edit-first behavior, generating lower-quality code, worse adherence to conventions and overall reduced reliability for long sessions.
Anthropic has already responded to the research with a multi-faceted explanation. Claude Code’s Boris explained that the redact-thinking-2026-02-12 setting only hides reasoning from the UI, and does not actually reduce reasoning.
The company also introduced adaptive thinking with Opus 4.6, with there model deciding dynamically how long to think to improve performance and efficiency.
“Some people want the model to think for longer, even if it takes more time and tokens,” Boris added. “To improve intelligence more, set effort=high via `/effort` or in your settings.json.”
With medium effort, or effort=85, now default for users, Anthropic has promised to testing higher effort for Teams and Enterprise users for them to “benefit from extended thinking even if it comes at the cost of additional tokens & latency.”
“I appreciate the depth of thinking & care that went into this,” Boris also noted, crediting AMD’s Laurenzo for the analysis.
Follow TechRadar on Google News and add us as a preferred source to get our expert news, reviews, and opinion in your feeds. Make sure to click the Follow button!
And of course you can also follow TechRadar on TikTok for news, reviews, unboxings in video form, and get regular updates from us on WhatsApp too.
https://cdn.mos.cms.futurecdn.net/Zk5on5LX67HGEKda5J7HzD-2560-80.jpg
Source link
