Developers scratched their heads. Claude, Anthropic’s prized AI, seemed to stumble. Responses grew repetitive. Credits vanished faster. Complex tasks faltered. Whispers of ‘AI shrinkflation’ spread across GitHub, X, and Reddit. Users paid the same, got less sharpness—or so it felt. Talk Android captured the frustration in a detailed breakdown published May 3, 2026.
Anthropic dug in. Their report pinned it on the software layer, not the core model. Three bugs toppled like dominoes. First, on March 4, engineers dialed reasoning effort from ‘high’ to ‘medium.’ Goal: quicker replies amid a compute crunch. Users sensed laziness right away. Claude skimped on depth. Solutions felt shallow.
Then March 26. A cache tweak backfired. Meant to clear stale context after an hour idle, it erased short-term memory with every fresh turn. Forget restating prompts. That burned tokens double-time. Repetition spiked. Workflows ground down.
April 16 sealed it. New prompts curbed wordiness. Coding output dipped 3%. Stella Laurenzo, an AMD director, logged the reasoning slide. Third-party benchmarks confirmed Claude lagged peers. No model rollback. Just infrastructure slips stacking up. Platformer noted similar gripes with Opus 4.6, where users called it ‘regressed’ and token-hungry.
Anthropic moved fast post-probe. They reset limits, refunding wasted credits. Employees now ‘dogfood’ the public version—no insider perks. Every prompt tweak runs model-tailored tests. And @ClaudeDevs launched on X for straight talk on fixes. But trust frayed. X posts erupted. One user measured Opus 4.6 thinking 67% less—median reasoning characters plunged from 2,200 to 600. API retries ballooned 80-fold. Developers fumed over hidden logs masking the drop.
Runtime chronicled outages piling on. March 2026 saw Claude Code logins fail, Sonnet models glitch. A database I/O snag post-maintenance crippled sign-ins. Pro users raged on Reddit: hamsters on strike, productivity torched. Runtime. Reliability? Still shaky.
So why the mess? Compute scarcity. GPUs cost a fortune. Firms like Anthropic trim where they can—latency tweaks, cache squeezes. But haste bites back. Claude’s woes echo industry pains. OpenAI faces parallel heat. Users demand the full model, not throttled shadows.
Anthropic insists no nerfing. Bugs, pure and simple. They reversed the reasoning cut by April 7. Effort scores jumped from 20 to 85. One X tip: force it via custom instructions. ‘Always reason thoroughly. Treat every request as complex.’ Simple fix. Yet silence fueled conspiracies—hoarding power for Mythos, the next beast.
Reality check. AI isn’t just weights. It’s prompts, caches, configs. Fragile stack. One tweak ripples. Benchmarks hold steady on core smarts. Software shadows the shine. Developers adapt. Add those lines. Monitor logs. Push for transparency.
Outages linger. April’s Opus 4.7 dropped amid nerf claims—better vision, coding, but token thirstier. Platformer again. Users test relentlessly. Trust rebuilds slowly. Claude rebounds. But the scare warns: even top models trip on their own scaffolding.
WebProNews is an iEntry Publication