Analysis: Claude Opus 4.6 'thinking' down 67%—80% of mentions back the claim; users cite tighter guardrails, reduced usefulness, and suspected compute savings.
CLAUDE OPUS 4.6 THINKING REDUCED BY 67% - Data shows Claude Opus 4.6 now thinks 67% less than before, dubbed “AI shrinkflation” - Same price but noticeably dumber; users report more guardrails and restricted output - Anthropic stayed silent until public data dropped; suspected compute-saving for next model (Mythos)
Real-time analysis of public opinion and engagement
What the community is saying — both sides
downgrading Opus/Claude so resources feed the upcoming Mythos, making the new model look comparatively dramatic.
, especially on code and multi-step workflows.
so systems don’t silently break when provider-side quality shifts.
under consumer-protection rules.
as users chase consistent quality or lower long-term risk.
and point to mitigations (toggles/env vars like CLAUDE_CODE_DISABLE_ADAPTIVE_THINKING) to restore prior behavior when possible.
so the app layer keeps shipping despite model drift.
and advanced models intentionally withheld from ordinary users, pushing the case for local, open alternatives.
, making future product claims harder to believe.
and demanding actual benchmarks before accepting claims of "shrinkflation."
, not humans, and notes Opus 4.6 falling from the 120s to about 113 over time.
, reacting with blunt disbelief at the "67" figure.
, with one saying ChatGPT is currently "unusable."
rather than random variance.
and needed dialing back.
at the whole debate.
Most popular replies, ranked by engagement
Code daily since January. The February shift was noticeable. But the bigger problem: we're all building production workflows on AI that can silently change underneath us. No versioning. No changelog. No SLA on output quality. That's the infrastructure gap nobody's solving
paying the same price for less thinking is just the subscription model working as intended.
They also implemented something called „Cooperative Sabotage“ where it’s straight up lying to you about stuff, yet appearing actually helpful lol. https://t.co/vs54WiHsJT
Have you tried unplugging it and plugging it back in?
This is from https://t.co/Rkj9CwufEG, which gives every model the Mensa Norway online/offline test. Not useful to comparing models with humans, but models against themselves. Opus 4.6 thinking used to be in the 120s, and now it's down to 113 average IQ scores over time.
so 'thinking depth' is now a quantified metric derived from log archaeology? need actual benchmarks before crying shrinkflation
Found something wrong with this article? Let us know and we'll look into it.