For weeks, a rising variety of builders and AI energy customers have been claiming that Anthropic’s flagship mannequin is shedding its edge. Customers on GitHub, X, and Reddit reported a phenomenon described as follows: "AI shrinkflation"—Claude was famous to have a deterioration during which he appeared to have decreased steady reasoning means, elevated susceptibility to hallucinations, and elevated waste of tokens.
Critics argued that the mannequin had moved on from its earlier state and pointed to seen modifications in conduct. "Analysis first" strategy to lazy folks, "edit first" This model was now not dependable for advanced engineering.
The corporate initially pushed again towards the next claims, however "nerf" Fashions for managing demand, rising proof from distinguished customers and third-party benchmarks have created a big belief hole.
As we speak, Anthropic instantly addressed these issues and revealed a technical autopsy that recognized modifications in three product tiers that have been answerable for the reported high quality points.
"We take studies of degradation very severely." Learn Anthropic’s weblog submit on this difficulty. "We by no means deliberately degraded the mannequin, and we might rapidly affirm that the API and inference layer have been unaffected."
Anthropic claims to have resolved the problem by altering the inference work and reverting redundant prompts whereas fixing the caching bug in model v2.1.116.
Proof of continued deterioration
This debate gained momentum in early April 2026, fueled by detailed technical evaluation by the developer neighborhood. Stella Laurenzo, senior director of AMD’s AI group, revealed a radical audit of 6,852 Claude code session recordsdata and over 234,000 device calls on Github, exhibiting worse efficiency than in earlier utilization.
Her findings steered that Claude’s depth of reasoning declined quickly, resulting in reasoning loops and an inclination to decide on. "best repair" Not the proper one.
This anecdotal grievance seems to have been verified by third-party benchmarks. BridgeMind reported that Claude Opus 4.6’s accuracy dropped from 83.3% to 68.3% in testing, and its rating plummeted from 2nd to tenth.
Some researchers have argued that these explicit benchmark comparisons are flawed resulting from inconsistent check protection, however Claude’s argument is that "silly" It grew to become a viral subject. Customers additionally reported that their utilization limits have been being depleted quicker than anticipated, resulting in suspicions that Anthropic was deliberately throttling efficiency to deal with surging demand.
trigger
Anthropic revealed in a submit after Morem Swamp that whereas the burden of the underlying mannequin had not been regressed, three particular modifications had been made to the mannequin. "harness" Issues across the mannequin have been unintentionally hindering efficiency.
Default Inference Effort: On March 4th, Anthropic modified the default inference effort for Claude Code from Excessive to Medium to deal with UI lag points. This alteration was meant to forestall the interface from being displayed. "frozen" The mannequin thought, however the consequence was a big lower in intelligence for advanced duties.
Cache logic bug: Cache optimizations aimed toward eliminating stale ones shipped on March twenty sixth "thought" Errors from idle periods contained a essential bug. The mannequin’s thought historical past is misplaced as a result of the thought historical past is cleared on every subsequent flip as a substitute of being cleared as soon as after an hour of inactivity. "quick time period reminiscence" They change into repetitive and forgetful.
Redundancy Limits for System Prompts: On April 16, Anthropic added directions to system prompts to maintain textual content between device calls to lower than 25 phrases and last responses to lower than 100 phrases. This try to cut back redundancy in Opus 4.7 backfired, leading to a 3% drop in coding high quality scores.
Impression and future safety
The standard difficulty prolonged past the Claude Code CLI, impacting the Claude Agent SDK and Claude Cowork, however not the Claude API.
Mr. Anthropic mentioned these modifications made the mannequin "intelligence turns into decrease," They acknowledged that that is not the expertise customers ought to anticipate.
To revive person belief and stop future regressions, Anthropic is implementing a number of operational modifications.
Inner dogfooding: Extra inner employees is predicted to make use of public builds of Claude code precisely to make sure the identical expertise of the product because the customers.
Enhanced analysis suite: The corporate will now run a broader analysis suite for every mannequin, "Ablation" Isolate the results of particular directions for every change in system prompts.
Tighter controls: New instruments are constructed to make it simpler to audit speedy modifications, and model-specific modifications are strictly restricted to the meant goal.
SUBSCRIBER COMPENSATION: Given the token wastage and efficiency friction brought on by these bugs, Anthropic has reset spending limits for all subscribers as of April twenty third.
The corporate plans to make use of the brand new @ClaudeDevs account on X and GitHub threads to supply deeper reasoning behind future product selections and keep a extra clear dialogue with its developer base.
