Detailed Analysis
A Reddit post from a self-identified Max subscriber ($100/month, 5x tier) has become a focal point for user frustration following the launch of Opus 5, raising pointed questions about how Anthropic is managing extended thinking controls across its model lineup. The core complaint centers on a granular but consequential UX regression: Sonnet 4.6 offered a "thinking toggle" where the model could choose to reason, Opus 4.6 offered an "extended" mode that guaranteed deep reasoning on every output, but Opus 5 has reportedly collapsed this into a single "effort levels" control with no explicit guarantee of extended reasoning. For power users who build multi-step workflows, internal scaffolding, or project-based prompt chains that depend on predictable reasoning behavior, this distinction is not cosmetic — it determines whether they can trust the model to reliably apply deep computation without manually babysitting outputs for missed steps or shallow responses.
The complaint also touches on a less quantifiable but frequently cited issue: perceived personality drift. The poster and various forum commenters describe Claude 4.5/4.6 as more pleasant and collaborative to brainstorm with, while Sonnet 5 and Opus 5 are characterized as curter or more resistant during open-ended ideation. This kind of subjective shift matters commercially because much of Claude's differentiation from competitors like ChatGPT and Gemini has rested on its conversational tone and collaborative feel, not just raw benchmark performance. When a model update changes how "agreeable" or "engaged" the assistant feels during brainstorming, it can erode the trust and rapport that heavy users — writers, researchers, coders leaning on Claude as a thinking partner — have built into their workflows, even if underlying capability metrics improve.
The post's most substantive observation is structural: Anthropic's own deprecation policy, which typically guarantees model availability for about a year, has left Claude Sonnet 4.6 and Opus 4.6 both protected from deprecation until February 2027, even as Opus 5 (and previously other models) ship as ostensibly superior successors. The poster speculates this signals that a meaningful share of users and enterprise workflows remain anchored to the 4.6 generation — whether for stability, cost, personality, or the reasoning-guarantee behavior discussed above — and that Anthropic can't yet afford to sunset it. Running multiple "frontier" models concurrently is compute-intensive, so keeping older versions alive despite newer releases suggests either strong customer lock-in to specific model behaviors or internal caution about forcing migration too quickly, both of which point to friction in how confidently Anthropic can deprecate models even when marketing frontier releases as strict upgrades.
This tension reflects a broader pattern across the AI industry as foundation model providers iterate rapidly: shipping cadence increasingly outpaces the ability of power users to validate that new "flagship" releases are unambiguous upgrades along every dimension that matters to them, whether that's raw capability, cost efficiency, reasoning transparency, or interaction style. Anthropic, OpenAI, and Google have all faced pushback when default behaviors, personality, or reasoning controls shift between versions, since professional and prosumer users often build fragile, high-value workflows around very specific model quirks. The fact that a paying subscriber is openly weighing a switch to Gemini — which offers explicit thinking toggles across both free and pro tiers — underscores how competitive the reasoning-model market has become on precisely these usability dimensions, not just leaderboard scores. For Anthropic, sustaining trust with its highest-paying subscriber tier likely requires either restoring granular reasoning controls or communicating more transparently about how effort-level settings map to guaranteed versus optional reasoning behavior.
Read original article →