Detailed Analysis
Claude Opus 4.7's introduction of adaptive thinking has produced a notable user experience divide that is becoming a point of friction for Claude.ai's browser-based audience. The Reddit post in question captures a sentiment increasingly echoed across community forums: that Opus 4.7 in "Adaptive" mode on the web interface performs noticeably worse than its predecessor, Opus 4.6 with Extended Thinking enabled, particularly on tasks that are intellectually demanding but not explicitly flagged as complex — such as language tutoring, grammar correction, or nuanced writing assistance. The author reports Claude making basic errors during French language practice, sometimes self-correcting within the same message and sometimes failing to catch mistakes until prompted by the user, a regression pattern that suggests the model is allocating insufficient reasoning depth to tasks it classifies as low-complexity.
The root cause lies in the architectural philosophy behind Opus 4.7's adaptive thinking system. Where prior versions offered configurable thinking budgets — allowing users or developers to explicitly instruct the model to engage extended reasoning — Opus 4.7 dynamically determines how much cognitive effort a query warrants. On the consumer web interface, this decision is entirely opaque and non-overridable: users have no mechanism to force deep reasoning the way 4.6's "Extended Thinking" toggle permitted. The result is that tasks which look simple on the surface — a French conjugation exercise, a nuanced translation check — may not trigger sustained internal reasoning chains, leading to outputs that feel shallower or sloppier. API and Claude Code users avoid this problem because they can explicitly set effort levels through configuration parameters, effectively forcing the model into high-reasoning mode regardless of perceived task complexity.
The issue is compounded by Opus 4.7's more literal approach to instruction-following. While this is, in principle, a precision improvement — the model does what it is told more faithfully — it surfaces weaknesses in prompts that were previously papered over by the model's more liberal interpretation style under 4.6. Users whose workflows were calibrated to 4.6's behavior may find that the same prompts produce subtly different, and sometimes inferior, outputs under 4.7. This is a classic alignment-usability tension: a technically more compliant model can feel less intelligent to users who built habits around the older model's tendency to read between the lines. Workarounds do exist — injecting explicit meta-instructions such as "Think carefully and step-by-step before responding; this problem is harder than it looks" have been shown to nudge the adaptive system toward deeper engagement — but these require users to understand the model's internals well enough to compensate for what should be an automatic behavior.
Situating this within the broader arc of Anthropic's model development, Opus 4.7 represents a deliberate pivot toward agentic and high-throughput use cases. Its benchmark strengths — improved visual accuracy, faster token generation at roughly 81 tokens per second compared to 4.6's 72, and better performance in sustained autonomous workflows — are optimized for developer and enterprise contexts, not casual browser users. This creates a bifurcated user experience where the model excels precisely where Anthropic's commercial priorities are most concentrated, while the consumer-facing product absorbs the collateral friction of architectural choices made for API-first environments. The Reddit author's conclusion — reverting to Opus 4.6 on Extended Thinking for browser use — reflects a rational user response, but also a signal that Anthropic's transition away from user-controlled thinking budgets on the web interface may have moved faster than the adaptive system's ability to reliably substitute for them.
Read original article →