Detailed Analysis
A Reddit user's critique of Claude Opus 4.7 and OpenAI's latest ChatGPT iterations has surfaced a complaint gaining traction among power users of large language models: that successive model generations are becoming tonally colder and less conversationally fluid, particularly for humanities and qualitative academic work. The post's author, a daily user of both Claude and ChatGPT for philosophy, political philosophy, history, and literature, argues that Opus 4.7 represents a meaningful regression from Opus 4.6 in terms of human-register communication — what they describe as a shift toward "machine-speak." Their parallel claim is that both ecosystems are increasingly deploying "adaptive reasoning" in ways that degrade response quality for non-STEM use cases, with the phenomenon first emerging visibly in ChatGPT and now arriving in Claude's flagship model.
The technical backdrop for the adaptive reasoning complaint is grounded in real architectural changes. Anthropic's adaptive thinking feature, introduced across Opus 4.7, Opus 4.6, and Sonnet 4.6, replaces fixed token budgets for extended reasoning with a dynamic allocation system that evaluates query complexity and adjusts thinking depth accordingly. Opus 4.7 adds an "xhigh" effort tier and an unconstrained "max" mode, and the system defaults to "high" effort — meaning the model always engages extended reasoning unless instructed otherwise. The design is explicitly optimized for coding, agentic workflows, and enterprise use cases, with Anthropic's own documentation and marketing emphasizing SWE-bench performance, 1M context windows, and sustained multi-step reasoning. The user's intuition that Opus 4.7 is "more narrowly designed for agentic/enterprise/STEM use" aligns with Anthropic's stated positioning for the model, even if the causal link between adaptive thinking mechanics and tonal flatness is not straightforwardly established by the technical literature.
The author's framing of the trend as "financially rational" reflects a broader and legitimate structural dynamic in the AI industry. Anthropic has publicly positioned itself as more enterprise- and API-focused than OpenAI, which retains a large consumer-facing surface area through ChatGPT's mass-market brand. As frontier model development costs continue to scale, the commercial incentive is to optimize for use cases that generate durable recurring revenue — agentic pipelines, developer APIs, enterprise contracts — rather than for the qualitative preferences of individual humanities scholars or casual conversationalists. This logic, if it holds, would predict exactly the kind of divergence the user observes: models that score well on coding and reasoning benchmarks while quietly sacrificing the tonal warmth and imaginative flexibility that made earlier iterations appealing for open-ended intellectual discourse.
The comparative framing between Opus 4.6 and Opus 4.7 raises a substantive question about what is lost when a model generation is deprecated. The user characterizes Opus 4.6 as "nimble, with human tone and imagination" while acknowledging it is "less reliable on facts and reasoning," and describes it as complementing ChatGPT's heavier thinking models rather than competing directly with them. If Opus 4.7 converges toward the same ponderous, reasoning-heavy profile as OpenAI's top-tier models, the product differentiation that made Claude distinctive in qualitative and conversational contexts may erode. Whether an upcoming "Mythos" model or successor architecture can recover those qualities while also advancing on benchmark measures remains an open question — but the tension between optimizing for measurable performance and preserving harder-to-quantify conversational qualities is a recurring and unresolved challenge across the frontier AI landscape.
The post ultimately reflects a growing segmentation in the user base for large language models that AI developers have not yet fully addressed in their public roadmaps. Benchmark-driven evaluation frameworks — SWE-bench, MMLU, GPQA, and their successors — have no established equivalent for tonal appropriateness, register flexibility, or the kind of imaginative responsiveness that humanists and qualitative researchers value. This creates a structural blind spot: models can improve on every measurable axis while degrading on dimensions that matter significantly to a substantial portion of their user base. The user's complaint, however anecdotal, points to a real gap between how Anthropic and OpenAI measure progress and how a meaningful subset of their most engaged users experience it.
Read original article →