← Reddit

Opus 4.7 is a regression on claude.ai, but improvement on claude code.

Reddit · Familiar_Text_6913 · April 17, 2026
Opus 4.7 improved performance on Claude Code but regressed on claude.ai compared to version 4.6, with adaptive thinking paired with 4.7 proving unreliable compared to extended thinking with 4.6. Without thinking capabilities, the Opus model performs poorly. Users experiencing this degradation remained on version 4.6 and reconsidered their subscriptions.

Detailed Analysis

Anthropic's Claude Opus 4.7, released on April 16, 2026, has produced a sharply divided reception depending on how users access it — a divergence that reflects deliberate architectural trade-offs rather than a simple across-the-board upgrade. On claude.ai, the model's general-purpose conversational interface, users report a noticeable regression in reliability and reasoning quality compared to Opus 4.6, while developers using Claude Code — Anthropic's agentic coding product — are seeing some of the most significant benchmark improvements in the Claude lineage. Opus 4.7 jumped from 80.8% to 87.6% on SWE-bench Verified and from 53.4% to 64.3% on SWE-bench Pro, surpassing GPT-5.4 and Gemini 3.1 Pro in head-to-head comparisons. Cursor, a prominent AI coding tool, reported their internal benchmark rising from 58% to 70% after integrating Opus 4.7.

The regression on claude.ai is not anecdotal — it is corroborated by specific benchmark data. Long-context retrieval dropped dramatically from 91.9% under Opus 4.6 to just 59.2%, a collapse that would directly impair users relying on the model for document analysis, extended research threads, or complex multi-turn conversations. BrowseComp, which measures multi-step web research and synthesis, also fell by 4.4 points. The model's new "adaptive thinking" feature — designed to allocate more cognitive effort to harder tasks — appears to introduce inconsistency in everyday claude.ai use cases where Opus 4.6's more stable "extended thinking" mode had been reliable. Users also note that Opus 4.7 exhibits more literal instruction-following and less spontaneous tool use, behaviors that can benefit structured coding workflows but feel brittle in the open-ended conversational contexts that define typical claude.ai sessions.

The trade-off reflects a broader strategic realignment at Anthropic. Opus 4.7 was clearly engineered around professional software engineering, enterprise agentic workflows, and precision-heavy tasks — domains where it delivers genuine gains in visual reasoning (CharXiv rising from 69.1% to 82.1%), financial analysis (General Finance benchmark: 0.813 vs. 0.767), and legal work (90.9% on BigLaw Bench). The model's 3x increase in image resolution support (up to 2576px / 3.75MP) also signals a push into vision-heavy enterprise applications. These are high-value, high-margin use cases, and Anthropic's investment in Claude Code as a product — deeply integrated with tools like Cursor — suggests the company is prioritizing the developer and enterprise market over the general consumer experience on claude.ai.

For individual subscribers, the practical implication is a product that no longer serves all use cases equally well. The user's decision to remain on Opus 4.6 and reconsider their subscription highlights a real tension Anthropic faces: as models become more specialized and task-optimized, the one-size-fits-all promise of a flagship "Opus" tier becomes harder to sustain. Versioned access and the ability to pin to older models provides a short-term workaround, but it also signals that the model roadmap is no longer tracking uniformly forward for all users. The gap between Opus 4.6 and 4.7 on long-context retrieval — more than 32 percentage points — is large enough that it is almost certainly a deliberate architectural choice rather than an oversight, suggesting Anthropic accepted this trade-off knowingly.

This pattern mirrors broader trends in frontier AI development, where leading labs are increasingly shipping models tuned for specific workflows rather than general-purpose excellence. OpenAI's o-series models, for instance, prioritize deep reasoning chains at the cost of conversational fluidity, and Google's Gemini lineup has similarly fragmented into task-specialized variants. Anthropic's move with Opus 4.7 suggests the era of a single model serving as the best option across all contexts may be ending — replaced by a portfolio approach where users must actively select the right model for their use case. For Claude, this means that "Opus" is no longer a proxy for "best overall," but increasingly a signal of "best for agentic coding and precision enterprise work," a distinction that Anthropic will need to communicate more clearly to avoid subscriber churn among its broader user base.

Read original article →