Claude Opus 4.7 Drops: Anthropic Finally Responds to 'Dumbing Down' Critics

I hesitated before opening Claude last week.

Not because of any sensitive topic, but because the past month has been brutal for Anthropic. Developer communities were flooded with complaints about Claude’s ‘dumbing down.’ Opus 4.6 started strong, but gradually, more users noticed the model becoming increasingly ‘conservative’ on complex engineering tasks—abandoning multi-step workflows mid-stream, delivering answers that looked plausible but were fundamentally wrong.

And let’s not forget Anthropic’s series of baffling decisions: making adaptive thinking the default on February 9, dropping effort levels from maximum to 85 on March 3, and quietly accelerating the 5-hour rate limit consumption on March 26.

So when Claude Opus 4.7 dropped, my first reaction was: finally.

What Actually Improved?

Short answer: coding capabilities got significantly better.

According to Cursor team’s benchmarks, Opus 4.7 hit 64.3% on complex coding tasks—a noticeable jump from 4.6. More importantly, its stability on long-context, multi-file engineering projects improved substantially.

I tested it myself by asking it to refactor a several-thousand-line Python project. Version 4.6 would frequently ‘forget’ its earlier design decisions halfway through. This time, 4.7 completed the task in one go without me reminding it of any context.

Vision capabilities also got a boost. I used to get incorrect component relationship identifications when analyzing technical architecture diagrams with 4.6. 4.7 is noticeably more accurate here.

But the biggest surprise? No price hike. Still $5 per million input tokens, $25 per million output tokens. In an era where everyone is raising prices, Anthropic showed some restraint.

The Truth Behind the ‘Degradation’ Controversy

Now, back to that sensitive topic.

I’ve been using AI coding tools daily for the past month—Claude, GPT, Gemini, domestic models, rotating through them all. Honestly, I felt that frustration of ‘the model got dumber’ firsthand.

But wait—is it really that simple?

Thinking it through, I noticed a few interesting points:

First,所谓的 ‘degradation’ often stems from shifting expectations. Months ago, we wanted AI coding assistants to ‘help complete my code.’ Now we expect them to ‘refactor my entire project.’ Task complexity increased, but the model’s capability boundaries remained the same—creating a perception gap.

Second, Anthropic did adjust the ‘effort level’ parameters, but that doesn’t mean the model itself got worse. More likely: they’re A/B testing or controlling costs. After all, full-power Opus inference is genuinely expensive.

Third, and most importantly: we anthropomorphize AI too easily. Models don’t have ‘intentions’ or ‘laziness.’ Performance variations stem from training data issues, sampling randomness during inference, or context management problems—not from the model ‘deciding’ to underperform.

My Recommendations

If you’re a heavy AI coding tool user like me, here’s my advice:

Don’t put all your eggs in one basket. Claude excels at code understanding and long context. GPT wins on tool ecosystem and plugins. Gemini leads in multimodal. Switch flexibly based on task type rather than loyalty to one model.

Learn to manage context. Even the strongest models have limited context windows. Close irrelevant files, mark key decision points with comments—it significantly improves AI performance.

Finally, maintain reasonable expectations. AI coding assistants are ‘assistants,’ not replacements. They accelerate your work but can’t think for you.

The Opus 4.7 release shows Anthropic is still serious about the product. Whether this wins back user trust depends on sustained performance. I’ll keep watching—and welcome your experiences in the comments.