Make money doing the work you believe in
I’ve been a heavy Claude user for a long time, especially for writing, architecture, debugging, and production development. I’m also building an ML-heavy, infra-heavy IQA platform with VLM enrichment, so my usage is not casual. I depend on these tools for real engineering work.
That is why the contrast between Opus 4.6 with 1M context and Opus 4.7 has been so frustrating.
Opus 4.6 was excellent for my workflow. It handled long context well, stayed aligned with the task, and felt like a genuinely strong co-pilot for complex work.
Opus 4.7 has been a huge disappointment. Over the past couple of weeks, the experience has clearly degraded: weaker context retention, more hallucinations, and less consistency. It feels like a regression rather than an improvement.
Today was a particularly disappointing example. I stepped away from an in-progress task, came back later, and asked what was currently running. Instead of understanding the task state, Claude ran ps and listed services. That kind of response is not what I expect from a model I rely on for production work.
Due to these inconsistencies, context loss and hallucinations, I had to temporarily use Cursor and explore GPT-5, Gemma 4, and Qwen Coder just to keep production moving. That was not the plan, but being blocked in the middle of delivery is worse.
What makes this harder to accept is that the degradation has continued for more than two weeks. At some point, users stop wanting promises about the next big thing and start wanting the current product to work reliably.
The message is simple: fix what is broken before celebrating what comes next.
Here are my couple of learnings shared on working with Opus 4.6 and Opus 4.7
