Why I Went Back to Opus 4.6
By Miguel Torrez, founder of Genial Agency
You will understand why I moved back from Claude Opus 4.7 to Opus 4.6 (the new adaptive thinking mode, not the model's intelligence), how to switch back and control reasoning yourself, and how to test any new model before relying on it.
What you need
A Claude account (claude.ai) with access to Opus 4.7 and the older Opus 4.6
One or two real tasks from your usual workflow to compare the models on
Step by step
Take launch benchmarks with a grain of salt
When a provider publishes benchmarks showing its new model beats Opus 4.6, Gemini or GPT, treat them with caution. You do not know how things were measured behind the scenes. Opus 4.7 is a good model, but being the latest is not a reason to adopt it immediately.
Remember that a smarter model is not automatically better
Any current model can already handle most cognitive tasks: analysis, data work, dashboards. Most real productivity gains come from connecting your model to the world (the features around it), not from the model itself. Judge an upgrade on your own results.
Spot adaptive thinking in the Claude app
Open the Claude application and select Opus 4.7. Look at the thinking setting: it is now called adaptive. This applies both in the chat application and in Claude Code. It is the main reason 4.7 can underperform 4.6 for you.
Understand what adaptive mode takes away
Claude reasons with hidden reasoning tokens before answering. With 4.6 you could tell it to put as much reasoning as possible into a hard task. With adaptive mode, Claude classifies your question as easy, medium or hard itself and may take the easy route even when you want deep thinking.
Know where you can still force more reasoning
In the chat application you can let Claude choose its own reasoning level, but you cannot force it to think a lot. In Claude Code you can. If you need heavy reasoning on hard tasks in chat, 4.7's adaptive mode may hold you back.
Switch back to Opus 4.6 with extended thinking
In the model picker, go to the older models and select Opus 4.6. Its setting is called extended thinking, which lets you choose when the model reasons more or less for the task. This is what I went back to because adaptive mode was hurting my usual workflow.
Note when your model performed best
Keep track of the dates you used a model and how well it did. The same model family can behave differently across versions. When you build with these models, they carry timestamps (dated versions), so you can select the exact version that gave you your best performance.
Test every new model for two to three days
When a new model drops, run your own tests on your real work for at least two to three days before drawing conclusions or deploying it. Do not fall into the AI hype: newer does not always mean better.
Watch out for
Adopting a model just because it is the newest or tops the provider's own benchmarks.
Asking Opus 4.7 for deep reasoning in the chat app: adaptive mode decides the effort itself and may pick the easy route.
Judging a new model in a day: I say you need at least two to three days of tests before concluding.
Not noting when a model worked best, so you cannot tell whether a change in quality came from a different dated version.
