4 Things That Changed in Claude Opus 4.8 | Choosing AI Effort Levels | (Claude, Opus 4.8, AI)
Claude's new AI, "Opus 4.8," has arrived. Let's briefly summarize what has changed.
In short, what has changed?
On May 28, 2026, Anthropic released a new AI called "Claude Opus 4.8."
The impression is that the pricing remains the same as the previous version, while the internal capabilities have been quietly upgraded. It seems that "intelligence," "honesty," and "speed" have all improved together.
Highlight 1: You can choose the AI's "effort level"
Personally, the most interesting feature is the "effort" setting.
You can now choose in stages how much effort you want the AI to put into thinking. For Opus, there are 5 levels: Low, Medium, High, Ultra-High, and Maximum, while Sonnet has 4 levels without "Ultra-High," so the range seems to vary slightly depending on the model you choose. The default for both was "Maximum."
I feel that being able to switch between quick responses for light questions and higher effort for deep thinking can also help reduce credit consumption. Moreover, it seems to be available on all plans, not just special ones.
Highlight 2: Fewer "lies" and more honesty
When using AI, we are sometimes troubled by plausible-sounding lies (hallucinations).
This area has reportedly been improved, and the probability of silently overlooking mistakes has been reduced to roughly one-fourth of the previous version. The explanation was that it has moved closer to saying "I don't know" when it doesn't know something.
If I don't have to doubt everything every time, I feel this is a welcome change. If it stops writing incorrect code, there will be no need for retries, which I hope will lead to credit savings.
Highlight 3: Faster and capable of reading longer texts
In addition, speed and reading capacity seem to have improved.
In Claude Opus 4.8, a "Fast mode" has been prepared for the API, and you can expect much faster responses than usual. However, this is not a feature that anyone can use in standard Claude.ai chats; at this moment, it is positioned as a Research Preview for the API.
Also, supported models can now handle long contexts of up to 1 million tokens, making them suitable for reading long manuals, documents, or large codebases all at once. However, the scope of use varies depending on the model, plan, and usage environment.
Highlight 4: More features for developers have also been added
This part is a bit technical, so I'll keep it brief.
For programmers, it seems they've added features that allow for large tasks to be completed automatically from start to finish. This isn't something I have much experience with yet, so I'll touch on it again if I have the chance to use it.
Actually trying it out
I immediately tried comparing the same question (propose 3 dinner menu ideas) across the 5 levels of effort.
Honestly, for light questions like this, I didn't feel a huge difference even when increasing the settings. It felt like just a bit more information was added to the response, while the content of the suggestions themselves remained similar. I feel like keeping it low for daily use and only raising it when it really counts is probably the right balance.
On the other hand, I haven't tried it yet for more elaborate tasks like creating games or apps. I imagine that the difference in effort levels might become more apparent with such complex tasks. I'd like to verify how much and in what ways things change as I continue to use it.
Summary
To summarize roughly, these were the updates.
・Price remains the same, but the content has been improved
・You can choose the AI's effort level in stages
・Fewer lies and missed errors, making it more honest
・Increased speed and longer readable text length
I've only just started using it, so if I notice anything else as I continue, I'll be sure to add it.
If you'd like, I'd be happy if you could follow or like this post. I'll continue to record my casual experiments with AI little by little.
Thank you for reading. See you in the next article.
#AI #noteupdate #Claude #GenerativeAI #TriedWithAI #AINews #Anthropic