Honestly, I'm not usually the type to get excited about AI model updates. Cheering for "Oh, 0.3% performance improvement!" isn't really my thing—my sensibilities have grown too numb for that. But Claude Opus 4.6 is different.

This isn't just a minor patch; it's the kind of upgrade that makes you think, "Ah, finally, this is actually usable."

What the Numbers Tell Us (and What They Don't)

Let's look at the benchmark scores. It's 144 Elo points ahead of GPT-5.2 and a whopping 190 points higher than its predecessor, Opus 4.5. But honestly, how many people actually know what an Elo point is? (I know it's a chess ranking system, but I have no intuitive sense of how significant a 190-point difference is in the AI context.)

The real eye-opener is elsewhere. On the MRCR v2 test, it scored 76%, compared to Sonnet 4.5's 18.5%. This isn't just "improvement"—it's a different league entirely. The fact that its long-context retention ability has improved this much means you no longer have to ask the AI, "Remember what I said earlier?"

1 Million Token Context Window: What's the Point?

They've added a 1M token context window (beta). If you can't grasp how big that is, consider that an average novel is about 100,000 tokens. In other words, it can remember and converse with the equivalent of 10 novels' worth of information simultaneously.

You might ask, "So what?" How often do you really need to throw 10 novels' worth of content at an AI? But when you actually use it, those moments come. Analyzing large codebases, reviewing multiple lengthy research documents simultaneously, maintaining all the context for complex projects... In these tasks, the elimination of "context rot" (the phenomenon where it forgets what it said earlier) is a genuine game-changer.

Adaptive Thinking: The AI Now Knows "How Much to Think"

One of the most interesting features is "adaptive thinking." Simply put, the AI judges the difficulty of a problem and decides how deeply to think about it.

There's no need to spend five minutes pondering a simple greeting like "Hello," right? But a question like "How should I migrate this legacy codebase?" genuinely requires deep thought. Opus 4.6 understands this difference.

Of course, sometimes it might overthink things. That's why they've added the /effort parameter. You can directly control "how hard to think" with four levels: low, medium, high, and max. It's like telling a student, "You can phone this assignment in" or "Really put your heart into this one."

Real-World Performance: Beyond the Numbers

Benchmarks are just benchmarks. What really matters is how useful it actually is.

Looking at early test results:

• Better results than the previous version in 38 out of 40 cybersecurity investigations

• Completed multi-million-line code migration in half the time

• 90.2% accuracy in legal and financial content analysis (BigLaw Bench)

The improvement in coding tasks is particularly noticeable. Its self-error detection and correction capabilities have been enhanced, so when you ask "Why isn't this code working?" it can actually find the problem and fix it. It no longer feels like a debugging assistant—it feels like an actual debugger.

Safety: Getting Smarter While Staying Safe

As AI becomes more powerful, safety concerns naturally grow. But what's interesting about Anthropic is that they've achieved both performance and safety improvements simultaneously.

It has the lowest over-refusal rate among Claude models. This means there are fewer cases of it refusing legitimate requests as "unsafe." At the same time, it's better at refusing actually dangerous requests.

If you understand how difficult this balancing act is... When you ask the AI to "find code vulnerabilities," it's that delicate judgment of actually finding vulnerabilities without teaching you how to exploit them, rather than refusing with "That's hacking, so I can't help."

Pricing: Performance Up, Price Stays the Same

The price remains unchanged from before: $5/$25 per million tokens. Premium pricing ($10/$37.50) kicks in beyond 200k tokens, but considering you get access to the 1M context window, it's not bad.

This is actually surprising. Usually, the business logic is "it's better, so it should be more expensive," right? But they've significantly boosted performance while keeping the price the same. It's evidence of fierce competition, and from a user's perspective, it's just good news.

My Experience Using It

I've been using it for a few days. What impressed me most was that it "doesn't lose context." Even in long conversations, even in complex multi-step tasks, it remembers the details I mentioned at the beginning all the way through.

For example, when writing a blog post, if I say "Oh, explain that in the context of that project I mentioned earlier," it actually remembers what I said earlier and explains accordingly. Previous versions had this "uh... what did you say again?" feeling in the middle, but not anymore.

Code work is also much smoother. Especially in tasks like large-scale refactoring or migration, its ability to make changes "while maintaining the overall structure" has improved. It's not just changing A to B—it understands "why it was A" and considers "where changing it to B will have impact."

Still Not Perfect (And That's Normal)

Of course, there are limitations. It still makes weird mistakes sometimes and occasionally gives overly confident answers. Adaptive thinking sometimes thinks too deeply and gives long answers to simple questions.

But hey, there's no perfect tool. What matters is "Is it generally useful?" And the answer is "definitely yes."

Conclusion: AI Finally Feels Like a Real Assistant

Claude Opus 4.6 isn't simply a "better version." It's closer to "the proper version at last."

Long-context retention, autonomous problem-solving, situation-appropriate thinking depth adjustment... When these things come together, AI feels like a "real assistant" for the first time. It's no longer "Should I ask the AI? Or should I just do it myself?" but rather "Let me just hand this to the AI."

It's not perfect. It will continue to improve. But right now, at this moment, Opus 4.6 is the most "practical" AI model I've used.

And honestly, isn't that everything? What matters more than benchmark scores is "how often you actually use it."