Claude Opus 5: 98% of Flagship Power for Half the Price

William Palmer

So I was watching this video clip the other day and it is the absolute perfect snapshot of where we are with AI right now. This developer gives the new Claude Opus 5 a single prompt, walks away, and the model just... takes over the browser. It navigates to the OpenAI site, logs in, clicks through the settings, and actually cancels his ChatGPT Pro subscription. Talk about a brutal product demo.

William Palmer

But here is the really fascinating thing. Anthropic is being incredibly modest about this model. Their official release notes say it only quote, "comes close" to their absolute top-tier model, Fable 5. But the developer community is losing its mind because, in practice, this thing feels like a massive leap forward. And when you look at the economics, you start to see why everyone is buzzing. This is a July 2026 launch, and they have priced Opus 5 at five dollars per million input tokens and twenty-five dollars per million output. Now, that might sound like a bunch of abstract numbers, but here is the key: that is exactly half of what it costs to run Fable 5. Literally fifty percent off.

William Palmer

And yet, if you look at the hard data from Epoch, their Capabilities Index—the ECI—scores Opus 5 at 159. Fable 5 is at 161. That is a microscopic two-point difference. And on the software engineering-specific benchmark, the SWE-ECI? It is a dead heat. Both models score exactly 161. So you are getting 98% of the flagship's brainpower for half the price. It makes you wonder why Anthropic is being so quiet about it, right? Well, they know that Fable still has what developers call that "big model smell"—this elusive, heavy-parameter coherence over massive, complex projects that is incredibly hard to measure in a standard test but you know it when you see it.

William Palmer

Think of it like race cars. I love getting out on the track, and this is the perfect way to picture it. Fable 5 is like a heavy, raw V12 supercar. It has got massive, brute-force torque, and on a long, straight highway, it is just going to pull away. It has got that weight. But Opus 5? Opus 5 is a lightweight, highly tuned track car. It is stripped down, it is agile, and on a tight, technical autocross circuit—like coding and tool use—it is going to pull the exact same lap times as the V12, but it uses a fraction of the fuel to do it. It is just a highly optimized machine for the specific work most people actually need to do day-to-day.

William Palmer

But when you start pushing this lightweight track car to its absolute limits, you run into some really bizarre behavior. Evaluators testing Opus 5 on the FrontierCode benchmark noticed this massive anomaly. The model actually performed better on coding tasks when it was set to "medium effort" than when it was allowed to use "high effort." Now, that is deeply weird. In machine learning, the golden rule lately has been that if you let a model think longer—if you give it more "test-time compute" to search for answers—it should get smarter. More effort should equal better results, right?

William Palmer

But it turns out, just like a human driver overthinking a simple corner, models can get in their own way. When you force a reasoning engine to spend extra compute on a relatively straightforward software engineering task, it starts second-guessing itself. It overcomplicates elegant solutions. It gets stuck in these endless, pedantic self-correcting loops where it edits perfectly functional code until it actually breaks it. It falls into what we call search-space traps, chasing its own tail because it has been told it *must* keep thinking. Sometimes, a quick, intuitive reaction is just better than a five-minute internal debate.

William Palmer

Now, where this optimization really pays off is in how they have tuned the safety systems. This is huge for anyone doing security work. Anthropic designed Opus 5's cybersecurity classifiers to intervene roughly 85% less often than they do for Fable 5. If you have ever tried to use a frontier model to scan proprietary source code for vulnerabilities, you know the pain—the older models would constantly panic, thinking you were trying to build a cyberweapon, and just shut down the conversation. With an 85% reduction in those false-positive blocks, defenders can actually use this thing for deep security audits, while the model still holds the line against generating actual, malicious binary exploits.

William Palmer

What this launch really shows us is that the frontier of AI is shifting. It is no longer just a brute-force war of who can build the biggest, most expensive V12 engine. The real battle is about efficiency, tuning the search effort so the model does not overthink, and relaxing the safety friction so developers can actually get work done. Delivering that 98% flagship performance at a daily-driver price point? That is how you actually win the race. Anyway, that is the state of the track today. Catch you next time.