Claude Opus 5 got better once reviewers deleted their setup

Anthropic released a frontier model on a Friday afternoon, which tells you something about how fast this summer is moving.

@claudeai announced Claude Opus 5 with a simple pitch: it "comes close to the frontier intelligence of Fable 5 at half the price." That works out to $5 per million input tokens and $25 per million output, the same price as the Opus 4.8 it replaces, with a one million token context window. The company posted that it went live the same day across paid plans and the API, as the default model on Claude Max, and in a Fast mode running around 2.5 times the normal speed.

Launch-day claims from a lab are marketing until somebody outside the building checks them. Somebody did.

The independent numbers mostly held

@ArtificialAnlys, which benchmarks models for a living and evaluated this one before release, called Opus 5 "narrowly the most intelligent model" on its Intelligence Index, with comparable intelligence to Fable 5 "at 26% lower Cost per Task." At maximum effort it scored 61 on that index, effectively tied with Fable 5 at 60, and just ahead of GPT-5.6 Sol at 59 and Kimi K3 at 57.

Read those numbers again. The top four models on a respected independent index sit within four points of each other. The real spread is in what each one costs to run. That is why the loudest number in the launch post was a price, not a benchmark.

The most useful review was the least flattering

Then came the post actually worth reading. @danshipper of Every, whose team spent a week putting Opus 5 through coding, writing, knowledge work, and their own internal agent, called it "a hard model to love." He described a model that argued with instructions, stopped before the work was finished, and did not play well with the skills and plugins his team had already built.

That is a bad review. But buried in it is the most interesting thing anyone published this week. His team deleted their existing skills and started from scratch. Stripped of the elaborate workflows built for earlier models, Opus 5 got dramatically better.

The scaffolding was the problem. Not the model.

Your workarounds have an expiration date

This generalizes well beyond one lab's release. Every prompt template, rule file, custom instruction, and clever workaround you have written is a fossil. It records exactly what a specific model was bad at on a specific date. You wrote "always double-check your math" because the model was bad at math. You wrote six paragraphs of formatting rules because it would not hold a format. Those instructions felt like accumulated wisdom. They were patches.

When the underlying model stops having the weakness, the patch does not become neutral. It becomes friction. You are now spending tokens and attention telling a capable model to compensate for a flaw it no longer has, and in the Every team's case, actively making the output worse.

The effort setting points the same direction. Anthropic ships a dial that runs from low up to max, and Opus 5 defaults to high. The instinct is to crank it. Shipper's team found the opposite, that lower effort levels often produced better results, since more thinking time gave the model more room to do the annoying things. More compute is not automatically more useful.

What this means if you run AI in your business

You probably do not care which model tops which index. Here is the part that lands on your desk.

If you put an AI process into production more than a few months ago, some of it is now dead weight. Three things worth reviewing:

• Instructions that fix old weaknesses. Try removing them and compare the output honestly, rather than assuming the longer prompt is the safer one.
• Settings turned up to maximum out of caution. Test the middle. You may be paying multiples for a worse result.
• Tools bolted on to patch a gap. Some of those gaps closed while you were not looking.

The deeper point is that a working AI process is not a thing you install once. It is closer to a piece of equipment that needs a service interval. The model underneath it changes every few weeks, and a setup that was tuned well in March can be quietly costing you money in July.

That is also why we tell people not to start with the tool. Start with the process. The two or three repeatable things eating your week are stable, and they are worth automating no matter which model is on top this month. If you want an outside read on where those are in your operation, New Face Design does a free process audit for businesses in the Fox Valley and beyond. We map what your week actually looks like and point at the spots where automation pays for itself.

Anthropic shipped a better model on Friday. The more useful lesson came from the people who had to throw out their old one to see it.