What got better?
Anthropic emphasizes coding, knowledge work, and long, multi-step agent tasks. The model is said to check its own work more often and stop less in the middle of a complex task.
According to the company's own and partner measurements, it made significant progress in several tests examining coding and office work. Important small print: most of the results were published by the manufacturer himself, so it is worth waiting for independent measurements.
What does the effort setting mean?
Developers can control how much "thinking work" the model puts into a response. It can be faster and cheaper at a lower level; at a higher level, he continues to work on the more difficult problem. In Hungarian: it has a "say it quickly" and "sit down, think properly" switch.
How much does it cost?
The base price of the API is $5 per million input tokens and $25 per million output tokens. Fast mode promises about 2.5 times the speed, but for twice the base price.
Who might be interested?
Primarily for programmers, researchers and companies where AI does not have to give a single answer, but must handle documents, tables, code and tools in several steps. It's probably too much to write a shopping list — unless the list is a complete supermarket redesign.
What makes this more than a simple model update?
For long tasks, it is not only important that the model gives a good answer once. It's more important to stay on target over several steps, spot your own mistake, and then correct it without going in a completely different direction. That's why Anthropic emphasizes verification and retry: a complex program, research brief, or corporate workflow is rarely perfect the first time around.
However, the higher effort setting is not magic. It can mean more calculation time and generally more costs, so it is worth choosing according to the difficulty of the task. A short email doesn't need max level, but when debugging a large code base, a more patient mode may be worth it.
What is worth observing in your own test?
Don't just look to see if the first answer is pretty. Give him a task in which he has to plan, execute and check. Then ask him to list what he couldn't prove for sure. This way, it quickly becomes clear whether he is really more thorough, or whether he is just expressing himself more confidently.
Claude Opus 5's first day at work
At 8:00 you will receive the entire codebase. At 8:02 you find the error. At 8:04 a.m. he fixes, tests, documents, and creates a recovery plan for it. At 8:05, the boss leans in: "Very good, but I just wish the button was a little more blue."
Claude politely resets the button, then analyzes in max effort mode on twelve pages what exactly "a little bluer" might mean. The Fast mode, meanwhile, does the same thing two and a half times faster, but at twice the price – in other words, it works exactly like the emergency mechanic on a Sunday night.