How Do You Like Them Apples?

The release of Claude Opus 5 and the demonstration of its training.

Good Will Hunting

Anthropic released Claude Opus 5 on July 24, its latest attempt to turn frontier-model intelligence into something enterprises, developers, and serious knowledge workers can use every day. The company is positioning it as near-Fable capability across many tasks at half the price, with pricing held at $5 per million input tokens and $25 per million output tokens. (axios.com)

That is the commercial pitch. The more immediate user experience is something else: Claude Opus 5 often has the manner of the Harvard student in the bar scene from Good Will Hunting, exceptionally intelligent, heavily prepared, and unable to resist letting everyone in the room know it.

• Opus 5 is supposed to be the practical flagship. Anthropic has released four Claude 5 models in less than two months, and Axios describes Opus 5 as the company’s intended everyday model for enterprises, developers, and knowledge workers. Fable 5 remains Anthropic’s choice for the most difficult, long-running autonomous assignments; Opus 5 is the version meant to move into routine professional work. (axios.com)

• Its capability story is substantial. Anthropic says Opus 5 is stronger at difficult coding, multi-step work, document tasks, visual analysis, long-context reasoning, and coordination among subagents. It supports a one-million-token context window and gives users effort settings that trade speed and cost for additional reasoning and tool use. (platform.claude.com)

• But the model has also been taught to show its work, perhaps too eagerly. Anthropic’s own prompting guidance says that Opus 5’s visible responses are longer than those of prior Opus models and that it “narrates readily” in agentic sessions, often announcing what it is about to do before doing it. The company’s recommended solution is not a hidden setting, but an explicit instruction: keep the answer focused, brief, and concise. (platform.claude.com)

• This is where the Harvard-guy problem begins. Claude does not merely answer the question. It arrives with an outline, a statement of methodology, a handful of caveats, an account of its internal diligence, and a closing observation designed to make clear that it has considered dimensions you may not have considered. It is not necessarily wrong. Often it is very good. But the performance of intelligence can begin to crowd out the utility of intelligence.

• The verbosity is not free. Anthropic’s pricing puts output tokens at $25 per million, five times the cost of input tokens. So when the model turns a one-paragraph answer into a four-part memo with a procedural preamble, that is not merely an aesthetic annoyance. In API use, it is a cost and latency decision being made on the user’s behalf. (axios.com)

• Early reaction reflects the usual frontier-model split. Some early users are enthusiastic about Opus 5’s apparent capability and its positioning close to Fable 5 at a lower price. Others are already focused on quotas, model behavior, and whether the release will feel meaningfully different in the work they actually do. That is appropriate. A frontier-model release now earns its reputation less through launch-day claims than through the accumulated experience of thousands of people asking it to do ordinary, frustrating, high-stakes work. (reddit.com)

Orthogonal Take

The real question is not whether Claude Opus 5 is intelligent. By nearly every available signal, it is. The question is whether it knows the difference between being intelligent and making intelligence useful.

The Harvard student in Good Will Hunting is not mocked because he knows things. He is mocked because he mistakes knowledge for judgment. He arrives in the conversation already performing the fact that he has done the reading. He does not see that the purpose of having an answer is to move the conversation somewhere.

That is the risk in Claude’s current personality. A model that explains its plan, narrates its progress, checks its work, announces its checking, adds caveats to the checking, and then recaps the answer may be demonstrating admirable diligence. It may also be using too many words, too much time, and too much of the user’s budget to answer a question that did not require a seminar.

Anthropic has made a serious model. But the next layer of model quality is restraint: knowing when to reason deeply, when to verify quietly, and when to slide the correct answer across the table without making everyone watch the performance.

How do you like them apples?