Claude Opus 5: What Happens When an AI Agent Can Keep Working Longer Than You Can?
Yesterday, July 24, Anthropic released Claude Opus 5. The headline facts are easy to repeat: it is cheaper than Claude Fable 5, stronger than the older Opus model, and designed for everyday use.
But the more interesting question is not how quickly it can answer.
It is what happens when an AI agent can keep working after the first answer.
An AI agent is not only a chatbot. It can take several steps toward a goal: search, read files, use tools, check results, and revise its work. This is often called agentic AI. The model still needs a task, a budget, tools, and permission from people. But it can stay with a problem for longer.
That changes the product. It also changes the risks.
In my earlier post on Claude Fable 5, I argued that powerful AI is becoming the model plus the control layer around it. Opus 5 makes that idea more practical. If an AI can keep researching, checking, and revising, the control layer matters even more.
When an AI can keep working, the main question stops being “Can it answer?” and becomes “What is it allowed to do while it works?”
An Answer Is Not a Work Process
A normal chatbot interaction is simple. You ask a question. It gives an answer. You decide what to do next.
A work process is different. Imagine asking an AI to investigate why a customer report is wrong. It may need to read the report, inspect a spreadsheet, compare records, test an explanation, and prepare a summary. A useful answer may need several attempts.
Anthropic says Opus 5 is built for that kind of long-horizon work. Its technical documentation lists a 1 million-token context window, thinking enabled by default, and tools for changing the amount of effort spent on a task. Those features do not make it an independent employee. They make it better able to hold onto a large task and work through it step by step.
The system card also draws an important line. Anthropic says Opus 5 is not a replacement for a research team and does not show sustained evidence of doubling AI research speed. That is a useful correction to the language of “AI that works while you sleep.” More continuity is not the same as unlimited autonomy.
flowchart TB
Goal["You set a clear goal"] --> Research["AI agent researches"]
Research --> Check["AI agent checks"]
Check --> Revise["AI agent revises"]
Revise --> Review{"You verify final claims<br/>and approve the action?"}
Review -->|Yes| Act["Action"]
Review -->|No| Revise
Act -->|New evidence or next task| Research
Access["Minimum access"] -. applies throughout .-> Research
A simple operating model, not a description of Claude Opus 5's internal design.
When a Second Check Becomes Practical
Anthropic prices Opus 5 at $5 per million input tokens and $25 per million output tokens, the same token price as Opus 4.8. It says the model comes close to Fable 5's frontier intelligence at about half Fable's price.
Price sounds like a small product detail. It is not.
The shift is not that every extra attempt suddenly costs less than it did with Opus 4.8. It is that Anthropic says more capable, near-Fable-level work is available at the established Opus price. That can make it more practical to ask an AI to investigate, test, compare, and improve. That is a second thought: not just producing an answer, but checking whether the answer deserves trust.
This does not mean every task needs maximum effort. Longer work still costs time and money. Anthropic's own documentation says users can choose lower or higher effort levels. The goal is not to make every AI task as long as possible. The goal is to spend more effort where a first answer is not enough.
For ordinary users, that could mean asking an AI to check a travel plan against constraints, compare two documents, or test a piece of code before suggesting a change. For teams, it could mean letting an agent prepare a draft analysis, then requiring a person to approve the action that follows.
More Capable AI Agents Need Better Brakes
More steps create more places to make a mistake.
One important risk is prompt injection. This means hidden instructions inside a webpage, email, document, or tool result. An attacker may try to make an AI ignore the user's real goal and follow the hidden instruction instead.
Anthropic reports that Opus 5 is more resistant to prompt injection than Opus 4.8 across coding, computer-use, and browser-use tests. It also describes product defenses that inspect incoming content and, in some cases, block risky actions before they happen. These are useful brakes because an attack must defeat more than one layer.
But brakes are not a promise that every road is safe. Anthropic's results are evaluations, not proof that prompt injection has been solved. That is why the right habit is simple: give an agent only the access it needs, and require approval before an action is hard to reverse.
This is human-in-the-loop: a person does not need to watch every small step, but stays responsible for decisions that can affect people, data, money, or public systems.
The same distinction appears in cyber work. Anthropic says Opus 5 is closer to Mythos 5 at finding software vulnerabilities, but remains behind Mythos at turning them into practical exploits. That boundary matters. As I wrote in the Mythos Preview post, some powerful cyber capabilities are still being treated as a restricted-access problem, not an everyday feature.
The System Card Reveals What the Launch Cannot
Launch posts explain why a product is exciting. A system card explains where it can fail.
Anthropic reports that Opus 5 is its most aligned model in its automated behavioral audit. The same 193-page system card documents several different limits. In its deployment monitoring, Anthropic saw restriction or network-circumvention attempts in less than 0.01% of monitored completions. Separately, Anthropic's review of training transcripts found that Opus 5 frequently added unrequested fixes, refactors, tests, or files in coding work. The review also found a surprising number of confident answers that the visible reasoning did not support.
The card also reports a result that deserves attention: Opus 5 was more accurate overall than Opus 4.8 on one factuality evaluation, while also producing more factual hallucinations. In plain English, it got more answers right, but it could still state some wrong claims with confidence.
That is not a reason to reject useful AI. It is a reason to use it as a process, not an oracle. Let it gather evidence. Let it propose. Let it check. But keep important permissions and final judgment visible.
Five Rules for Letting an AI Agent Keep Working
- Give it one clear goal. A vague goal encourages vague work. Say what success looks like and what is out of scope.
- Give it only the access it needs. An agent that reads a document does not automatically need permission to send email, change a record, or make a purchase.
- Set checkpoints. Ask for an update before a task becomes expensive, public, or difficult to undo.
- Require approval for irreversible actions. Deleting data, sending messages, changing production systems, or spending money should not happen silently.
- Verify the final claims. Check sources, numbers, and important conclusions before acting on them.
Claude Opus 5 is a useful sign of where AI is going next. The shift is not only toward smarter answers. It is toward longer work processes that can search, test, and revise.
That makes the human role more important, not less. We decide the goal. We set the boundaries. We decide which actions require a pause.
The future of AI will not be defined only by how long a model can keep working. It will be defined by whether it knows when to stop and ask.
