AI in the Agency: What Do I Actually Do with Claude Opus 5?
Anthropic released Claude Opus 5 on 24 July 2026: an effort dial, agents that can work for hours. What I actually use AI for in a production agency, what I don't use it for, and what that dial has to do with craft.
Anthropic released Claude Opus 5 on 24 July 2026. It's the first Opus model of the fifth generation and the company's most advanced Opus.
Two things catch my attention.
First: the effort dial. You decide how much computing power the model devotes to a task. At lower effort it spends fewer tokens, runs cheaper, and preserves a significant part of its performance.
Second: agents that can work for hours, even overnight. When it gets stuck it finds a way around, when it makes a mistake it recovers, and it reaches the objective.
On pricing, nothing changed: 5 dollars per million input tokens, 25 dollars per million output tokens. The same as the previous Opus.
This article isn't a model review. It's the honest inventory of someone running a production agency: what I use these tools for and what I don't.
Where I genuinely use it
Multilingual site content. I run this site in five languages: Turkish, English, Russian, Ukrainian, German. For someone working in Antalya that isn't a luxury, it's a necessity — my client base really does come from those languages. Holding a consistent brand voice across five languages wasn't a job I could do alone. Now it can be done.
But let me add this: I write the Turkish text. The model translates, then I check the tone. If it worked the other way round — if I told the model "write a blog post" — what came out wouldn't be mine.
Proposals and scope documents. A hotel sends us a brief. Turning it into a shooting plan, a crew list, a day breakdown and delivery items is mechanical work. Mechanical work can be delegated. I put in the numbers, the model builds the structure.
Shooting plans and call sheets. Location, hour, person, equipment. Putting these into a table used to take an evening; now it takes half an hour.
My own software work. I run this site, client sites and my internal tools myself. I'm not a professional developer. Five years ago this wasn't possible; I would have had to hire a developer or use an off-the-shelf template. Now I can build my own infrastructure. That's the single thing that has changed my agency's cost structure.
Archive and tagging. I have a ten-year image archive. Finding what I shot was a harder problem than keeping it tidy.
Where I don't use it
Image generation. No frame I deliver to a client is generated by AI. That isn't just an ethical position; it's a business position. They hire me because I record a hotel's real room, real light, real texture. We're not selling a generated room.
Creative concept. I can tell the model "give me an idea for a brand film for this hotel" and it'll give me ten. All ten will be the average of things already seen. An average idea is failure in my work.
The moment of decision. Which frame I shoot on set, which model I choose, which light I build — these don't get delegated. If they could be, I wouldn't have a profession.
Talking to a client. The model doesn't write my emails to clients. A relationship isn't a delegable thing.
The effort dial is really a metaphor for craft
To me the most interesting part of Opus 5 is that dial. You decide how much power a task gets. You don't spend full power on a simple job; you don't hold back on a hard one.
That's exactly what I do on set.
For a hotel's standard room: twenty minutes, four lights, two frames. For the same hotel's honeymoon suite: two hours, six lights, fifteen frames and a stylist. Both are part of the same job, but they don't deserve the same effort.
The biggest mistake beginning photographers make is not knowing this setting. Either they give maximum effort to everything and can't finish the day, or they give minimum to everything and deliver a mediocre set. Craft is knowing how much to give to what.
I like that a tool presents this as a setting. It asks the right question.
Be careful about this
The better a model gets, the more convincing its output looks. Looking convincing and being correct are not the same thing.
I test this constantly when working with numbers. If you hand the day count in a proposal, the crew cost, the delivery items to the model and don't check them, one day you will genuinely sign off on a wrong number. Checking is your job, and that job isn't delegable.
The same goes for text. The model writes politely, fluently, cleanly. But my brand isn't polite. I read every text and harden it. If I didn't, within a few months this site would look like everyone else's.
Conclusion
In my work, AI accelerates what comes before and after production. It doesn't touch the middle part — the set, the decision, the eye.
That's good news for me. Because my agency's value was never email-writing speed or spreadsheet-filling skill. The value is in knowing which frame to shoot.
However powerful the tool gets, nobody answers that question on your behalf.
The ordinary talks about what the tool can do. The rebellious knows which job they took the tool out of.