Updated August 31, 2026. This was written on March 31, 2026, when Claude 4 was the current model and the daily driver described below was the one I actually had. It is kept as a record of what changed at that moment, not as current advice. My orchestrator today is Claude Fable 5, which carries the script, the sheets, the storyboard and the direction of the generations. The observations here about where a model drops context, and about the metric a non developer actually counts, still hold. The version number does not.
Anthropic released Claude 4 (Sonnet and Opus), and I've been running it daily in my actual production work for weeks now.
Not in demos. Not in benchmark comparisons. In the things I actually need to get done.
Here's what changed for me.
It Stopped Arguing With Me
The older Claude versions had this habit of over-qualifying everything. You'd ask for something specific and it would give you three versions with a paragraph explaining why it chose each one. Helpful sometimes. Exhausting most of the time.
Claude 4 makes a decision and explains it once, briefly, only if the decision isn't obvious. That's a significant shift in working rhythm.
When you're building something (a script, a pipeline, a campaign), you don't want a committee. You want a collaborator who has an opinion and moves.
The Reasoning Is Different
I'm not a developer. I never wanted to be. But I now run automated video pipelines, content systems, and client-facing tools that I built with AI assistance.
With previous models, there was a ceiling. Complex logic would break down. You'd get something that looked right but failed in production.
With Claude 4, the ceiling moved. I've given it multi-step problems ("build this system that does X, then Y, then handles this edge case"), and it holds the full context without dropping threads halfway through.
For non-technical people trying to build actual things with AI, that reliability is the whole game.
What I Use It For (Specifically)
In the last month: automated video production pipelines (Remotion + ElevenLabs + Gemini), a client-facing website for a medical AI agency, cold email sequences for a lead gen product, and social content systems that run on their own.
None of that was possible for me two years ago. Not because the ideas didn't exist. Because the execution required a technical team I couldn't afford.
Claude 4 is not just a better chatbot. It's the difference between having an idea and actually shipping it.
The One Thing It Still Gets Wrong
Context management over very long sessions. If you're working on a large codebase across multiple hours, it still loses detail from early in the conversation. The work-around is disciplined documentation: keeping explicit notes that you feed back in when needed.
It's a real limitation. It's also manageable if you structure your work correctly.
I've learned to treat every session as a handoff. Summarize where you are, what decisions were made, what comes next. It takes two minutes and prevents a lot of frustration.
The Honest Assessment
I've tried GPT-4o. I've tried Gemini. I keep coming back to Claude.
Not for any feature that a spec sheet would highlight. For how it writes. How it reasons through ambiguity. How it handles creative direction instead of just technical instruction.
I direct commercials. I write sketches. I compose music. The model I use most is the one that sounds least like a machine when I ask it to help me think.
Claude 4 is that model right now.
That can change. It probably will. But today, this is what I'm using, and that's worth saying out loud.
Where I use it that nobody talks about
The places I get the most value out of Claude 4 are not the obvious ones. They are not coding assistance, even though I do that too. They are creative-judgment moments that used to require a peer on the phone.
Reading a treatment I wrote three days ago and asking it whether the second act earns the ending. Submitting a client's brief and asking what is missing from their stated goals. Looking at a transcript of a creative call and asking which decision the team avoided making. Pasting a piece of copy and asking what a senior creative director would push back on.
None of these are tasks. They are second-opinion checks. The ones I used to call a friend for at 11 pm and feel guilty about. Claude 4 is good enough at this kind of read that I do not feel like I am wasting the question, which is the bar for whether a tool gets used or quietly abandoned.
What I do with that read changes when the job is paid. On a commercial there is an approval chain, and the version I send is the version I have to defend in a room full of people who were not there when I made it. So on client work I stopped taking the first call the model makes, however good it sounds at midnight. I put my pass and its pass in the same document, mark which one is mine, and pick in the morning. Plenty of times I pick its version. What I cannot afford is to walk into the meeting with a cut I like and no memory of why it is that way.
The boring metric that actually matters
The benchmark I care about is not on any leaderboard. It is the ratio of times I close the chat satisfied to times I close it frustrated. With previous models, that ratio was somewhere around 60-40 in favor of satisfaction. With Claude 4, it has shifted closer to 85-15. Three percentage points off is the difference between a tool I reach for and a tool I avoid.
That is the metric to watch as you evaluate any AI tool: not what it does in a demo, but how often you actually want to use it once nobody is watching.
Sources: Anthropic, Claude | Anthropic News
