A single email doesn’t need the same firepower as a full market analysis — and Hollywood already learned that lesson the hard way.
When Denis Villeneuve’s team needed to keep the Fremen’s blue eyes consistent across roughly 1,000 shots in Dune: Part Two, they didn’t hand the whole problem to one all-purpose AI system. According to Foundry, they trained Nuke’s CopyCat tool on artist-made assets for that one specific job — eye color continuity — and left everything else, from color grading to full VFX composites, to different tools built for those tasks. Around 40% of the AI-trained shots needed no further cleanup at all.
That’s the approach Hollywood has quietly settled into industry-wide. Netflix’s $600 million AI filmmaking partnership splits its tools by task: custom models for visual effects, separate systems for automated color grading and dialogue cleanup, and a different set of tools entirely for pre-visualization. Lionsgate has deployed AI across more than 80% of its workforce, but not as a single tool — it runs Copilot, ChatGPT Enterprise, and Snowflake side by side, each one handling a different kind of work. Nobody in the industry is trying to get one model to do everything.
It turns out that’s also the exact mistake a lot of everyday ChatGPT users make.
An Email Isn’t a Market Analysis
Anyone who uses ChatGPT regularly has likely run into both situations. A task that sounds complicated gets finished in a few seconds. Another prompt, one that seemed simple at first, needs several rounds of adjusting before it lands. The prompt isn’t always the problem. Often, it comes down to which model is handling the task.
That’s the case for thinking not just about what gets asked of ChatGPT, but which model is doing the answering — and whether that model actually fits the job.
Consider a typical workday. The morning brings an inbox message that’s far too long, when all that’s really needed is the three most important points and a short reply. Speed matters most here, and a model built for complex reasoning is often overkill for a job this small. By the afternoon, the task looks different: comparing several documents, spotting the differences between them, and drawing conclusions from what changed. Here, a few extra seconds of processing time matters far less than whether the model actually understands how the information connects.
That’s why it’s worth testing different ChatGPT models rather than defaulting to the most powerful option out of habit. The better approach is choosing based on the task itself: a short summary, a longer analysis, coding work, or something where several steps need to build logically on one another.
ChatGPT tends to be most useful on tasks made up of several smaller steps rather than one large, vague one. A business report is a good example. The instruction “analyze this report” is quick to type, but it leaves far too much undefined.
A short workflow usually works better. First, ChatGPT extracts the key metrics and statements. Next, it identifies what changed compared to the previous year. Only in a third step does it move on to possible explanations or the questions those numbers raise.
Breaking the work into stages has a second benefit: interim results can be checked immediately. An error spotted in the extracted numbers can be corrected before it works its way into a full analysis — not unlike the way Hollywood’s VFX teams check whether an AI-trained shot needs manual cleanup before it moves further down the post-production pipeline.
A Good Answer Can Still Be a Wrong One
Even a highly capable model shouldn’t get a free pass on verification. Numbers, program code, and anything a later decision will rest on are worth double-checking before they’re trusted outright.
The U.S. National Institute of Standards and Technology takes the same position. Its framework for generative AI addresses how organizations can deploy AI reliably while managing the risks that come with it — which, depending on the application, can include human review, testing, and verification of whatever the model produces.
The Takeaway
In practice, the rule is a simple one. A short, clearly defined task can start with a fast model. But when the job involves large amounts of information, more complex relationships, or reasoning that has to build across multiple steps, it’s worth reaching for a model actually designed for that kind of work.
Hollywood didn’t need a $600 million deal to arrive at that idea — just a willingness to match the tool to the shot instead of running everything through the same system. The same logic holds at a desk with ChatGPT open: the fast model for the email that needs a quick reply, the more capable one for the report that actually needs untangling.
