Opus is the model to use when brand voice actually matters

The most interesting AI buying decision I’ve watched recently came from a legal tech company called Legora. They picked Claude Opus specifically because it could match the style of an existing document, or the style of a particular lawyer, better than anything else they tested. They called it “fine drafting.”

That phrase has been bouncing around my head for a week. Because the thing legal teams need (precise voice replication that survives across long documents) is exactly the thing most marketing teams are failing at with AI.

The generic voice problem nobody fixed

Every marketer I know has the same complaint about AI copy. It sounds like AI. You can paste in five brand guidelines, three example posts, and a customer persona, and the output still reads like a LinkedIn ghostwriter with a thesaurus.

The fix most teams reach for is more prompt engineering. Longer style guides. More examples. A “do not use these words” list. It helps a little. It does not solve the problem.

What Legora figured out, and what most marketing teams haven’t, is that this is partially a model selection problem. Not every frontier model is equally good at this. Opus, in particular, is unusually strong at picking up the rhythm and quirks of a sample and carrying them through a longer piece of writing without drifting back to a default voice.

I’ve been testing this against my own writing for the last month. Same prompt, same examples, run through GPT-5, Gemini 2.5, and Opus. Opus is the only one that consistently keeps the sentence fragments. The others smooth them out by paragraph three.

What “fine drafting” actually means for marketers

In legal work, fine drafting means the small choices: which clause comes first, whether you say “shall” or “will,” how a particular partner phrases a carve-out. The substance is mostly fixed. The style carries a lot of weight.

Marketing copy has the same structure more often than we admit. The substance of a product launch email is mostly fixed (what shipped, who it’s for, what to do next). The style is where brand actually lives. And the style is what generic AI output flattens.

If you reframe AI copywriting as a fine drafting problem instead of a content generation problem, the workflow changes:

You stop asking the model to invent. You start asking it to render. You provide the structure and the substance, and you ask Opus to write it the way a specific voice would.

A workflow that actually works

Here’s what I’ve been doing for client work, and what I’d suggest if you’re trying to get past generic output.

First, build a voice corpus per brand or per writer. Not a style guide. Actual writing samples. 8-15 pieces, varied in length and purpose. Email, long-form, social, internal Slack messages if you can get them. The Slack messages matter because they show how the person writes when they’re not performing.

Second, write the brief and the structure yourself. Bullet points, key facts, the argument, the call to action. Don’t ask the model to figure out what to say.

Third, hand Opus the corpus plus the brief and ask for the draft in that voice. Give it permission to use fragments, contractions, weird transitions, whatever shows up in the samples.

Fourth, and this is the part most people skip, run a diff. Pick three lines from the output and three lines from the corpus. If a reader couldn’t tell which is which, you’re there. If they can, your corpus is too small or your brief is doing too much of the writing.

Where this stops working

Opus is good at matching a voice that exists. It’s not good at inventing one. If your brand voice document is aspirational (“we want to sound like Patagonia mixed with Liquid Death”), the output will be mush. You need actual samples of the voice you’re trying to hit, written by humans, in the wild.

It also degrades fast on very short outputs. A 30-word headline doesn’t give the model enough room to settle into a style. Voice matching shows up most clearly in pieces over 300 words.

And cost matters. Opus is expensive per token compared to Sonnet or GPT-5 mini. If you’re generating 200 product description variants, this is the wrong tool. If you’re drafting a CEO’s quarterly letter, it’s worth every cent.

The real shift here is treating model selection as a creative decision, not just an ops one. For years, the question was “which model is cheapest for this task.” For voice-critical work, the question is “which model preserves the human under the prompt.” Right now, for long-form drafting in a specific voice, Opus is the answer. I’d build the workflow around that assumption, route only the voice-critical pieces to it, and let the cheaper models handle the volume work where sounding like everyone else is fine. The catch most people will miss: this only pays off if you actually have a voice worth preserving. If your samples sound generic, no model can save you.