Most people pick one Claude model and use it for everything. Usually it’s the most powerful one they can get. If you’re paying for Opus, why would you write with anything else?
Because for content, the answer to “Opus 5.5 or Sonnet 5.5?” is both, at different points in the work. I use Opus 5.5 to think and Sonnet 5.5 to create. When I let Opus draft, the writing comes out over-engineered: titles that don’t land, caveats nobody asked for, warnings the reader doesn’t need. Sonnet writes the plainer draft. Opus does the work that decides whether the draft is worth writing at all.
If you run your company’s content with AI, that split is the most useful model decision you can make right now.
Two models, two different jobs
Anthropic released Opus 5.5 on September 22 and Sonnet 5.5 six days later. The company is pretty clear that they’re for different work. It describes Opus 5.5 as “built for complex work requiring careful judgment.” Sonnet 5.5 is “strongest at well-scoped everyday tasks,” including “creating polished documents.”
Opus 5.5 is a better writer than earlier Opus models, too. Anthropic says it communicates more clearly and puts the most important information up front. That’s a big part of why people default to it. But a model that’s great at careful judgment brings that judgment to every sentence, and a blog post doesn’t need every sentence weighed like a legal memo.
Price matters too. On the API, Sonnet 5.5 costs half of what Opus 5.5 does per token, $2 versus $4 per million input tokens and $10 versus $20 for output, according to Anthropic’s pricing page. If you’re running drafts all day, that adds up.
None of this is obvious from the model picker. Even Unite.AI’s recent Claude review lists it as a con: plans, models and modes change so fast that it’s hard to know which one to use.
Use Opus 5.5 to decide what to say
The hard part of company content happens before the draft. What’s the point of this post? Who is it for? What do we actually know that the reader doesn’t? Is that claim true, and where’s the source? That’s research, strategy and judgment, and it’s exactly the work Opus 5.5 was built for.
Say you’re launching a new product feature. Opus is the model you want reading the customer interviews, the support tickets and the competitor pages, then telling you which angle holds up and which claims you can back. It’s also the model you want building the things you’ll reuse: your writing skills, your content plan, the brief that every draft starts from.
Use it again at the end, as the checker. Hand it the finished draft and the brief and ask what’s wrong: a claim without support, a point that drifted, a paragraph your buyer won’t care about. Careful judgment is a super useful thing to have on both sides of the draft.
Use Sonnet 5.5 to write it
Once the thinking is done, writing the draft is a well-scoped task. You have the point, the reader, the facts and the structure. You need clear sentences that get to the point and stay there. That’s Sonnet’s lane, and Anthropic says Sonnet 5.5 “writes more clearly than our previous generation of models.”
When I draft with Opus, I get the extra stuff: a hedge on a claim that didn’t need one, a “keep in mind” paragraph, three title options that each try to do too much. When I draft with Sonnet from the same brief, I get the post. It still needs an edit, like any draft. It just doesn’t need a rescue.
I run AI-powered news operations, and the old newsroom split still holds. The person who decides what the story is and the person who writes the copy aren’t always the same, and neither of them should be skipping the editor. AI makes that split cheap enough for a startup to run: one model plans and checks, another drafts, and a person still edits.
More thinking isn’t always better
Both models now come with effort levels, from low to max. More effort means the model reasons longer and checks its work more. It’s tempting to turn everything up.
Anthropic’s own Sonnet 5.5 results show why you shouldn’t. On one of its coding benchmarks, Sonnet 5.5 scored lower at max effort than one step down. At max, it more often spun up a whole multi-agent review, which sometimes ran out of time or made changes nobody asked for. That’s code, not prose, but it’s the same thing I see in drafts. Extra deliberation turns into extra output, and extra output isn’t the same as a better post.
So set effort by stage, the same way you pick the model. Turn it up for research, strategy and fact-checking. Leave it at the default, or lower, for drafting. Anthropic’s own advice is that lower settings suit routine work, and a draft written from a good brief is closer to routine than most people think.
How to decide for your own content
Don’t take my split as a rule. Test it. Write one solid brief for a real piece, then draft it twice: once with Opus 5.5, once with Sonnet 5.5. Read both as your buyer would. Which one gets to the point faster? Which one has fewer sentences you’d delete? Which one sounds like your company and not like a model being careful?
That’s the same approach I argued for when comparing Opus 5.5 with GPT-6 Sol: benchmarks narrow the field, but your own work decides. Run the test again every time a new model ships. The right assignment changes as the models do.
Whatever you pick, keep the decisions with a person. The model doesn’t decide what your company believes or what goes out under your name. If the knowledge belongs to your best expert, they supply it and stand behind it, and someone edits before anything ships. The model choice just decides how much cleanup that editor has to do.
Opus to think, Sonnet to write. Start there, test it on your own work, and change it when the results tell you to.
