Anthropic's New Model Doubled Its Benchmark Scores. Your Marketing Copy Won't Notice.

Claude Opus 5 shipped on July 24, and Anthropic led with the scoreboard. On one frontier reasoning test the new model more than doubled the score of the version it replaced. On a test of abstract problem-solving it scored roughly three times its nearest rival. The price held at $5 and $25 per million words in and out, the same as the model before it.

Every one of those leaps is in the same kind of work: reasoning, coding, running tools on its own. None of them is about writing.

For a marketing team that reads each model release as a copy upgrade, that gap is the story.

The place a newer model shows up for a marketer isn’t the blog draft. It’s the research and the checking underneath it.

Start with the writing, because that is what most teams reach for first. On the public leaderboard that scores models on creative writing, the top of the field is packed into a narrow band: Gemini 3.1 Pro at 1487, Claude Opus 4.6 at 1468, GPT-5.4 Pro at 1461, inside about 50 rating points of each other. On coding and reasoning the same models are separated by wide margins. The editor who runs that leaderboard, writing as Glevd, names the reason plainly:

“Writing quality is harder to benchmark than coding or math. There’s no SWE-bench equivalent for prose.”

There’s no hard test for prose, so no lab optimizes for it the way they chase the benchmarks they can win. The result is a writing frontier that barely moves between releases while the reasoning numbers jump. A blind test run this spring found the gaps that remain are small and situational: across eight prompts and 134 readers, Claude took four rounds, Gemini three, ChatGPT one, and the test used cold, short prompts with no brand voice or brief attached. Claude has a real edge on tone. It’s narrow, and a decent brief closes most of it.

Lisa Peyton, who teaches at the University of Oregon’s School of Journalism and Communication and tests these models on real content work, put the skill that actually matters in one line:

“The job quietly becoming the most valuable one isn’t picking the smartest tool. It’s knowing which to reach for, and catching what they miss.”

So where does the upgrade land? In the work that sits around the writing. The marketing consultancy Your Marketing People, reviewing Opus 5 the day it launched, found the real change was that “the model got better at holding a large, messy pile of context in view while chaining several dependent steps and checking its own work along the way.” That’s the unglamorous middle of the job: pulling a quarter of campaign data into one view, running a multi-step competitor teardown, catching the number that contradicts the other number before it reaches a slide.

The mechanism is straightforward once you separate the two kinds of work. Short marketing copy is easy for every frontier model, so a better model clears a bar they all already cleared. Reasoning over a large, contradictory pile of inputs is hard, so a better model moves work that used to stall. The first is where teams look for the payoff. The second is where it lands. Anthropic’s own marketing team ran into the same split when it automated its weekly reporting: the work that survived the automation was checking the numbers, not writing them up.

None of this makes the model a strategist. Your Marketing People is blunt on the limit: “Opus 5 does not make strategy obsolete. It makes the consequences of weak strategy arrive sooner.” Point a faster, sharper model at a weak brief and it produces a weak deliverable faster, with more confidence. The brief is still the input that decides the output, which is why the leaderboard editor defines real writing work as “match this brand voice, keep it under 800 words, use this structure, don’t use these phrases” rather than “write me something creative.” That instruction set is yours to build, not the model’s to invent.

For the budget, the correction is small and specific. Stop pricing the next model release as a content upgrade; the copy your team ships already sits at the ceiling the whole field shares, the same reason it rarely pays to rebuild your marketing around whichever model just shipped. When OpenAI cut frontier prices again this month, the useful response was to point the new headroom at the work that stalls: the research passes, the data reconciliation, the self-checking that used to eat an analyst’s afternoon. Re-standardizing on the cheapest writer just buys a faster version of a draft that was already good enough. Budget for the work under the draft. The draft was never the part that was going to get better.