Frontier AI Just Got Cheaper. Here's What That Does to Marketing

What is happening

Anthropic released Claude Sonnet 5, described as the most agentic Sonnet model they have built. The headline is not the benchmark chart. It is the price. Launch pricing sits at $2 per million input tokens and $10 per million output tokens through the end of August, rising to $3 and $15 after that. Performance is close to Opus 4.8, the far more expensive model, and it beats it outright on agentic search and computer use tasks.

It is the default model on the free and Pro plans, and it is available across Max, Team, Enterprise, Claude Code and the API. Anthropic also flagged lower hallucination rates and better resistance to prompt injection than the previous Sonnet.

The pattern here matters more than the release. Roughly every few months, the capability that used to sit at the top of the price list becomes the thing you get by default.

What I learned from this

The interesting shift is not that the model is smarter. It is that a model good enough to actually finish a task now costs almost nothing to run.

For a couple of years the honest position on AI in marketing was that it drafted well and executed badly. It would write you three ad variations and then fall over the moment you asked it to pull the data, check the landing page, cross reference the search terms and come back with a recommendation. Early access partners testing Sonnet 5 reported the opposite: it completes complex tasks where previous versions stopped short. Agentic performance, browser and terminal use, checking its own work. That is the part that changes a workflow rather than a first draft.

Cheap and capable together is what unlocks volume. When a competent pass over an account costs pennies instead of pounds, you stop rationing it. You stop asking whether a task is worth the token spend and start running the analysis on every campaign, every week, instead of on the three accounts that were shouting loudest.

What I keep coming back to is what this does to the value of the human sitting above it. When execution is cheap and abundant, execution stops being where the margin is. The scarce thing becomes knowing which question to ask, recognising when the output is confidently wrong, and having the commercial context to decide whether the recommendation is right for this business at this moment. I have watched AI produce a flawless technical answer to completely the wrong question. It does that faster now, and cheaper.

The other thing worth noticing: nobody has to upgrade to benefit. If your team is on Pro, this is already the default model. The capability landed whether you planned for it or not.

What I recommend for your business

Stop treating your AI tooling as a fixed decision. If you scoped your workflows around what a model could do six months ago, those assumptions have expired. Go back to the tasks you decided AI could not handle and test them again.

Start with the work that is repetitive, requires multiple steps and currently eats analyst hours. Search term reviews. Asset performance audits. Competitor landing page comparisons. Weekly account health checks across a portfolio. Those are the jobs that were previously too expensive or too unreliable to automate and are now neither.

Then build the review layer, because this is where teams get it wrong. Faster output with no verification step just means you distribute errors more efficiently. Decide who checks the work, what they check, and what triggers a stop.

Falling AI costs do not make your marketing better. They make it faster. Whether that is an advantage depends entirely on whether the thinking behind it was any good in the first place.