The price of intelligence fell twice in two days

Between Sunday and Monday this week, the cost of hiring a very capable mind dropped again. For what AI help costs overall, from model prices to hours and hires, see What AI help costs.

On September 21, xAI released Grok 4.7. The next day Anthropic released Claude Opus 5.5, and about an hour later OpenAI released GPT-6 Sol and GPT-6 Luna. Simon Willison put the prices side by side the same day. Per million tokens, input then output:

Model	Input	Cached input	Output
GPT-6 Luna	$0.10	$0.01	$0.50
GPT-6 Sol	$2.00	$0.20	$10.00
Grok 4.7	$2.00	$0.50	$6.00
Claude Opus 5.5	$4.00	$0.20	$20.00

Anthropic says Opus 5.5 is cheaper per token than Opus 5.0 and carries the intelligence of its larger Fable 5.1 model. The cached-input price on Opus 5.5 fell 60 percent. Luna is among the cheapest models OpenAI has shipped. Those are the companies’ claims, reported by Willison. Nobody outside has finished measuring them yet.

Why a business of one should care this week

A company with a hundred people pays for work in salaries. A business of one pays for most of its work in tokens. When the price of a token falls this far, the list of jobs worth handing to an agent gets longer overnight: first drafts, research sweeps, support replies, reconciliation, code.

But the sticker price is not the bill. Willison ran Opus 5.5 at its highest reasoning setting on a small drawing task. The run went nearly twenty minutes, hit the 128,000-token output limit, failed, and cost him $2.56. A cheap token spent on a runaway job is still an expensive job.

The number that matters is cost per finished job: what you paid for the work you actually shipped, including the failed attempts and the minutes you spent checking it.

What to rerun, and how

Pick one real workflow you already run with AI. Not a demo. Something that happens every week, like turning a call recording into a proposal or turning last month’s invoices into a cash summary.

1. Run it three ways. Once on the model you use now, once on the cheapest new model that might handle it (Luna is the obvious candidate), and once on the strongest.
2. Score the output, not the model. Would you send it as is? How many edits did it need? Did it fail?
3. Count the whole bill. Tokens, retries, and your own review time. Divide by the number of jobs that shipped.
4. Look at your cache. Long agent sessions resend the same context again and again. If most of your input is cached, the 60 percent cache cut on Opus 5.5 may matter more than the headline price.
5. Write the date on your results. This price sheet will look old within months. Treat it as a snapshot from September 22, 2026, not a ranking to keep.

The larger pattern

Every few months since ChatGPT went public in November 2022, the same thing has happened. Capability goes up, price comes down, and the set of work one person can direct gets larger. The owners who benefit are not the ones who guess the winning model. They are the ones who can swap models in an afternoon because their systems were built to expect change.

That is most of what we teach: build the loop so the model is a part you can replace, measure the work it finishes, and keep the judgment for yourself.