Read time: 4 minutes

Leader, welcome back.

There are two ways to use AI. You can ask it a question, or you can hand it a job.

As of Monday, one of those costs up to 45% less than it did last week. It's the one almost nobody in finance is using.

Claude Fable 5.1 landed on Monday. The headline price didn't move.

What got cheaper is rereading. On a long job the model reads your brief, your documents and its own previous answers over and over. That rereading is billed separately, and it just fell 75%.

The saving isn't spread evenly. A quick question costs what it cost last week. A job that runs for hours costs up to 45% less.

If you're on a flat monthly fee rather than billed per use, you won't see a price change at all. You'll see it in your limits. Long jobs eat your allowance more slowly now.

Read the numbers properly.
Nothing on the rate card moved. That 25 to 45% is modelled on Anthropic's assumptions about how people work, not yours. Anthropic also claims 5.1 reads dense filings and charts with fewer confident errors, which is the claim that matters most to you and the one nobody outside Anthropic has tested. Try it on a filing you already know inside out before you trust it on one you don't.

If your risk team blocked Claude, ask them again.
When Fable 5 launched, Anthropic dropped zero data retention and moved to a 30-day window. Regulated firms couldn't use it. Microsoft restricted internal use.
That's now reversed. Enterprise Frontier Safeguards, announced Tuesday, keeps monitoring data in your own cloud account under your own encryption keys, and nobody at Anthropic reads it. It's free, opt-in, and rolls out in phases this autumn, with zero data retention on Fable 5 and 5.1 for eligible customers until then.
The thing your risk team said no to last quarter is not the thing on offer now.

The change nobody is covering.
Anthropic has published the off switch for Claude's writing voice. Its own prompting guide admits 5.1 writes denser prose, then hands you the fix:
tell it to remove all mannered prose.
Four words in your prompt. The output changes immediately.

Anthropic wrote that instruction because it thinks the style is a problem. Use it.

One more.
Get the expensive model to write the plan, then hand the plan to a cheaper one to execute. You stop hitting limits by lunchtime!

What this means for you: the cheaper half of AI is the half where you describe the whole job up front. The barrier is the describing, and that's a writing job.

Which is where most people stop.

You know this cycle.

You write a decent prompt. The output is wrong. You fix it. The third attempt is excellent.

Then the chat scrolls away. Next month, same task, you start from nothing.

That's a storage problem. You solved the prompt, you just didn't keep it.

Every major assistant now has skills. You teach a task once, save it, and call it back by name. The instruction moves out of a conversation and into somewhere you can find it again.

Here is what that looked like last week.

A team had run discovery calls across 5departments, all recorded in Granola. The job was to turn that pile into proper project documentation in Confluence.

By hand that's most of a week. Sit through recordings, pull out what matters, reshape it into the house format, repeat five times.

Instead we connected Claude to Granola and Confluence, wrote the instruction once with the exact structure the pages needed, checked the first output properly, and saved it as a skill.

The 6th department now takes minutes. So will the 12TH, in November, because the instruction isn't sitting in a chat that scrolled away in August.

For a finance team the obvious candidates are the jobs you do every single month in the same shape: variance commentary, the board pack narrative, the standard contract review, the weekly cash summary.

Easiest way in: next time a conversation goes well, ask for that chat to be turned into a skill before you close it. Ninety seconds. Almost nobody does it.

In GPT you can access the Skills section which looks like this:

In Claude Skills page look like this:

To add tools as a connection in GPT, select Plugins:

To connect tools in Claude select Customise and then Connectors:

Here's the part most people miss. A skill doesn't have to stay yours.

On Claude's Team and Enterprise plans an admin can push a skill out to the whole organisation. One person writes the variance commentary properly, once, and everybody runs that version. One good analyst becomes a company standard.

And you can see whether it landed. The admin analytics on those plans list your skills with the number of uses and the cost per use, so you know which ones stuck and which ones nobody has touched since the day you built them.

Which gives you a leaderboard. Publish the numbers monthly, put the author's name on every skill, and people start competing to write the one everyone else uses. Cheapest adoption lever I know.

What this means for you: if you have explained the same job to AI three times, you have paid for it three times. Save it once instead.