I ran Fable 5.1 for a day. Three things changed.

I gave Fable 5.1 my whole Tuesday. Client reporting, two automation fixes, a proposal draft, a pile of research, and the usual "read this thread and tell me what to do" tasks. Same setup I run every day, same prompts, same folders. The only change was the model.

By evening I had three clear findings and a short list of habits that make the model work harder for a company your size. The findings first, because they decide whether you should bother.

Finding 1: it still costs like a truck on a subscription

Anthropic cut the price of cached read tokens by 75% with this release (Anthropic's Fable page). Cached reads are the part of a conversation the model has already seen, like your system prompt, your uploaded docs, and the first 40 turns of a long chat. Most real work runs on cached reads, so on paper this is a big cut.

On a Claude subscription it made no difference to my quota. The usage meter drained at the same rate as Fable 5. I asked around, and the answer I got was that subscription pricing already runs at a discount, so the price cut on the API side never reaches it.

If your team uses Claude through the API, through a tool your developers built or through a vendor that bills you per token, this cut shows up straight in your invoice. For a long working session where 80% of the tokens are cached reads, Fable 5.1 now comes in cheaper than Opus, the smaller model most companies default to for cost reasons. Cheaper and smarter. Ask whoever runs your AI bill to pull last month's cached-read share and do the math.

Finding 2: it is fast when the task is small

Fable 5 thought hard about everything. Ask it for a two-line email and it would reason for 15 seconds before writing two lines.

Fable 5.1 answers the small stuff straight away, even when I leave the reasoning setting on high. "Rename this file", "give me three subject lines", "what's 18% of $42,000" come back in a second or two and in one short paragraph.

The bigger effect sits underneath. On the same task, 5.1 used fewer tokens than 5 to reach the same result. It stops thinking when thinking stops helping. Across a day that adds up to more work per dollar and less time watching a spinner.

Finding 3: the "Claude speak" is gone

You know the tone. "Great question! Let's dive in." Three bullets, each bolded, each ending with a summary of itself. A closing paragraph that repeats the opening.

I did not see it once in eight hours. Answers read like a sharp colleague typed them. Short, direct, no padding, no cheerleading. When I asked for an opinion I got one, with a reason, and it held the position when I pushed back.

The judgment held all day too. It chose the right file, the right tool, the right level of detail, without me steering. And it never fell back to Opus, which Fable 5 did a few times a day on my account.

So I'm running Fable 5.1 as my main model for the next two weeks. Six habits made it earn that spot.

How to use it better: six habits that took me a day to learn

Each one below has a plain-English reason and a line you can paste. None of them need a developer. The full set of prompting notes sits in Anthropic's prompt engineering guide, and the five-part brief I use for every prompt is in an earlier piece.

1. Start on high, then test lower

Fable 5.1 has five effort levels: low, medium, high, xhigh and max. Effort controls how long it thinks before answering, which sets speed, cost and quality all at once.

Most people leave it wherever it landed. Do this instead. Run your ten most common tasks on high for a week. Then run the same ten on medium. If you can't tell the difference on a task, move that task to medium for good.

The numbers matter here. Fable 5.1 on medium roughly matches Fable 5 on high, at lower cost. On low it beats Opus and Sonnet on most tasks while costing about the same. So the old rule of "use a small model for cheap work" now reads "use Fable 5.1 on low for cheap work."

Two things to watch at the ends. On low, it searches less and answers from memory more, so keep anything that needs fresh facts on high. On xhigh and max, it thinks so long before a big document that it can draft the whole thing in its head, then write it again. Save those two levels for problems where you have measured a quality gain, and keep long writing jobs on high.

2. Tell it to talk while it works

On long jobs, 5.1 goes quiet. Fable 5 narrated every step. 5.1 does the work and reports at the end, and the final message often covers only the last step. If someone on your team is watching an agent run for ten minutes with no output, they will assume it is stuck.

The fix is one line in your instructions:

"Before you start, say in a line what you're about to do. Give brief updates while you work. Close with a short recap that stands on its own: what you found, what you did, and what's next."

Also delete any old instruction that says "hold everything for the final answer." Those were written for chattier models and they now make 5.1 fully silent.

3. Tell it to finish

This is the one that changes the most for a founder who delegates. 5.1 can run very long tasks on its own. But left alone it sometimes stops early and describes what it would do next, or asks "shall I apply this?" about a step you already asked for. You reply "yes, go ahead," and you have just become the bottleneck in your own automation.

Paste this at the top of any task you want done end to end:

"You are operating on your own. I'm not watching in real time and can't answer questions mid-task. For reversible steps that follow from this request, proceed without asking. Stop only for destructive actions or a real change of scope. Before you end, check your last paragraph. If it's a plan or a promise, do that work now."

That first sentence does most of the lifting. Keep it.

One exception worth adding: when you are thinking out loud or describing a problem, say so. Then it reports and stops instead of fixing things you only wanted to discuss.

4. Tell it what not to touch

Ask 5.1 to add one feature and it may also fix a nearby bug, extend something you never mentioned, and add four test files. Every extra is well intended and every extra is something you did not ask for and now have to review.

This applies well beyond code. Ask it to tighten one section of a proposal and it will rewrite the intro too. So say where the edges are:

"Change only what this task needs. If you notice something else worth fixing, list it at the end as a follow-up. Don't fix it."

Anthropic's own tests show this line cuts the extras sharply with no drop in how often the main task gets done.

5. Ask for plain writing, and say when you want structure

Two writing quirks pull in opposite directions.

5.1 writes denser than 5. Longer sentences, fewer paragraph breaks, and a taste for metaphor where a plain word would do. "A dial worth turning" instead of "a setting worth testing." If you see that, add this:

"Say what you mean. When a literal phrase is available, use it. No metaphor, no flourish."

At the same time, it uses less formatting than earlier models. Fewer bullets, less bold, fewer headers. If your old instructions say "no bullet points ever," they were written to hold down a model that over-formatted. 5.1 does the opposite, so replace that rule with one that says when structure helps:

"Use lists when the content has several parallel parts. Use plain prose for anything conversational."

One more writing note. When it summarises a document, 5.1 sometimes lifts a sentence straight from the source without quote marks. If you publish those summaries, ask it to reword everything except short marked quotes, and give it one example of what a good summary looks like.

6. Ask for small edits, and give it room for big ones

Two habits for anyone who lets the model edit files.

First, 5.1 likes to rewrite a whole file for a two-line change. The output is usually fine but it costs time and tokens. Say:

"Edit only the lines that need to change. Don't rewrite the file."

Second, if you do ask for a long deliverable on xhigh or max, tell it the budget:

"Everything you produce in this reply, including reasoning, counts toward one limit. Don't draft the full document in your head and then write it again. Use the reasoning to settle structure and check inputs, then write once."

That single note made a 3,000-word rewrite come back in a third of the time on my setup.

Three things I got wrong on day one

I want to be honest about the misses, because they are the ones you will hit too.

I ran everything on xhigh for the first two hours because "more thinking is better." Wrong. Half my tasks came back slower with no visible gain. Medium did the job on most of them.

I left an old "keep replies short, no updates" line in my setup. 5.1 obeyed it completely and I spent twenty minutes wondering if a job had died.

I pasted a whole 60-page PDF and asked for one number. It found the number, but it also quoted three paragraphs verbatim as if it had written them. The next time I gave it the six relevant pages and the example line above, and the summary came back clean.

What this means if you run a $5M to $100M company

Your team is already using Claude. Most of them are using it the way they used the last version, with the same prompts and the same expectations. That leaves money on the table in three places.

If you pay by the API, check your cached-read share and re-price your model choice this month. If you pay by subscription, nothing changed on cost, so grade the upgrade on speed and judgment instead.

Give every person on your team the six lines above. Put them in the shared instructions once, and the whole company gets the model that finishes tasks, speaks plainly and stays inside the lines. At NuVision Auto Glass, the $48M company where I run AI and growth, that shared file is the difference between automations the team trusts and ones they quietly route around.

And run the effort test. One week on high, one week on medium, the same ten tasks. Most companies I've sat with never do this, and they pay for thinking that changes nothing in the output.

I'll post the two-week update with the numbers. The model has held the main seat for one day so far. That's already more than Fable 5 managed in its first week.

If your team runs Claude every day and nobody has touched the shared instructions since the last model, that's the cheapest fix here. Send us the file your team runs on and we'll rewrite it for 5.1, with the six lines above in place and the effort test set up, inside a week. If you'd rather talk it through first, book 20 minutes. And if the bill is the part that worries you, the earlier piece on what AI actually costs a company your size covers the API-versus-subscription math in full.