Fable 5.2: Initial Thoughts and Prompting Advice
- ai
- claude code
- prompting
Fable 5.2 came out yesterday, and we have been playing with it ever since.
In our experience so far, the improvements aren’t monumental. But they are certainly noticeable.
It’s also cheaper per token than Fable 5.1. Fable 5.2 costs $8 per million input tokens and $40 per million output tokens, down from $10 and $50.
Read on as we share all that we have learned so far, and how you can prompt the best out of it.
It Gets Hard Builds Right the First Time
Last night, we built a multiplayer game with Fable 5.2. It got it right on the first attempt. The animations were smooth, and the physics behaved exactly the way I expected it to.
A few weeks ago, we tried a similar build with Opus 5.5. The output came up short on assets and characters, and it needed a lot more work before it felt like a real game. Fable 5.2 didn’t need that work.
And there are others reporting similar results.
Lex Toumbourou (staff engineer at Canva) gave Fable 5.2 his nephew’s game idea, a pile of drawings, and an earlier version he’d built in June.
He says that Fable “threw away most of my code and rewrote the game”. It came back with a 12-level, four-player game, with roughly 175 images and 140 audio clips generated along the way.
The interesting part is how it got there. Before handing the game over, Fable wrote a bot to play it, ran the bot through all twelve stages in a browser, and checked screenshots for problems.
Videos That Are Ready to Ship
Opus 5.5 was already off to a great start with video. In our experience, it got a video about 80% right on the first try. Then you worked with it to get it over the line.
With Fable 5.2, you are spared that back-and-forth. Its videos are much more accurate on the first go, and they look production grade straight away.
Anthropic now lists “animations, and educational videos with sound in code” among the things Fable 5.2 is built for. Others who had early access agree. Matt Shumer calls its video work “a notch above Opus”.
It’s still code-based video, so think launch videos, product demos, and animated explainers rather than realistic footage. Here’s a prompt you can try:
Video prompt
Make a 30-second explainer video for my app, [app name], at 1920x1080.
Open with the problem it solves: [problem]. Then show these three features one at a time: [feature 1], [feature 2], [feature 3]. End on [your URL]. Use these brand colors: [colors]. Add a simple background track and sound effects that land on the transitions.
Before you show me the video, watch your own render. Fix anything that's off, such as timing, overlapping text, or elements cut off at the edges.
The Price Dropped, but Your Bill Might Not
Per token, Fable 5.2 is 20% cheaper than Fable 5.1. It’s also up to 30% faster.
At high effort, Fable 5.2 scores about the same as Opus 5.5 at max effort on the FrontierCode 2.0 coding benchmark, for roughly half the cost per task.
But when you run both models at the same effort level, the picture flips.
Atomic Agent ran that test on an early build of Fable 5.2. They gave both models the same job: fix 13 planted bugs in a Python codebase and build one new feature. Every run of both models passed all 59 hidden tests. Here’s the cost breakdown:
| Model (effort) | Time | Cost per task |
|---|---|---|
| Opus 5.5 (medium) | 4.2 min | $0.97 |
| Opus 5.5 (high) | 9.6 min | $1.70 |
| Fable 5.2 (medium) | 4.2 min | $1.87 |
| Fable 5.2 (high) | 26.6 min | $7.70 |
On a task both models could handle, Fable at medium cost about twice as much as Opus at medium. At high, the gap grew to more than four times.
In our experience, low and medium effort are more than enough for almost everything. So keep Fable on medium unless the task is genuinely hard. In Claude Code and the Claude apps, medium is already the default. On the API, the default is high, so set it yourself.
Read more: Opus 5.5 vs Sonnet 5.5: Which Model to Use for Coding
Is Fable 5.2 Included in Your Claude Plan?
Here’s how much of Fable 5.2 you can use, and how you pay for it, based on your plan:
- If you’re on the Pro plan, or have a standard seat on a Team plan: Fable isn’t included in your plan’s usage limits. You pay for it with usage credits, which you buy separately from your subscription.
- If you’re on the Max or Ultra plan, or have a premium seat on a Team or Enterprise plan: Fable is included in your weekly usage limit, but it can only use up to half of that limit. Once you reach it, you can switch to another model, or keep using Fable and pay with usage credits.
- If you’re on the Free plan: You can’t use Fable at all.
If you’re on Max or Ultra, or have a premium seat, Anthropic is also giving you one to three free Fable resets, depending on how long you’ve been subscribed. You can use them until March 8, 2027. You only get resets if you were already paying for Claude before Fable 5.2 launched.
When Opus 5.5 Is Still the Better Pick
Unless you have an infinite money/reset/limit glitch, you will still have to pick your moments with Fable 5.2. Here’s how we recommend you split the work between Fable 5.2 and Opus 5.5:
- Use Fable 5.2 when you hand off a big job and walk away. A whole game, a full feature with many moving parts, a video, a large rewrite. This is where getting it right the first time saves you hours.
- Use Opus 5.5 when you’re in the loop. Small fixes, UI tweaks, questions about your code, and anything where you check each step before the next.
Hit a bug that neither model can crack? Check out our debugging and bug fix service.
How to Prompt Fable 5.2
Here are some pointers on how to get the best out of Fable 5.2:
Name the Actions That Need Your Approval
You should name the exact actions that need your approval. For example: deploying to production, deleting data, sending emails, or spending money.
Be specific, but don’t overdo it. In Anthropic’s testing, one broad rule about messaging people outside the conversation made the model hold back about half of its routine messages to colleagues.
Tell It Whether the Code Is Yours to Change
Without guidance, Fable tends to build around your existing code and report nearby problems instead of fixing them. If you want it to clean up as it goes, add a line like this:
“This codebase is yours: what it’s for and how it’s built are your call, and you’ll keep building on it.”
Leave that line out when you want small, contained changes. Because it gives Fable more room to change things, it also makes the model write more.
Ask It to Finish the Whole Job
Fable can sometimes stop before finishing a long task, especially at low and medium effort. Anthropic suggests adding a short instruction telling it to finish the job. In their tests, this cut the number of unfinished tasks in half.
Ask It to Keep You Posted
On long tasks, Fable 5.2 can work for a long stretch without telling you much. Come back an hour later, and you’ll see a lot of activity but very little explanation.
Ask it to say in one line what it’s about to do before it starts, and to end with a short recap. It costs a few more tokens, but you’ll always know what happened during a turn.
Give One Agent the Whole Project
You don’t need a team of AI agents to build with Fable 5.2. One agent with a clear goal usually does better.
Fable already starts helper agents on its own when it needs them, and those helpers can make up a large part of a session’s cost. So don’t add more on top. Give one session the whole project and a clear definition of done.
Delete Your Old Prompt Crutches
Lines written for older models can hurt Fable 5.2. Anthropic recommends removing:
- Instructions to think less or “answer directly without deliberating.” Lower the effort level instead.
- Hard style limits like “answer in 20 words or fewer.”
A One-Shot Build Prompt You Can Use Right Away
Here’s a prompt that incorporates all the best practices we discussed above. You can use it to build just about anything with Fable 5.2:
One-shot build prompt
Build [what you're building] for [who it's for]. It must [the three to five things it has to do]. It's done when [your definition of done].
This project is yours: how it's built is your call, and you'll keep building on it. Match the style of these references: [links or screenshots].
Test your own work before you hand it to me. Use the app the way a real user would, check screenshots for visual problems, and fix what you find.
Don't stop while any part of the task is unfinished. If you have to stop, tell me exactly what's left.
Keep me posted as you work. Before each major step, tell me in one line what you're about to do. When you're done, give me a short recap: what you built, how you tested it, and anything I should check myself.
Ask me before you deploy anything, delete data, send messages to anyone, or sign up for a paid service. Everything else, go ahead without asking.
Read more: Anatomy of a Senior Engineer’s Prompt
Wrapping Up
Fable 5.2 is inarguably the best vibe coding model available today. Maybe not in terms of cost, but definitely in terms of output quality and accuracy.
If you qualify for resets, use them on the biggest thing you’ve been putting off. Fable is at its best when you hand it a hard job and let it run.
Frequently asked
Is Fable 5.2 included in the Claude Pro plan?
No. On the Pro plan, Fable 5.2 runs on usage credits, which you pay for separately. This is the same setup Fable 5.1 had.
How much does Fable 5.2 cost?
On the API, Fable 5.2 costs $8 per million input tokens and $40 per million output tokens, down from $10 and $50 for Fable 5.1.
Is Fable 5.2 better than Opus 5.5?
For large, complicated tasks, yes. Fable 5.2 scores higher on most of Anthropic's benchmarks and gets complex projects right on the first try more often. For everyday coding where you check the work as you go, Anthropic still recommends Opus 5.5, and it costs a lot less per task.
Which effort level should I use with Fable 5.2?
Medium for most work, which is already the default in Claude Code. Move to high for your hardest builds. Avoid max unless you're testing the model's limits.