Goosefeathers home

2026-11-11

Fable 5.2 is overrated, except on low effort

In this post (5 sections)
  1. What Anthropic says
  2. How I ran it
  3. Test 1: Squishy gummy bears
  4. Test 2: A 3D biology textbook
  5. So is Fable 5.2 overrated?

People across social media are hyping the living hell out of Anthropic's new model, Claude Fable 5.2. I'm not sure it deserves the hype. I gave it two tests I've run on other models and put it up against Claude Opus 5.5, which has been my go-to model.

Two gummy bear demos from the same prompt. Opus 5.5's bears are glossy and detailed. One of Fable 5.2's bears hangs from its stretched neck with an oversized rear end.

Opus 5.5 is on the left, built for $13.48. Fable 5.2 is on the right, built for $40.26. Look at the red bear's rear end.

What Anthropic says

On paper, Fable 5.2 is a beast. In Anthropic's launch post it came first on 10 of the 12 benchmarks they ran it on, often by a decent margin.

The bigger claim is about hardware. Anthropic says it gave Fable 5.2 one paragraph asking for three real devices: a home assistant hub, a speakerphone and a handheld virtual pet. Fable wrote the specs, picked every part from US suppliers, designed the circuit boards and cases, and wrote the software. Anthropic had the devices made, and all three worked. It also hooked Fable 5.2 up to robots that unscrewed a bottle cap and solved a Rubik's cube.

It isn't cheap. Fable 5.2 costs $8 per million input tokens and $40 per million output tokens, twice the price of Opus 5.5. Anthropic says it's faster, uses fewer tokens and does better at lower effort levels, so a whole task can still cost less than it would on Opus. I had my doubts about that.

How I ran it

Same prompts, word for word, each in an empty folder with nobody answering the model's questions. Every run was in Claude Code, one run per model and effort level, so don't read too much into small gaps.

Costs are what each run would cost at list API prices, worked out from the tokens it used, even where a subscription covered the run. Fable 5.2 on extra-high hit the Claude session limit partway through the gummy bears and was told to keep going. Opus 5.5 ran the Cell Atlas prompt twice; these are its second-run numbers.

Test 1: Squishy gummy bears

The prompt asks for see-through gummy bears on a table, in a single HTML file, that you can grab, stretch, squash, throw and pour. It tests two things: how good the bears look, and how good the physics feel.

Model Effort Time Cost
Claude Opus 5.5 Extra-high 49 min $13.48
Claude Opus 5.5 High 35 min $7.78
Claude Fable 5.2 Extra-high 87 min $40.26
Claude Fable 5.2 Medium 35 min $10.39
Claude Fable 5.2 Low 30 min $9.18

Opus 5.5 on extra-high is the best I've seen on this test. The bears look great, the colors pop and the physics are really solid. It's the gold standard.

Fable 5.2 on extra-high made smaller bears that also look good, and the physics are solid too. But the bears' rear ends are ginormous, and it doesn't look right. Fable 5.1 had the same problem on this test, and the Fable models are the only ones I've seen do it. So it's good, but definitely worse than Opus, and it cost three times as much and took almost twice as long.

On medium, the bears don't look as good and they still have the same oversized rears.

Low is where it gets interesting. The colors don't pop as much, but the physics are really good, the bears still look good and they don't have the big rears. Fable 5.2 on low did better than it did on medium or extra-high, and it cost $9.18, less than Opus's $13.48. It still isn't as good as Opus on extra-high.

Opus 5.5 on high is the cheapest of the lot at $7.78. Its physics aren't quite as good as Fable on low, but I think its bears look a little better. We're splitting hairs at that point.

Play them yourself: Opus 5.5 on extra-high, Fable 5.2 on extra-high, medium and low, or all eleven builds.

Test 2: A 3D biology textbook

The second prompt asks for Cell Atlas: an interactive 3D biology explorer with an animal cell, a plant cell, a white blood cell, a neuron and a DNA strand, every model built from code. You can spin each one, zoom in and click a part to learn about it.

Two 3D animal cells, each beside a glass panel listing its parts. Fable 5.2's is packed with colorful parts. Opus 5.5's is a translucent sphere cut open to show the nucleus.

Fable 5.2 is on the left and Opus 5.5 is on the right, both on extra-high.

Model Effort Time Cost
Claude Fable 5.2 Extra-high 89 min $37.77
Claude Opus 5.5 Extra-high 93 min $18.17
Claude Fable 5.2 Medium 32 min $11.46

Fable 5.2 on extra-high looks awesome. There's an astonishing amount of detail, and clicking around the cells, zooming and turning them all works well. I wish I'd had something like it in high school.

Opus 5.5 on extra-high is a bit different: a little more translucent, with plenty of detail of its own. If I had to pick, Fable did a little better here. It shows more detail and its colors pop more. But Opus wasn't far off, and it cost about half as much.

I also ran Fable 5.2 on medium. It came in at 32 minutes and $11.46, cheaper than Opus this time and about a third of the time. It's not as good as Fable on extra-high, and I think it's slightly behind Opus on extra-high too, but it's not far off.

One caveat: the prompt asked for correct biology, and nobody has checked every part.

Try them: Fable 5.2 on extra-high, Opus 5.5 on extra-high, Fable 5.2 on medium, or all three side by side.

So is Fable 5.2 overrated?

Kind of. It's a great model, and I think it's more powerful than Opus 5.5. But in my testing the margins are slim. It feels like an iPhone upgrade: a better camera and a faster chip, but the same phone at the end of the day. On extra-high it lost one test to Opus and won the other by a little, at two to three times the price.

What interests me most is how it does on the lower effort levels. I never use low, and I rarely use medium, because the quality usually drops a lot. Fable 5.2 is the first model I've seen deliver solid quality on low.

And those settings are priced well. Its best gummy bears came from low, at $9.18 instead of $40.26 on extra-high, 77% less. Its Cell Atlas on medium cost $11.46 instead of $37.77.

So if you use Fable 5.2, try it on low or medium before you reach for extra-high. Every build is linked above, and the prompts are on GitHub if you want to run them yourself.