Fable 5.2 excels at ambitious builds, but its token costs add up

I judge a new model by whether it makes me try things I would not otherwise attempt. Fable 5.2 does that for some tasks, but its strengths are uneven: for ordinary work, I might still choose Opus 5.5.

Every’s frontier team receives early access and free tokens from major labs. We test models on our own coding, writing, design and knowledge work, rather than relying only on public benchmarks or short demonstrations.

I rate the model especially highly for coding, including building games that are actually fun. For writing I give it a B or B-plus: it can bury the lead and its prose can be clunky. Design earns an A. Knowledge work can be excellent for the right user, but for routine tasks it can be slow and expensive.

My colleague Kieran Klaassen built EveryCut, the video editor we now use, with Fable 5.2 through prompting. It combines the particular editing, transcription and recording features our team uses from several products into a tool fitted to our workflow.

Klaassen sees this as a change in professional software: organizations can build tools around their specific workflows. This was not a two-hour project, however. I say the editor took weeks and thousands of dollars, which may still make sense for a serious creative team.

Tyler Nishida calls Fable an overnight model. I found it useful to send substantial work to a cloud computer, leave it running and return hours later. For interactive back-and-forth, lower effort settings or Opus may be a better fit.

The model can give an estimated completion time. In one example it warned Tyler to plug in his laptop because the task would outlast the battery. Klaassen prefers Opus 5.5 for thinking together and Fable for work he can hand off.

Klaassen also used the model on creative tools including a music application. He says its longer autonomous coding runs feel more trustworthy than earlier versions, which sometimes spent tokens on unnecessary work.

Our colleague Jack Cheng tests models by photographing toys in his children’s room and asking for a game. In the comparison shown, the Astra game has limited controls, while the Fable version uses the toys as racers and adds a track, items and more playable controls.

For routine documents and email, I think the extra cost and power are unnecessary. But for people willing to change the form of their work, Fable can make a research result into something more than a document or slide deck.

Our head of evaluations, Mike Taylor, researched how to get AI to write like him. Instead of presenting only a document, he used the model to turn footage of his explanation into a short documentary about the process and findings.

Taylor described the three-week test as exceptionally productive. His work involves model experiments: he could give the model ideas to run in parallel, attend to other responsibilities and return to results later in the day.

At recording time we did not know Fable 5.2’s exact price. The documentary example used about 1.4 billion tokens; we estimated almost $2,000 at Fable 5.1 API rates and $862 at Opus 5.5 rates. These are estimates using other models’ prices, not a bill or a confirmed Fable 5.2 price.

For writing, our staff writer Katie found the prose easy to read, emotionally aware and funny, but also sometimes condescending or inclined to stop early at medium effort. I prefer Gemini 4 Argon for writing, with Astra close behind; in my tests Fable can bury the most interesting point.

For design, Tyler Nishida saw a new ceiling and found the model easy to steer. Higher effort sometimes produced too much, so he suggests trying lower effort settings. I recommend varying effort, not only switching between model tiers.

One detail stood out in my recurring task of making a film from Plato’s Protagoras: the model gave the younger Socrates a dark beard. Other models often defaulted to an older, gray-bearded depiction. I see it as an example of the model attending to details in a creative brief.

My conclusion is to reserve this model for ambitious coding and creative work where the potential result justifies the tokens. For ordinary tasks or some coding jobs, I found faster, more efficient alternatives such as Astra 6 could be preferable.