AI Coding Weekly

Fable 5.2 wins Agent Arena but burns a $200 plan faster

Two scoreboards, one model, opposite answers.

Arena's Agent Arena results put Claude Fable 5.2 (Max) at #1 with +30.4% net improvement across 4.7k real-world agentic sessions, about 1.6x the #2 Opus 5.5 at +19.5%. It took #1 in Code, Chat and Work, with Code jumping from #6 to #1. Theo's reaction to the spread was that Anthropic's current dominance is, in his words, the biggest lead any lab has had since GPT-3.

The same week, Theo ran the other scoreboard. Analyzing over 1.1 million real requests to the Anthropic APIs on his $200 Ultra subscription, he found Fable burns faster per dollar than Opus, and that Fable is capped at 50% of the weekly limit, so using only Fable means half the allowance goes unused. That math was the basis of his earlier claim about what each model is worth on a monthly plan.

Work done per week, not per task

Theo's preferred measure is work done per week. Exclusively using Opus, he estimates roughly 1,200 turns and 230 threads in a week. Exhausting the Fable limit and then switching to Opus gets about 900 turns but slightly more threads, which he attributes partly to Fable being more likely to finish work in a single prompt. He flags the numbers as early and notes a lot of his usage was benchmarking the new model, so the ratio may move.

Arena's cost figures run the same direction. Median cost per task was $7.3 for Fable 5.2 against $5.5 for Opus 5.5 and $1.5 for Gemini 4 Carbon (High), which landed third at +17.0%. Fable did get cheaper than 5.1, with API pricing dropping from $10 in and $50 out per million tokens to $8 and $40, and median cost per task falling 6%.

For working engineers, the two rankings are not in conflict, they are measuring different things. Arena scores success per session, weighted toward signals like steerability, where Fable 5.2 led at +33%, and confirmed success at +22%. A subscription meters tokens, and a model that reasons more per answer costs more of your weekly budget per answer even when it is right more often. Theo's own game test pointed the same way, where Fable built a Fishslop run in 76 minutes for about $42 but he judged the Opus version more playable.

Theo - t3.gg
@theo
X
The $200 Claude Ultra plan gets you $15,000 of Opus 5.5 in a month, but only around $3,000 of Fable 5.5.
Nov 10, 2026 · View on X

Have the good opinion, not the loud one

One email a day. Under three minutes. What working engineers found out about coding with AI.

Get the next one by email

Coding with AI, read daily so you do not have to. The experiments, the receipts and the arguments from engineers shipping real software. Not a changelog.