Fable 5.2 costs 2×. Is it worth it?
Welcome back to Zero to MVP Weekly. One signal, one tool, one MVP idea — five minutes, every Saturday. Four new models this week, from 50 cents to 40 dollars per million output tokens. I only had one real question: should I switch from Opus to Fable for anything, and pay double?
1. The signal of the week
Price per task, not per token.
Four releases in four days, at very different prices:
- Claude Fable 5.2 (Anthropic): $8 per million input tokens and $40 per million output, down from Fable 5.1, and up to 30% faster. That's still twice the price of Opus 5.5. But Anthropic makes an interesting claim: on its FrontierCode benchmark, Fable at High effort roughly matches Opus 5.5's best score "for roughly half the cost per task." Anthropic itself still recommends starting with Opus 5.5 for most coding work.
- Gemini 4 Hydrogen (Google): $1.50 in, $7.50 out, a 1M context, and 57% on Terminal-Bench 4.1, by Google's count. It's in the free tier of AI Studio.
- Muse Ember (Meta): a small model for agent subtasks like tool calls, routing and extraction, at $0.10 in and $0.50 out. That's the same price as Claude Haiku 5.5, which it beats on some of Meta's benchmarks and loses to on others.
- Gemini 4 Carbon (Google) is now open to developers at $2 in and $10 out, but only until November 30. From December 1 it costs $4 and $20, the same as Opus 5.5. The same goes for Gemini 4 Argon.
What to do with it: the range is now 80×, from 50 cents to 40 dollars. Before you switch, run a few of your own tasks on two models and compare the bills, not the price lists. And if you're on Gemini Carbon or Argon, put December 1 in your calendar.
One honest caveat: "half the cost per task" is Anthropic's number, on one benchmark, and nobody outside has checked it yet. Google's and Meta's numbers are their own too.
2. My take
I think the price per million tokens is becoming a less and less useful way to choose a model. A cheap model may need several attempts and fixes, while an expensive one solves the task right away. What matters is the cost of a successfully completed task, including both the run time and the developer's time if they have to step in.
But to me, something else matters even more here: the shift from one all-purpose model to systems of several AI agents. Cheap models handle the simple tasks, powerful ones the hard tasks. Knowing how to distribute tasks well and optimize the total cost of the system is becoming more important than choosing any particular LLM.
3. Tool of the week
Fable 5.2 vs Opus 5.5 on a real code review.
- The test: GridCore, my scheduler for local AI models, written in Go. Both models got the same prompt: find races, deadlocks, leaks, error-handling bugs and security issues, change nothing, and write a report. They reviewed about 10,000 lines of code.
- What they found: nothing critical from either. The four most noticeable problems were found by both. Opus found 7 issues, Fable found 11, several of them in video memory accounting, the most important part of the scheduler.
- The real difference: for 5 of its 11 findings, Fable wrote throwaway tests on its own and checked that the bug actually reproduces. Opus drew all its conclusions from reading the code.
- Where it falls short: one task and one run per model, and I didn't measure cost. Most of Fable's findings only show up in specific setups, like two runtimes or a full disk.
- My verdict: Opus handles everyday work well. Fable is worth paying double only when you know a task needs maximum effort.
The video: Fable 5.2 — a routine update or a real step forward?
4. MVP idea of the week
Cost per passed task.
- Problem: price lists compare tokens, but you pay for tasks. Nobody knows whether Fable at High effort is cheaper than Opus at Max on their own code, or whether Hydrogen is good enough for their subtasks.
- Solution: point it at 5 to 10 of your own tasks with hidden checks. It runs each model at two or three effort levels, three times each, and reports dollars and minutes per passed task, not benchmark scores.
- For whom: small teams and solo builders choosing a default model, and anyone on Copilot whose models disappear on December 14.
- Minimal v1: a command-line tool around one coding agent, a YAML file with the tasks, and a Markdown table as the output. No UI.
- Demand check: publish one table, for example a cheap model and a frontier model on the same tasks with the bill for each, and ask who wants it for their own repo. Count the replies that come with a task attached.
- Why now: Anthropic itself now argues in cost per task, prices span 80× this week, and Gemini's prices double on December 1.
5. Useful links
- My DeepSeek V4.1 Flash test — the cheap end of the menu: six backend tasks, 18 runs, about 66 cents in total.
- Google — Gemini 4 Hydrogen — the new cheap tier: $1.50 input, 1M context, free in AI Studio.
- Meta — Muse Ember — a Haiku-priced model for agent subtasks; read the table, including where it loses.
-
Apiiro — Coding agents are leaking your secrets
— use Claude Code? Check your
settings.local.jsonand update to 2.1.323.
— Nick
P.S. Every video on the channel is the polished end of a much longer process. On Zero to MVP Lab, my Patreon, I share that process as it happens: local LLM benchmarks, configs, prompts, code, experiments, failed approaches and early conclusions. patreon.com/ZeroToMVP