Fable 5 is back, but pricey — here's what else actually works

Quick summary

• The value pick is GLM-5.2: open-weight, roughly one-sixth the cost, and level with the closed flagships on third-party coding benchmarks.
• Opus 4.8 is the closest stay-in-Claude option, and was never part of the suspension.
• New 'mixture' tools (Hermes MoA, OpenRouter Fusion) try to beat any single model by combining the ones you can access — promising, but the standout numbers are self-reported.
• Pick by your actual job — Fable 5 is back, but many jobs never needed the priciest model.

The honest answer first

Fable 5 has no free public download, and no single model is a clean one-for-one swap. But several models you can use today get you most of the way — and one of them is genuinely close to the closed flagships, for a fraction of the price.

At a glance

The models you can run today — and where the buzzy orchestration tools actually land.

The value pick: GLM-5.2

GLM-5.2, from Z.ai (Zhipu), is open-weight (MIT), runs a 1M-token context, and on independent coding benchmarks lands right next to Claude Opus 4.8 and ahead of GPT-5.5 — at around one-sixth the cost. If you lost Fable 5 and want something strong and cheap today, start here.

One caveat: using GLM through its hosted API raises data-handling questions for some teams. Because the weights are open, you can also run it yourself to avoid that.

Closest to Claude: Opus 4.8

If your work is built around Claude — Claude Code, the same prompts and review steps — Opus 4.8 is a solid everyday route at half the token price of Fable 5's post-July-7 credits, and it was never part of the suspension. Keep the same repo and brief so any comparison with Fable 5 stays meaningful.

The mixture route: combine the models you CAN use

A newer idea is getting attention: instead of one model, route a task through several and let an aggregator synthesize the best answer — aiming to beat any single model you can access. Two to watch:

Hermes Agent's Mixture-of-Agents (Nous Research, just shipped) exposes 'MoA presets' as virtual models; it claims to beat Opus 4.8 by about 8% and GPT-5.5 by about 11% — but on its own, not-yet-released benchmark, so treat that as self-reported. OpenRouter's Fusion does something similar server-side. Both still run on the models you already have, and both add cost and latency. A third, Sakana's Fugu, exists but has drawn poor reviews for speed and price.

• Where these tools land in the timeline what's real vs hype, attributed

Pick by your job

• Long coding / agentic workGLM-5.2 or Opus 4.8; try a mixture route if you want to push past a single model.
• Cost-sensitive or high-volumeGLM-5.2 (open weights, low cost), or a smaller model with good scaffolding.
• Stay-in-Claude continuityOpus 4.8.
• Need the top model on a hard jobFable 5 is back — use it with prompt caching and the Batch API to blunt the premium credit rate.

Want it tailored to your case?

These are the options in general. If you tell our AI advisor what you're actually building and your budget, it'll name the single cheapest setup that works for you — specific models, rough monthly cost, and how to wire it up. Free, and a live preview of Coily.

Open the AI advisor → · GLM-5.2 vs Opus 4.8, head to head →

From the Field Guide

Stop guessing which model to use

Weekly verified workflows for the credits era: which model in which seat, what it costs per task, and copy-ready prompts — plus an 8-runbook Starter Pack you can download once and keep. Every claim checked against published pricing, sources shown.