As Fable Gets Pricier, Developers Split Coding Work Across Models Like GLM 5.2

Fable is widely regarded as one of the best coding models to date in terms of raw quality. The problem is price. Models available around the same time — Opus, GPT-5.6, K3 — could already handle most coding tasks just fine, so there wasn't much reason to route everything through Fable. dbreunig summed up the situation with a simple line: the free lunch is over.

That phrase traces back to a 2005 essay by software engineer Herb Sutter. Back when Moore's Law was still holding, you could just wait 18 months and CPU performance would double, so there was little incentive to optimize code by hand. But once single-thread performance stalled in the mid-2000s, developers were forced to think seriously about parallelization, architecture, and memory layout. dbreunig argues something similar is now playing out in the AI coding model market.

GLM 5.2: an alternative that showed up the same week

GLM 5.2, released in the same week as Fable, sits at the center of this shift. It reportedly costs about one-ninth as much as Fable and roughly one-fifth as much as Opus 5. Its quality may fall short of Fable's, but according to dbreunig, it's more than good enough for the bulk of routine coding work. That's why he's landed on a split workflow: talk through design decisions with Fable, then hand off the actual implementation instructions to GLM.

Model	Relative cost (GLM 5.2 = 1)
GLM 5.2	1 bar:11
Opus 5	5 bar:56
Fable	9 bar:100

Some argue that as inference costs keep falling, everything will eventually consolidate back around the biggest, most capable model. But dbreunig pushes back on that: when costs drop, cheaper models like K3 and Qwen benefit just as much, and as coding harnesses — the tooling that feeds AI systems task context — keep improving, even weaker models can perform well as long as they're given sufficient context.

A second shock: data policy

Pricing isn't the only thing Fable has shaken up. According to dbreunig, the model also comes with access controls, dynamic performance degradation, and mandatory data retention policies — enough that some companies and governments are now reconsidering where their work logs and token processing actually get routed. In effect, choosing a model has become inseparable from choosing where your data ends up.

This trend also connects to the recent price cuts from OpenAI and Anthropic as they respond to competitive pressure from Chinese AI models.

Editor's take

This piece isn't a knock on Fable. If anything, it's pointing out a paradox: Fable is so good that it ended up fragmenting the market. That tracks with what teams actually running coding agents have found in practice. Using the smartest, most expensive model to set direction during the design phase, then handing off the repetitive follow-through to a much cheaper model once that direction is locked in — that's already become a natural routine at plenty of development teams.

For teams here, there are two practical takeaways. First, build a routing strategy ahead of time that splits coding work into "design" and "implementation" phases, each assigned to models at a different price point. Second, get in the habit of tracking, as a team, which data flows into which model. If you adopt a model like Fable — with its strict data retention terms — without reviewing the contract first, you risk running into problems later during an audit. That's not something to gloss over.

Over the coming months, expect more teams to build stacks that combine cheaper models — GLM, K3, Qwen and similar — with harnesses tuned to work well with them. Meanwhile, the most expensive top-tier models will likely get pushed into a narrower role: handling only the high-stakes stages, like design and verification, where mistakes are costly. The era of one model doing everything looks like it's coming to an end.