Drew Breunig says Fable's cost pushed teams back to cheaper models

Key facts

• Simon Willison's post on 23 August 2026 is a direct quote from Drew Breunig's article "Fable & The End of the Free Lunch," with no added commentary.
• Breunig writes that before Fable, teams had little reason to invest in coding harnesses or context strategy, because a new model kept arriving at the same or lower price and closed most of the gaps.
• Fable itself was "incredible" but expensive enough that Opus, plus systems Breunig names only as "5.6," "K3," and "GLM," remained good enough for most of the code his team needed.
• That cost gap is what made the team start deciding "what work went where," meaning routing different coding tasks to different models instead of defaulting to the newest one.
• The excerpt does not specify what Fable is, who makes it, or its price, nor does it explain what "5.6," "K3," or "GLM" refer to.

Why it matters

For teams building on AI coding assistants, Breunig's account marks a practical shift: the assumption that waiting for the next model release solves cost and quality together no longer holds once a genuinely better but much pricier model shows up. Fable raised the ceiling on quality, in his account, but at a cost that kept the previous generation, Opus and the systems he calls "5.6," "K3," and "GLM," as the practical choice for most work. That forces an explicit decision about routing tasks by cost instead of always defaulting to the best available model.

Who it affects

Anyone building or maintaining a coding harness or an agentic workflow around multiple large language models, and any team that has to decide how to split work between a top tier, expensive model and cheaper alternatives for routine coding tasks.

How to use it

The excerpt gives no price, tier, or configuration detail for Fable, Opus, or the other systems it names, so there is nothing concrete to copy from it directly. The practical takeaway is the framing Breunig applies: treat model choice as a per task routing decision, sending the hard or high value work to the expensive model and the bulk of routine coding to cheaper ones, rather than assuming one model should handle everything. Readers who want the reasoning behind that split should read Breunig's full essay, which this quote is drawn from.

How solid is it

This is a single quoted paragraph, reposted verbatim by Simon Willison, an established commentator on AI coding tools, from Drew Breunig's own essay. It is a first person account of one team's experience, not a benchmark or a dataset; the excerpt itself gives no cost figures, no release date for Fable, and no comparison numbers.

Risks and caveats

The passage never states what Fable is, a model, a product, or a company, nor what it costs; it only says the cost "was so high." It likewise never explains what "5.6," "K3," and "GLM" refer to beyond naming them next to Opus as adequate for the same work. Readers should treat this as one practitioner's anecdote about his own team's model choices rather than a general claim about pricing across the market.

“Prior to Fable, it felt silly to waste too much time improving your coding harness or context strategies. A new model would arrive at the same price (or cheaper!) and paper over most of your problems. But then Fable landed. It was (and still is!) incredible. But the cost was so high and Opus was good enough (as was 5.6, K3, and even GLM) for most of the code we needed. So we started to think about what work went where.”

— Drew Breunig, quoted by Simon Willison, Fable & The End of the Free Lunch