In an essay published Sunday, Drew Breunig describes using Anthropic's premium model to shape designs before handing written briefs to GLM 5.2 for implementation.
By RuntimeWire Staff · Published
Primary source: Drew Breunig
Why it matters #
Breunig's workflow turns model routing into a concrete cost strategy: use the strongest model to define the work, then give predictable implementation to a cheaper system. That shifts value toward coding products that prepare context, evaluate substitutes and decide which model receives each task.
Anthropic's Claude Fable 5 is pushing developers to think harder about which coding tasks warrant a frontier model. That is the case Drew Breunig makes in an essay published Sunday, two and a half months after Fable's release.
The shift cuts directly into the product strategy of Dario Amodei (@DarioAmodei), Anthropic's CEO and co-founder. Amodei previously led large language model development as OpenAI's vice president of research, including work on GPT-2 and GPT-3, before founding Anthropic with his sister, president Daniela Amodei, and other former OpenAI researchers. Anthropic was built around a thesis that increasingly capable models require stronger safety controls. Fable 5 puts a conspicuous price on that thesis.
Anthropic released Claude Fable 5 on June 9 at $10 per million input tokens and $50 per million output tokens. Anthropic described it as its most capable generally available model and its most capable model for ambitious coding. Fable-related traffic requires 30-day retention under Anthropic's API documentation, adding a data-policy constraint to the price calculation.
For Breunig, who previously ran strategy and data science at PlaceIQ and co-founded the Reporter app, Fable's performance justified using it. The bill discouraged using it indiscriminately.
The expensive model becomes the architect
Breunig's workflow is the useful detail: he uses Fable to interrogate and shape a design, then hands a written brief to GLM 5.2 for routine implementation. He estimates that Z.ai's GLM 5.2 costs roughly one-ninth as much as Fable for the usage he compared.
The comparison does not require GLM to match Fable on every task. It only needs to be competent enough for the larger volume of predictable work after Fable has supplied the plan and relevant context. Breunig says GLM 5.2 is sufficient for much rote coding, particularly when it receives strong context.
That division of labor creates an opening for coding products that can decide which model gets each task. The valuable layer moves toward harness design, context preparation, evaluation and routing. Developers can spend premium tokens on architecture, debugging and ambiguous decisions while sending mechanical edits, migrations and test generation elsewhere.
GLM 5.2 arrived in the same week as Fable, according to Breunig's account. His comparison concerns the economics of routine implementation after a stronger model has prepared the brief, rather than parity across every coding task.
Breunig borrows his framing from Herb Sutter's argument that the slowdown in single-threaded processor gains ended software's "free lunch." Faster chips could once compensate for inefficient programs. Engineers eventually had to pay closer attention to parallelism, memory and architecture.
AI coding followed a compressed version of that pattern. Rapid model releases allowed teams to expect that the next model would fix weak prompts, thin context and crude agent loops at a similar price. Fable's price made waiting for the next model a less convincing engineering plan. The work of deciding "what goes where," as Breunig puts it, returned to the application layer.
Evaluated fallbacks cover policy risk
Price-driven routing also gives customers somewhere to send work when a preferred model becomes unavailable. Anthropic requires 30-day retention for Fable 5 and Mythos 5 traffic, including API use. For organizations using zero-data-retention workspaces, Anthropic's documentation says retention must be enabled for a workspace using the covered models; prompts and outputs in that workspace are then retained for 30 days.
Anthropic's materials describe additional safeguards for cybersecurity and biology requests. That makes model identity and response behavior additional inputs for a router, alongside price and task difficulty.
Three days after launch, Anthropic disabled Fable 5 and Mythos 5 after the US government directed it to block foreign nationals from accessing the models. In its statement on the suspension, Anthropic said the directive cited national-security authorities without supplying specific details and forced a shutdown for every customer. Access returned on July 1.
A coding product tied exclusively to Fable could therefore lose its primary model because of a policy decision outside the customer's control. Evaluated fallbacks can preserve availability, even if the substitute changes output quality. That episode supports Breunig's routing thesis without changing its economic core: teams need to know which cheaper model can accept a prepared brief before the preferred model becomes too expensive or unavailable.
Funding does not remove the routing incentive
On May 28, Anthropic announced a $65B Series H at a $965B post-money valuation, led by Altimeter Capital, Dragoneer, Greenoaks and Sequoia Capital. Anthropic also said its run-rate revenue crossed $47B that month. Both the revenue figure and Fable customer results, including a claimed one-day migration across Stripe's 50-million-line Ruby codebase, come from Anthropic rather than independent audits.
Anthropic said the funding would help expand compute for training and serving its models. It does not erase customers' incentive to reduce inference spending. Every improvement in lower-cost models strengthens the economic case for routing routine work away from the premium option.
Coding platforms have already attracted large sums by owning the layer between developers and model providers. Cursor, an AI coding workspace, announced a $2.3B Series D at a $29.3B valuation in November 2025. That financing reflects investor demand for companies that control how developers assign work across AI models.
Fable may remain the model developers call when a problem is difficult, poorly specified or expensive to get wrong. That is a strong position, although it may mean handling a smaller share of the tokens in an engineering workflow.
Amodei's bet is that Anthropic can charge for exceptional capability while imposing controls intended to make that capability safe enough to distribute. Breunig describes a practical customer response: use Fable where its judgment earns the premium, then route routine implementation elsewhere. The better Fable becomes at writing the plan, the easier it may become for cheaper models to finish the job.