Claude Fable Cost $9 in One Coding Test. GPT-5.5 Cost $1.50. Model Triage Is Now a Platform Engineering Skill.
Cloud Platform EngineeringThe pricing headline is striking. Claude Fable 5 runs at $10 per million input tokens and $50 per million output tokens. GPT-5.5 is priced differently. One identical coding task — same prompt, same codebase, same output expectation — cost $9 on Fable and $1.50 on GPT-5.5.
The performance difference on that task was not 6x.
The Routing Problem
In a single-developer workflow, a 6x cost difference per task is a productivity consideration. In a production agentic system where a single user action spawns 20–50 model calls, that multiplier compounds across every request, every hour, every day. At scale, the choice of which model handles which task is not a configuration detail — it is a recurring infrastructure cost with compounding exposure.
The pattern The New Stack identifies is this: without a deliberate routing policy, individual developer habit silently moves an entire team’s AI spend onto the most expensive model for work that does not require it. The developer who defaults to Fable for boilerplate refactoring is not making a bad technical decision. They are making a decision that the team has not given them the tooling to make differently.
What a Routing Policy Actually Does
A model routing policy defines which model class handles which task type, enforced at the infrastructure level rather than left to individual discretion:
Frontier models (Fable 5, Mythos 5): novel architecture problems, complex multi-step reasoning, tasks where output quality has direct revenue or safety impact.
Mid-tier models (Opus 4.8, GPT-5.5): standard code generation, documentation, retrieval-augmented tasks where the model is retrieving and rewriting more than reasoning from scratch.
Efficient models (Kimi K2.7 Code, Claude Haiku 4.5): high-volume, repetitive subagent tasks — parsing, classification, summarization, structured extraction — where cost compounds hardest because task frequency is highest.
The routing layer sits between the orchestration system and the model API. It intercepts each agent call, classifies the task type, and dispatches to the appropriate model tier.
Why This Became a Platform Engineering Problem
Two years ago, AI costs were a line item in R&D budgets — variable, experimental, absorbed. As agentic pipelines move into production, the cost structure changes: AI spend is now recurring, predictable in aggregate, and tied directly to product usage volume. A pipeline that routes everything to frontier models because that was the default during development becomes a compounding exposure as usage scales.
The analogy to compute resource management is direct. Teams that would never allow developers to provision arbitrary instance sizes without a policy are allowing developers to call arbitrary model tiers without one. The cost profile is structurally similar; the governance gap is larger.
Anthropic’s Fable 5 suspension by US government order — three days after launch — adds a second dimension to this problem. Teams that built hard dependencies on a single frontier model learned that model availability is not a constant. A routing policy that distributes across multiple providers is also a resilience pattern.
So what: If your team runs agentic pipelines in production and does not have a documented routing policy that maps task types to model tiers, you have a cost governance gap. The investment to close it is small relative to the compounding exposure. Start with your highest-volume agent tasks — the ones that run on every user action — and ask whether each one needs frontier capability or whether a mid-tier or efficient model produces acceptable output for that specific task type.
Source: The New Stack
Content created with AI assistance and reviewed for accuracy.
Join the conversation
Stack Insiders is our free community for readers who want to go deeper — share resources, ask questions, and connect with others across every vertical we cover.
Join Stack Insiders →