uygnoey Balancing intelligence against cost across a large deployment
Frontier capability comes at a frontier price, and Yusuke Kaji, General Manager of AI for Business at Rakuten, is direct that cost decides how widely he can deploy: "As a large enterprise, we want to balance intelligence and cost."
This skill describes the routing pattern as the post presents it.
Instructions
1. Measure two numbers together, not one
Kaji's team measures task completion ratio alongside cost per task.
Neither number decides anything alone. A cheap model with a low completion ratio produces work that gets redone; an expensive model with a high completion ratio may still be the wrong choice for work that the cheap model finishes fine. The pair is the signal.
Record both per task class with templates/task-routing-record.md.
2. Route by whether extra capability changes the outcome
The rule the post states: send Fable 5 the work where the extra capability changes the outcome, and let smaller models keep the rest.
This is a per-task-class decision, not a per-organization one. The question is not "is this model better" but "on this class of task, does the better model produce a different result?" Where the answer is no, the routing decision is already made.
3. Count the second-order costs, not just the per-token price
Kaji names two things that make the math work in Fable 5's favor:
- It gets more done with fewer tokens and fewer wrong turns. A higher per-token price does not settle the comparison if the cheaper model spends its run recovering from a bad early path.
- It needs less hand-holding. Human check-ins are a real cost that does not appear in a token bill.
The reference for how these signals relate to the model's self-verification behavior is in references/cost-signals.md.
4. Re-run the comparison when a model launches
Rakuten has tested Claude models since September 2024, across nearly a dozen model launches. Routing is not a one-time decision: each launch can move a task class from the smaller model to the frontier one, or the reverse, as capability and price both move.
Examples
Rakuten's stated position. As a large enterprise, Rakuten wants to balance intelligence and cost. The team measures task completion ratio alongside cost per task, then sends Fable 5 the work where the extra capability changes the outcome and lets smaller models keep the rest.
The two arguments in Fable 5's favor, as the post states them:
| Signal | What it means for the cost math |
|---|---|
| More done with fewer tokens and fewer wrong turns | The run itself is shorter and less wasteful, offsetting a higher per-token price |
| Needs less hand-holding | Removes human check-in time, a cost the token bill does not show |
A worked routing shape (the structure, filled with your own classes):
| Task class | Completion ratio, small model | Completion ratio, frontier | Cost per task, small | Cost per task, frontier | Outcome changes? | Route |
|---|---|---|---|---|---|---|
| Long unattended multi-hour run | low | high | low | high | yes | frontier |
| Routine bounded task | high | high | low | high | no | smaller |
The post does not publish Rakuten's own numbers; the shape above is the decision rule it describes.
Source
Working at the frontier: How Rakuten builds agents overnight with Claude Fable 5 — Claude blog, July 20, 2026.