Read the claim at the right level
Anthropic's September 28 Sonnet 5.5 announcement reports faster generation and up to 30 percent lower cost per task in its testing, despite unchanged listed token prices relative to Sonnet 5. The vendor attributes the savings to completing work with fewer tokens. This is a useful distinction for buyers: unit price and task cost are different measurements. The published comparison is Anthropic's own result, not a guaranteed saving for every workload. A business must test the workflow it actually intends to buy.
Source: Introducing Claude Sonnet 5.5 · 2026-09-28
Accepted work is the useful denominator
Imagine two illustrative models preparing product summaries. One costs less per attempt but frequently omits required dimensions. The other costs more per attempt and usually produces an acceptable draft. Count all attempts, including repairs, plus the staff time required to inspect them. Divide that total by the number of summaries accepted for their intended use. This calculation can reverse an apparent bargain. It also creates a productive discussion with the team: which defects are costly, which checks are quick, and which tasks justify a stronger model?
Route by difficulty and consequence
A sensible routing policy begins with task classes rather than prestige. Formatting a known text, identifying an approved category, and drafting from a complete brief may suit a cheaper model. Reconciling conflicting evidence or resolving a sensitive customer issue may justify stronger reasoning and closer review. The route should also consider consequence. A low-effort task can still carry serious risk if it changes a payment instruction. Separate the ability to produce a draft from the authority to act on it. Price alone cannot determine that boundary.
Measure escalation as part of the design
A two-stage workflow can save money when the first stage handles routine work and reliably flags difficult cases. It can waste money when almost every item escalates or when reviewers must reconstruct everything from scratch. Track the escalation rate, duplicated context, and time spent transferring a case. Give the stronger stage the evidence already collected and the specific unresolved question. Do not simply ask it to redo the whole task. Routing should reduce the total work required to reach an accepted result, including the human work between stages.
Keep quality samples stable
Use a fixed set of representative tasks when comparing routes. Include easy, ordinary, and difficult examples, plus cases with missing information. Score factual correctness, completeness, format, and the decision to escalate. Rotate a smaller hidden sample to discourage overfitting to familiar examples. Evaluate the entire workflow after changing prompts, tools, or retrieval settings, because the model is only one component. A routing table that was sensible last month may become expensive after the input format changes or a new requirement increases the review burden.
Turn the pilot into a spending rule
Set a per-task budget and an acceptance threshold before the comparison. At the end, choose the cheapest route that consistently meets the requirement within the operational constraints. Keep the stronger route available for defined exceptions and monitor the cost of those exceptions. Report both the average cost and the worst ordinary cases so a quiet tail of expensive work does not disappear in the mean. The commercial opportunity in cheaper models is broader access to useful work. Realizing it depends on measuring what the organization accepts, rather than celebrating a lower sticker price.