A working illustration of architecture selection. Adjust the task and constraints, then inspect what remains feasible.
Performance per dollar
Every task has its own frontier.
Capability, cost and control belong in the same conversation. Change the task and its requirements. See which combination earns its place.
Illustrative study · Synthetic data
Evaluation tasks passed (%) Cost–quality frontier
Lower costCost per task (A$)
Minimum share of evaluation tasks passed
Routed system
Match effort to the task.
Routine work takes a compact path. Difficult cases take a more capable path. Evaluation determines where the boundary belongs.
A$0.029per task
95%tasks passed
Retrieve → Route → Verify
In a real engagement, your tasks set the test. Quality, response time and total operating cost set the bar.
Inspect the study data +
Example values illustrate architecture selection. Quality is the share of tasks passing evaluation; cost covers the complete route. The line joins candidates for which no alternative is both cheaper and at least as capable. The selection is the lowest-cost candidate meeting your requirements. This study runs entirely in your browser.
Synthetic measurements. Response time is p95.
System
A$ / task
Quality
Response
Compact model
A$0.007
78%
0.32s
Cache + compact
A$0.011
86%
0.44s
Retrieval + compact
A$0.016
92%
0.85s
Routed system
A$0.029
95%
1.35s
Specialist system
A$0.040
94%
1.90s
Large model
A$0.086
94%
2.70s
Large model + tools
A$0.125
98%
3.80s
Choose against the work.
The instrument selects the least expensive candidate that meets the required quality and optional response-time constraint. If none meets the requirement, it says so. The plotted frontier identifies candidates for which no alternative is both cheaper and at least as capable.
All values are synthetic. They describe no provider, model release or Form From production result. The study runs locally and makes no inference requests.
A real comparison would need representative tasks, a clear evaluation method, complete-route costs, latency distributions, versions, dates and operating assumptions.