Developers overpay in two directions at once: stacked subscriptions that each include model access, and API loops that quietly multiply the flagship rate by the number of steps. This lesson takes one of each.
Scenario 7: Cursor Pro and Claude Pro
Leo pays $20 for Cursor Pro, which includes model access inside his editor, and $20 for Claude Pro, which he opens in a browser tab to ask coding questions and occasionally to draft an email. Both bills renew monthly.
Which would you trust?
Which is the cheapest setup that still does Leo's job?
Scenario 8: every step of an agent loop on a flagship
A support team runs an agent that reads a ticket, searches the knowledge base, checks the order system and drafts a reply: the site's 12-step reference trace, 166,800 input and 3,600 output tokens per run. They run 1,000 tickets a month, every step on GPT-5.5.
| Setup | Per run | 1,000 runs a month |
|---|---|---|
| GPT-5.5 every step ($5 / $30) | $0.94 | $940 |
| Claude Sonnet 5 every step ($2 / $10) | $0.37 | $370 |
| Gemini 3.8 Flash every step ($0.75 / $3.75) | $0.139 | $139 |
| DeepSeek V4 Flash every step ($0.05 / $0.16) | $0.0089 | $8.90 |
| Routed: Gemini 3.8 Flash first, Sonnet 5 re-runs the 10% it fails | $0.139 + 0.1 × $0.37 | about $176 |
Verdict: the routed setup. Gemini 3.8 Flash follows strict tool schemas well enough to run unattended on routine tickets, and escalating the tenth that fails to Sonnet 5 keeps the outcome quality without paying the flagship for the nine that did not need it. That is $940 down to about $176 a month, over $9,000 a year, with prompt caching still to come (it roughly cuts the Sonnet portion by two-thirds). The frontier-for-every-step design is not buying reliability; the caps and the escalation are.