Every cost comparison I see starts and ends with the same column: dollars per million tokens. It is the only number a vendor publishes, so it is the only number that gets compared. It is also the number that predicts your invoice worst.
This week three conversations converged on the same point from different directions: an agent engineer counting merged pull requests, a writer pricing per-request crypto payments, and a quota analysis of a model that got cheaper per unit and more expensive per task. Here are the three numbers they were actually arguing about.
1. Retry rate: the multiplier you never put in the spreadsheet
A retry is billed twice for the same work. If a provider fails 8% of requests and your client retries them, your effective input price is 1.08x the sticker:
effective price = sticker price x (1 + retry_rate)
$2.00 / 1M input -> $2.16 / 1M at an 8% retry rate
Sixteen cents does not sound like much until you notice it is larger than the gap between the two models you were choosing between. A model that is 8% cheaper per token and 10% less reliable is not cheaper; it is the same price with a worse tail.
The number to instrument is trivial: count requests, count retries, count completed generations. The ratio between requests and completions is money spent producing nothing, and unlike cached-token savings it does not require you to change any code to measure.
2. Whether your cap is a stop or a suggestion
Prepaid balances are an underrated feature. They are a hard stop: an agent stuck in a loop hits the ceiling, fails loudly, and someone gets paged. The spend cannot exceed what you topped up.
Per-request settlement removes that stop. Brave's new agentic payments proxy bills $0.005 per search in USDC, with no account and no key — genuinely faster to ship, and a real improvement for ephemeral CLI agents. But a runaway loop no longer dies at a cap; it dies when the wallet empties or a human notices. Crypto has no natural daily boundary unless you build one.
The arithmetic is unforgiving because it is multiplicative rather than additive:
worst-case spend = rate x requests-per-minute x minutes-until-someone-looks
At $0.005 a call and one call per second, an unnoticed loop is $18/hour. At a chat-completion price it is four figures an hour. The cap is not friction; it is the only thing bounding the product of two numbers you do not control.
So when migrating off prepaid keys, carry the ceiling over explicitly — per agent, per day — and measure cost per successful call rather than cost per call. The difference between those two is your retry tax, again.
3. Merge-slot yield, for agents that ship code
One team counted 416 merged pull requests in a month from engineer agents, and were honest about the quality side: no person had read most of the code, so types, tests and model review were the gate.
Their real ceiling was not agent throughput. It was the merge pipeline: one pipeline per merge, about 44 minutes, which capped a long day at roughly twenty merges. Batching already-verified PRs into a single pipeline moved the number more than adding more agents did.
That gives the third number:
merge-slot yield = merges completed / merge slots reserved
If a broken PR occupies a 44-minute slot and then fails, that is not a 44-minute delay, it is a lost slot out of twenty. Running types and tests before a slot is reserved turns that failure into seconds. Same agents, same code, more merges — because the expensive resource was the queue position, not the model.
What this adds up to
The price per token is the only one of these four numbers you can copy from someone else's page. The other three you have to measure in your own system:
- retries per request,
- spend ceiling per agent per day,
- merges completed per slot reserved.
A model comparison that ignores them is not wrong about tokens. It is just answering a question nobody asked — what a unit costs, rather than what a finished piece of work costs.
If you want the unit prices themselves in one table, including which units each vendor actually bills on, I keep that here: aiapi-pro.com/chinese-ai-api-pricing. The three numbers above are the ones you still have to measure yourself.
Follow-ups on the same bill: the KV cache is the hidden cost of splitting prefill and decode, a quote lock's TTL is a cost parameter, and your fallback provider is the thing that breaks first.
Top comments (0)