"Get the subscription, it's cheaper" is the most common piece of AI cost advice, and it is almost always given without a number attached. Cheaper past what? Everyone nods and nobody checks.
So I checked, using 34 days of my own logged usage. The answer is a single figure you can hold in your head.
Why a blended rate, not a list price
You cannot answer this from a price sheet, because a price sheet has four numbers per model — input, output, cache write, cache read — and your bill depends entirely on the mix. My workload is agentic, so it is overwhelmingly cache reads: the model re-reading context it already has.
Across 9.96 billion tokens and 23,157 API calls, priced at published list rates per model, the effective blended rate came to $0.6492 per million tokens. Not the list input rate of $5. Not the output rate of $25. Sixty-five cents, because the cheap token class dominates the volume.
| Plan | Buys at API rates | Roughly equal to |
|---|---|---|
| $20 / month | 31M tokens | A few hours of agent work |
| $100 / month | 154M tokens | About two days |
| $200 / month | 308M tokens | About four days |
Calculated at $0.6492 per million tokens — the blended rate measured on this workload. A different token mix moves this number, sometimes a lot. See the method note.
The part that surprised me
154 million tokens sounds enormous. It is not. My own usage ran to 6.45 billion tokens a month — 41.9 times past the breakeven point.
The question isn't whether you're past the line. It's whether you noticed crossing it.
That gap is the entire argument. Once a workload is genuinely agentic — long context, many tool calls, re-read on every turn — token consumption stops resembling chat usage and starts resembling a data pipeline. Gartner puts agentic workloads at five to thirty times the tokens per task of a chatbot turn, and that matches what I see.
Where the subscription stops helping
Two limits, and the advice never mentions either.
A plan has a ceiling. It is not unmetered. Sustained agentic load runs into fair-use limits, and when that happens the cost saving converts into a throughput problem — which is worse, because now you are blocked rather than merely billed.
The breakeven moves with your mix. Mine is unusually low precisely because caching does so much work. A workload with poor cache hit rates pays closer to full input price, the blended rate rises, and the breakeven falls even further. Ironically, the worse your caching, the sooner the subscription wins.
How to work out your own number
- Take a week of real usage and total each token class separately — input, output, cache write, cache read
- Price each at its own published rate, not the headline input rate
- Divide total cost by total tokens: that is your blended rate
- Divide your plan price by that number. That is your breakeven, in tokens per month.
- Don't estimate from the input list price — on a cached workload it overstates cost by roughly 8×
If the answer comes out anywhere near mine, you are not deciding whether to subscribe. You are deciding how much to have overpaid before you checked.
Method & disclosure
Derived from the same dataset as the 34-day cost study: 381 session transcripts, 23,157 billable calls, 18 June to 3 August 2026. Token counts are actual, read from per-message usage records.
Costs are calculated, not invoiced — published per-model list rates applied to measured token counts, with cache reads at 0.1× input and cache writes at 1.25× on the five-minute TTL. The blended rate is total calculated cost divided by total tokens. It is specific to this workload's token mix and will differ from yours.
No commercial relationship with any vendor named. No affiliate arrangement on anything in this piece.