Two billion tokens in thirty days, for $16.64. That's the post in two numbers; the rest is what I learned while spending them.
The numbers come from DeepSeek's usage export, 2026-07-11 to 2026-08-09. The cheap model, deepseek-v4-flash, carried 1.96 billion tokens across 22,349 requests. The strong model from the same family, deepseek-v4-pro, got 42.8 million tokens across 914 requests. Ninety-eight percent of the tokens I generated last month ran on a model that lists at $0.14 per million input tokens — and 97% of its input hit the context cache, which DeepSeek prices at $0.0028 per million. That hit rate is what makes the math work; without it, the cheap tier isn't cheap. The flash spend for the month was $15.52 — about $0.008 per million tokens. The whole ledger — flash and pro together — was $16.64.
One honest caveat before the claim: my router also reaches OpenRouter and a couple of subscription gateways, so the actual frontier — Claude Opus, GPT-5.5 — shows up in other buckets with messier cutovers. The clean data is DeepSeek's own stack, and that's the slice I'm using here. The claim is only about this slice: within one vendor's stack, the cheap model carried 98% of a month's tokens.
That's the claim behind the title, with the receipts: the cheap tier carries the load. It isn't a compromise I'm tolerating; it's the default state of the system — provided the loop is closed around it.
The reason I can get away with it is the loop. More capable models are expensive because they carry a verification loop internally — they check their own work, and the self-checking is what you're paying for. An internal loop is the model reasoning about its own correctness; an external loop is the model being corrected. A cheap model with an external loop produces the same outcome for a fraction of the cost, because the check lives outside the model. Tests, type checks, a build that fails loudly, a reviewer that reads the output — a task with one of those is a task a cheap model can do. The more capable model stays where the loop has to close inside the model.
The proof is in how my subagent routing is configured. Scouts run on the cheap tier. Workers run on a code-tier model — kimi-k2.7-code — subscription-billed, so it never appears in this DeepSeek ledger. Review runs on a strong tier — glm-5.2, through OpenRouter, so it's outside this export too — the checker gets the intelligence, not the generator. And the no-loop jobs — the hard second opinion, the adversarial security pass — run on the reasoning tier, the actual frontier in my stack. It's gated: the orchestrating model asks before launching it, because it's the one that costs real money. Every token in this export came from the cheap tier or v4-pro. v4-pro is the direct main-session model — interactive sessions talk to DeepSeek outside the router, and the export shows all 914 requests came through that path.
This is the second version of this post. The first version claimed cheapness itself was the insight — "most agent work is bounded, so a cheap model is enough." I wrote it, sat with it, and deleted it, because it was the wrong interpretation of the same evidence. Cheapness was never doing the work; the loop was. A cheap model without a verification loop is just a cheap model, and it will confidently ship garbage. The configuration survived the rewrite. The claim didn't.
Here's what this costs, honestly. First: cheap hides waste. My router tracks cache statistics, not value — I can't tell you which of the two billion tokens earned their keep. The ledger proves volume, not quality. Second: the loop moves trust from the generator to the checker, and the checker is itself a model. When the reviewer misses, garbage ships with full confidence. There's no third layer — two is where I stop. The strong model's self-checking was what I gave up, and I replaced it with a cheaper checker that can also be wrong.
Two billion cheap-tier tokens became the month's work: six blog posts, a router migration onto the cluster, a NixOS config that produces its own agent policies, and a product pivot that cut sixty-six thousand lines. $16.64 for all of it — the v4-pro share of that bill was a dollar twelve.
The reasoning tier — the actual frontier in my stack — is invisible in this bill, its spend buried in prepaid subscription credit. The cheap tier carries the load; it just needs the loop closed around it.