There's a genre of AI cost article that ends with "so we negotiated enterprise pricing." This is not that article. Everything here runs on consumer-grade money: one mid-tier token plan, one ordinary $20 subscription, a $10 deposit. The yield comes from architecture, not negotiation — treating every subscription you already pay for as a capacity pool with an idle problem.
The fleet
Primary: the mid-tier workhorse
The platform's daily driver is MiniMax M3 on a mid-tier token plan. Not the flashiest model on the board — deliberately. The primary model's job is volume with quality above a floor: planning decomposition, code generation, memory operations, the thousands of routine calls an agent platform makes per day. A mid-tier plan prices that workload sanely, and modern mid-tier models clear the quality bar for most of it. Paying frontier rates for classification calls is burning money to feel good.
Secondary #1: the subscription you already forgot
The best deal in the stack was already paid for. An existing ChatGPT Plus plan includes Codex usage that would otherwise evaporate unused every month — because like most people, the subscription's chatbot isn't the workshop's daily driver, it's a thinking rubber duck. OAuth into Codex turned that idle allocation into a real agent endpoint: a genuinely capable coding model whose marginal cost is zero dollars, because the subscription earns its keep either way.
This is the whole arbitrage in miniature: a dollar buys either one thing or two things depending on whether you're paying attention. The subscription was always bought; routing agent traffic through its included usage is pure recovered value.
Secondary #2: free previews as burst absorbers
The third leg is free-tier capacity: models in free preview on OpenRouter (including whatever model happens to be reading this page's server logs, hello), plus limited-use free models unlocked by a small $10 account deposit — about a thousand calls a day across them. Free tiers are flaky by nature: rate-limited, occasionally degraded, subject to disappearing without notice. That's fine, because they're not load-bearing. They absorb burst and overflow — batch jobs, parallel worker fleets, overnight build loops — where "usually available and literally free" beats "always available and billed."
The routing discipline that makes it work
- Nobody is load-bearing. Every task class has at least two homes. When the OpenRouter key hit its credit cap mid-month, work moved to the other legs without ceremony — because moving was already the design.
- Tier by consequence, not by habit. Frontier APIs get the rare task where output quality is the product. Mid-tier gets the volume. Free tiers get the burst. The expensive path is a choice made per task, never a default nobody revisited.
- Subscriptions are capacity, not identity. "I have ChatGPT" and "ChatGPT is my platform" are different architectures. The first is an asset; the second is a leash.
- Track burn, not vibes. Rough monthly accounting — tokens by source, dollars by leg — is what turns "it feels cheap" into "$60 for 8 billion tokens." You can't optimize what you refuse to measure.
Why this pairs with local-first instead of replacing it
The manifesto (Why Local AI Matters) argues for owning your stack; this article is what that looks like after the hardware pivot. The local GPU still handles embeddings, TTS, and auxiliary inference — the always-on workloads where latency and privacy matter more than ceiling. Rented intelligence handles reasoning. The fleet keeps the rented part cheap and non-exclusive. Local-first was never "own everything"; it's "own everything that's cheap to own, rent nothing exclusively."
// the one-paragraph version
Audit every subscription you hold for idle capacity. Route agent work through it before buying new capacity. Add free tiers as shock absorbers. Keep one paid workhorse for the volume. Never let any single source become the only path. That's the entire strategy — it fits in a paragraph and pays for the hobby.
Caveats, honestly stated
- Free previews disappear. Treat them as opportunistic, never foundational.
- Mid-tier models fail differently than frontier ones — verification gates matter more when the average model is cheaper.
- Subscription terms change. The arbitrage exists until it doesn't; the architecture assumes legs will vanish eventually.
- Your numbers will differ. The ratios are the point, not the totals.