Coding Plan vs API: Stop Paying for Tokens Twice
7–10B tokens/mo on a plan, not an API bill the number that matters here
By Stan K.
A $10 coding-plan budget bought ~1.9B tokens/month at API prices, and our mid-tier agent plan moved 7–10B tokens in a month. The same volume through raw APIs is a car-payment bill. When the subscription model quietly wins — and the two cases where it doesn't.
The default mental model for LLM costs is the API price list: so many dollars per million tokens, metered, honest. And then there is the second model almost nobody prices correctly: the coding plan — a flat monthly subscription that wraps API-class models behind a quota of prompts or credits.
We run both, daily, for real work — agent sessions that orchestrate sub-agents, scan pipelines, write and review code. Our own billing history makes the comparison unambiguous: for volume users, the plan is not “cheaper per token”, it’s a different economic category. Here is the honest math, and the two cases where we still pay for raw API.
The counterfactual, in real numbers
Our most heavily used month on a mid-tier agent plan moved roughly 7–10 billion tokens — orchestrator sessions, sub-agent fan-outs, cached context re-reads, the lot. (The credit mechanics are documented separately — a 5-hour credit window, tokens weighted by input/cached/output.)
What would that month have cost at raw API prices? Even at the cheapest production-grade rates on the market — DeepSeek Flash at $0.30/1M input, $1.20/1M output (peak; off-peak is −50%), or GLM-5.3-Flash at $0.15/$0.50 — single-digit billions of tokens are a three-figure monthly bill. At frontier rates (Claude Sonnet 5.5 at $2/$10, GPT-6.1 Sol at $2/$10, GLM-5.3 at $1.4/$4.4) the same month is four figures. The plan that carried it costs less than a dinner for two.
×6 — the verified floorThe cheapest plan we tested
The $10 OpenCode Go tier includes a $60/month model budget on GLM-5.3-Flash — six times its own price in real API dollars, every month, with the budget denominated in actual dollars rather than abstract points. Community reports confirm $10 comfortably runs three parallel agents on DeepSeek Flash without exhausting it. The multiple is model-dependent (on GLM-5.3 proper the budget is $15 = ×1.5); on the discount workhorses it’s the best dollar-for-dollar ratio we’ve verified.
Why the plans can do this
Not charity — three structural reasons:
- Off-peak and pooled capacity. Plans route you into capacity the providers would otherwise waste; Z.ai’s own API prices drop 50% off-peak, and plans are effectively always priced near the off-peak floor.
- Weighted credits, not raw tokens. Cached input counts at a fraction of fresh output (the published formula weights cached reads roughly 4× cheaper than input and output 3.5× dearer). Agent workloads are cache-heavy, so the effective token count is far below the raw count.
- You are the workload shape they want. Coding-plan users generate exactly the long-context, high-cache-hit traffic that costs providers the least to serve.
The two cases where we still buy raw API
The plan model is not a universal win, and pretending otherwise is how people strand projects:
- Programmatic product calls. A plan quota lives inside coding tools and their validated CLIs. The moment tokens are served to your users — a feature in your product, a batch job on customer data — you need the API (or a broker), and its price list is the real one. Our video pipeline economics and TTS per-minute math are API-side costs precisely because they are product calls, not coding sessions.
- Guaranteed latency and rate limits. Plans are best-effort pools with 5-hour windows; an API with provisioned throughput is the contract you need when a customer is waiting.
The decision, compressed
If you touch an LLM every working day and the traffic is you-building-things, price the plan first — our Coding Plans hub ranks all ten by exactly this value math. If your traffic is your product calling a model, budget at API prices and read the fine-print first: the invoice traps — invisible thinking tokens, cache write charges, context multipliers — add 30–200% to naïve estimates.
The expensive mistake isn’t choosing the wrong vendor. It’s paying API rates for a coding workload a $10–30 plan already covers.