Tokens in 60 seconds
Models read and write in tokens — chunks of ~4 characters, about ¾ of a word. You pay for every token in (your prompt + files + history) and every token out (the answer plus hidden reasoning). A "small" question inside a big codebase session can silently carry 100k+ input tokens.
Context re-billing
Agentic coding tools resend large context on every tool call. Ten tool calls over a 60k-token workspace ≈ 600k input tokens for one task — before any output. Prompt caching can cut re-read costs ~10×, which is why cache-aware workflows feel dramatically cheaper.
Reasoning multiplier
High-effort reasoning modes generate long internal thought. It's common for thinking tokens to be 3–10× the visible answer. Output tokens usually cost ~5× input tokens, so reasoning is the most expensive lane on the highway.
Worked example: the 30-minute burn
Frontier model at $15/Mtok in, $75/Mtok out. An agentic session doing 60 tool-call turns:
Input: 60 turns × 80k context = 4.8 Mtok → $72
Output + reasoning: 60 × 6k = 0.36 Mtok → $27
Total ≈ $99 per half hour of heavy agent work.
Subscription plans meter this with rate limits rather than dollars, so a max-effort model can hit a 4-hour cap in minutes while an efficient one never does. Same wallet, ~8× the tokens per task.
Cutting costs without cutting quality
Route routine work to efficient models and save frontier reasoning for the hard 10%. Trim context: pass file excerpts, not whole repos. Cap reasoning effort per task. And watch cache hit rates — re-reading the same 60k tokens uncached is the single most common source of shocking bills.