The task catalog
| Task | Token sketch | Low | Likely | High | Uncertainty |
|---|---|---|---|---|---|
| Quick chat Q&A | 2k in, 0.5k out | $0.01 | $0.02 | $0.05 | ±2x |
| Long document summary (about 75 pages) | 100k in (cold), 2k out | $0.40 | $0.65 | $1 | ±1.5x |
| Long chat thread (30 turns on a 50k-token context) | mostly cache reads | $0.50 | $1.50 | $4 | ±3x |
| Agentic coding, 1 hour, one interactive agent | 20-60M tokens, about 96% cache | $2 | $8 | $25 | ±5-10x (effort level and number of parallel agents) |
| Agentic coding, 1 hour, headless or 3 parallel agents | as above x3 x1.49 | $10 | $35 | $75 | ±5x |
| Code review of a pull request | 30-80k in, 3k out | $0.20 | $0.50 | $1.50 | ±3x |
| Deep research report | many searches plus a long synthesis | $1 | $3 | $8 | ±4x |
| Codex local message (ChatGPT plans only) | one Codex message | counted as messages (ChatGPT only) | OpenAI's own range per model | ||
Ranges were drafted at an Opus-class list price of about $5 per M input and $25 per M output tokens (Sonnet about 0.6x, Haiku about 0.2x). Re-checked 20 Oct 2026: Claude Opus 5.5 lists at $4 / $20 per M (claude.com/pricing), about 20% lower, so the high ends are conservative for Opus 5.5. On GitHub Copilot plans the calculator rescales every row to the chosen model's per-token rate published by GitHub (see methodology).
Not modelled
These jobs are counted in vendor-specific credits, not API dollars, so the calculator leaves them out:
- Image generation: Counted in vendor credits (Google Flow credits, the X/Grok Imagine pool), not API dollars.
- Video clip (8 s): Counted in Google Flow credits.
Agentic coding dominates almost every power user's week, so its range matters most. See the methodology.