gencast-cost-explain

gencast cost-explain

Read a gencast cost.json and translate it into a plain-language breakdown of where the money went, plus concrete optimisations. Compares predicted vs actual cost when the original notebook is also provided.

Rate table for the bundled-default models

The rate table below is injected at skill-load time:

gencast estimate --rates-only --json 2>/dev/null | jq -r 'to_entries[] | "- **\(.key)**: $\(.value.input_per_1k * 1)/1k in, $\(.value.output_per_1k * 1)/1k out"' || echo "(gencast not installed or --rates-only failed)"

Version check (need gencast>=1.2.0): !gencast --version 2>/dev/null | python3 -c "import sys,re; v=sys.stdin.read().strip(); m=re.search(r'(\d+)\.(\d+)', v); sys.exit(1) if not v else print(v if m and (int(m[1]),int(m[2]))>=(1,2) else f'TOO OLD: need gencast>=1.2.0, found {v} — run: pipx upgrade gencast')" 2>/dev/null || echo "MISSING — install with: pipx install gencast"

Prerequisites

• A cost.json file produced by gencast generate. Located at <output_dir>/<notebook_name>.cost.json.
• Optionally, the original notebook.yaml (lets the skill compare predicted vs actual via gencast estimate).

Workflow

1. Read cost.json. Validate it has the expected schema:

   • Top-level total_usd: float and stages: [...] or stages: {name: {...}}
   • Each stage: {name|kind: str, model: str, usd: float, ...} (or compatible) If the schema does not match, stop and tell the user this does not look like a gencast cost.json.

2. Compute proportions. For each stage, calculate (stage_usd / total_usd) * 100. Order stages by descending USD.

3. If the original notebook is available, predict vs actual.

   gencast estimate <NB.yaml> --json
   

   Parse total_usd from the prediction. Compute delta % against actual. Print: "Predicted $X, actual $Y (delta Z%)."

4. Plain-language breakdown. Walk top-3 most-expensive stages:

   • "Transcript was 60% of cost ($0.18) because Sonnet 4.5 ran on 6 segments. Switching to Haiku 4.5 would save ~$0.13 (-72%)."
   • "TTS was 30% ($0.10) because tts-1-hd at $0.030/1k chars × 4,500 chars."
   • "Whisper was 10% ($0.04) — 6 minutes of audio × $0.006/min."

5. Suggest optimisations. From the rate table above + the cost breakdown:

   • Cheaper outline/transcript model. If using Sonnet → suggest Haiku. If using Opus → Sonnet.
   • Fewer segments. If num_segments was high in the notebook (visible in cost.json's per-segment count) → suggest 4–6 instead.
   • Standard TTS. If tts-1-hd was used → suggest tts-1 for 50% TTS savings.
   • --cache-llm flag. If multiple runs of the same notebook are likely → enable LLM cache.
   • Local LLM. When v1.3 ships local-Ollama profiles → mention as a future option (not yet shipped — flag as future-work).

6. Output format. 5–8 lines of prose, optionally followed by a 3-bullet "what to try next" list.

What to NOT do

• Do not modify cost.json. It's a record of what already happened.
• Do not invent rates. Use only the rates from the injected table above (they are sourced from litellm.model_cost, the same source gencast uses at runtime).
• Do not promise specific savings — predicted savings are heuristic. Always say "approximately" or "around".
• Do not run gencast generate automatically. This skill is post-mortem only.