CollageMaker — LLM Usage & Value Report

Effective Tokens = uncached input + output + reasoning (truly new work). Raw Tokens = all tokens as recorded by opencode, including redundant context re-sent each turn in multi-turn sessions. The difference is the estimated prefix-caching savings.

Period: 2026-07-01 to 2026-08-04 · 10 models · 18 agent roles · Generated: 2026-08-15

Cost estimates: cloud-equivalent per-model pricing (actual cost: $0, all models run locally via LM Studio).

$307.93 saved via prompt context caching (82.1% reduction, 94.2% cache hit rate).

48.1M tokens consumed across 25 commits producing 7,474 lines of code. (est. effective)

Notes: summary_files is only populated for sessions where opencode recorded file changes. Most sessions are exploratory, research, or cross-session coordination. Build agent productive sessions are estimated from token-ratio approximation (build tokens / total effective tokens per day) since per-agent daily session counts are not available. 10 days had 19.4M tokens with zero commits (e.g., research, debugging, cross-session coordination).

Tokens vs. Commits

Tokens per Commit vs. per Line Added

Cumulative Tokens vs. Cumulative Lines Added

Cumulative Estimated Cost

Rolling 7-Day Tokens per Commit Trend

Daily Token Breakdown by Agent Category

All token figures in this section are effective tokens (cache-adjusted). This timeline cross-references them with 233 session summaries written by the build agents in _agent_docs/project-timeline/sessions/ (2026-06-30 → 2026-08-03). Only build-* and plan-bdd sessions write summaries, so the “Summaries” column is a floor for daily activity — the “DB Sessions” column counts all sessions recorded for this directory.

Day-by-Day: Tokens vs. What Was Happening

Patterns Visible in the Timeline

• build-docs ran every day — skill refinement, code reviews, and plan authoring ran in parallel with feature work. It is the single largest consumer (13.5M effective tokens, 121 sessions); the “skill refinement” loop distilled each day’s learnings into reusable reference docs for the building-web-apps skill.
• plan-bdd specs preceded nearly every multi-phase feature — title/sidebar/home, crop overlay, undo/redo expansion, mobile responsiveness, bottom sheet, and the two multi-touch fixes all started with a written BDD plan before the first TDD session.
• Token intensity decayed as the codebase matured — from 5.3M effective tokens/day at the greenfield peak (Jul 5) to ~0.6–0.9M/day in the polish weeks, as change units shrank from 31-file features to targeted bug fixes.
• Zero-commit days were plan/debug days — Jul 4–5, 20, 24–26, 30, and Aug 2 had no commits; their planning and diagnosis cost shows up in the next day’s commits (e.g., the Jul 19 pre-merge review directly produced the Jul 20–22 undo/redo expansion).
• 233 of 517 sessions have written summaries — the remainder are subagent runs (explore, world-review, diff-review) and short sessions that were not summarized. Summary coverage skews to the build-* agents, which were required to write session summaries.

All models run locally via LM Studio — actual API cost is $0. Cloud-equivalent costs below show what these sessions would cost if run through commercial APIs at current per-model pricing.

$307.92 saved via prompt context caching (82.1% reduction, 94.2% cache hit rate).

Cloud-Equivalent Cost by Model

Assumed Pricing Rates

Per-model rates above are used for detailed cost calculations. Flat-rate tiers below are used for summary estimates.

Estimated Cloud Cost Over Time (Hypothetical)

Cache Approximation

Cache Impact by Model

Cache Impact by Agent

Cache Impact by Day