Measured · AI usage

What AI-assisted engineering actually costs

Every figure on this page is derived from the provider's own usage exports for2026-07-31 to 2026-10-09 — 24,800 requests, 3.2B tokens,$53.29. Nothing is extrapolated and no usage is excluded: it is the practice's complete AI consumption for the period.

What the work was

Concrete deliverable counts from version control across the same period. No project names and no per-repository split — the work is characterised by kind, not by client or product.

Repositories32codebases worked in
Commits1,47920.8 per day · 68 of 71 days
Cost per commit$0.03627.8 commits per $
Agent turns per commit17model turns
Generated per commit16,8752.2M tokens of context

Work profile — software engineering across independent codebases: a desktop application built on a Rust core with a GTK front end, web front ends, CLI tooling and developer libraries, alongside research, design and business operations.

Deliverables against spend

Cost tracks shipped work rather than idle consumption — both series measured over the same 71 days.

What it cost

Provider-reported spend. Cost is always taken from the rate actually charged on each line of the export, never recomputed from a price list — the model line was relabelled on 10 September 2026 (V4 Flash → V4.1 Flash) and prices moved during the period.

Spend$53.29over 71 days
Per day$0.751$0.0022 per request
Tokens per $60.8M$0.01646 per 1M tokens
Generated per $468,307130,573 tokens per request
Cache reuse98.94%of all input tokens
Requests24,800465 per $

Daily spend

$53.29 over 71 days — $0.751 per day, every day of the period used.

Per model line

One series per line, continuous across the 10 September relabel.

Model lineTokensGeneratedRequestsCostCache reuseTokens / $Generated / $
Flash tier (V4 → V4.1)3.1B23,425,01023,353$42.2999.0%72.2M553,904
Pro tier (V4 Pro)184M1,533,3121,447$11.0098.1%16.7M139,340

What the method bought

The practice's AI work runs on a convergent working profile: context that is stable, revised in place and reused, rather than rebuilt per request. The provider charges a small fraction of the input rate for context it can serve from cache, so a stable context is what makes long-context work affordable. These three charts are that effect, measured.

Cache reuse

Share of all input tokens served from the prompt cache — 98.94% over the period, and it rises as the profile converges.

What the tokens are

Re-read context (cache hits) against freshly sent input and generated text — three orders of magnitude apart on a log scale. 98.18% of all tokens moved are re-read context; 25M tokens were actually generated.

The same workload without cache reuse

Identical token volume with every input token charged at the observed cache-miss rate instead: 9.9× to 27.3× what was actually paid.

How to read these numbers

  • Cost per commit ($0.036). The denominator is non-merge commits in the same period, counted from version control. It is unweighted — a documentation commit counts like a feature commit — so it is a headline comparison, not a productivity metric.
  • Tokens per dollar is a cache figure. 98.94% of input is served from cache because the context is reused, so the blended60.8M tokens per dollar measures context reuse as much as work. The cache-independent figure is 468,307 generated tokens per dollar.
  • Nothing is filtered. All requests, models and clients in the period are included, including research, planning and failed attempts. A benchmark that removes the unproductive work is a brochure.
  • Source. Provider usage exports, ingested row by row; spend is the sum of the rates actually charged, cross-checked against the provider's own cost column to the cent. Figures regenerated 2026-10-12.