10 · Cost

What it cost to build

Priced the same way, the human route costs roughly 11–69× more, even after billing the author's hours at a senior engineer's rate ($140/h, which puts this project at $33k–$41k). The token and API figures are counted exactly from the session transcripts; the human figures are estimates and the author's hours come from the transcripts and the author, and the page says how each was made.

By phase

Prices used

Published Claude API list prices in dollars per million tokens: the same table the first version of this page used, with the two models that arrived since (Opus 5.5 and Sonnet 5) added at their published rates. Cache reads are 0.1× input except where listed.

By model

Spend by day

The human estimate

A rough comparison, not a quote. The scope a team would face is what shipped: a citro3d rendering backend, a 32-bit sweep, audio on the DSP, stereo, lifecycle, an owned native renderer sized to the Old 3DS's instruction cache, the 64DD Expansion Kit, a bottom-screen dashboard and menus, and 2–4 player multiplayer over local wireless and the internet, on two consoles.

A check against people who have done comparable work: Ship of Harkinian, the Ocarina of Time PC port, was built by a volunteer team from November 2021, when the decompilation was finished, to its first release in March 2022; 2 Ship 2 Harkinian, the Majora's Mask port on the same libultraship base, shipped in May 2024, six months after its decompilation was finished (Wikipedia). Those are PC ports with no memory or frame budget to speak of. The N64-on-3DS projects listed on Pushing the boundaries are hobby efforts that, by their own READMEs, mostly target the New 3DS at 30 fps or with interpolation. Nothing here contradicts the range above; nothing narrows it much either.

The author's time

The transcripts hold 777 prompts typed into this project's sessions on 49 days between 08-10 and 10-26 (US Eastern time). Grouping prompts less than 20–45 minutes apart into sittings, and adding 15 minutes to each, gives 126–157 hours at the keyboard. The author confirms the total was about 125–150 hours: hardware testing, SD-card copies, photographing screens and reading the docs leave no trace in the transcripts, but they largely overlapped with sittings or were short, so the total is close to the keyboard time. Prompts fall on every day of the week and at every hour from 8 am to 3 am, so this was not only evenings.

No code was written by hand. The author's work was direction (what to build next, what to drop), judgement on evidence (which experiments were real wins), and the hardware loop the agents could not run themselves.

Side by side

Cost per outcome

How the census was made

docs/postmortem/gen/pa_census.py re-implements the first version's method and can be re-run. It reads every transcript under the project's two Claude Code directories (~/Code/G-Diffuser through v1.0.0, ~/Code/gdx-3ds since; 1,881 files: orchestrator sessions, Agent-tool subagents and deep-research workflow agents). Each assistant message id is one API call; the 2–5 streamed records of the same message are merged by taking the largest value of each usage field, so nothing is counted twice. Sessions started from the home directory that only mention the project, and every other project on the same account, are excluded.

The first phase's per-subagent and per-tool tables are in source-census.md. Its clearest cost lesson still stands: two 900-turn agents of 08-14 were left running through a credit pause and each re-read a very large context hundreds of times. An agent that has finished its brief should be stopped.