08 · Cost

What it cost to build

Every API call the project made is in the session transcripts on this machine. This page counts them. Token counts are exact. Dollar figures apply Anthropic's published API list prices to those counts; cache writes are billed at the 1-hour rate, the cache TTL Claude Code used in these sessions.

Prices used

Published Claude API list prices (platform.claude.com, pricing page, fetched 2026-09-03), in dollars per million tokens. Cache writes are billed at the 1-hour rate (2× input), the cache TTL Claude Code used in these sessions; billed at the 5-minute rate (1.25× input) instead, the total would be $4,280. Cache reads are 0.1× input on every model except Fable 5.1, where they are 0.025×.

Where the tokens went

By model

Three generations of model did the work. Fable 5 carried the first three weeks; Opus 4.8 carried the week the Fable credits were out (08-15 to 08-20, when the fog fix landed); Fable 5.1 carried the final week and the multi-core push.

Spend by day

Subagents

147 background agents were launched with the Agent tool (130 by the orchestrator, 17 by other agents), and 10 deep-research workflows spawned 1,103 short-lived search, fetch and verification agents. The table below is the twenty most expensive background agents; between them they are most of the subagent bill.

The two 900-turn agents of 08-14 (GPU profiling telemetry and the stereo foundation) were left running through the first credit pause and each re-read a very large context hundreds of times. They are the clearest cost lesson of the project: a long-lived agent with a big context is expensive per turn, and an agent that has finished its brief should be stopped.

Tool calls

Reconciling with what /usage and /stats show

The user's impression was that the project cost more than this page says. Three things explain the gap, and one real omission was checked and found small.

What a human team would have cost

A rough comparison, not a quote. The scope a team would face is the scope this project faced: a new fixed-function GPU backend, a 32-bit sweep across three codebases, a ROM-side audio driver port, an ARM11 performance grind to native 60 on a software display-list interpreter, stereo, a multi-core render pipeline, lifecycle, menus and packaging. About 20,000 lines of low-level code plus the research and the hardware loop.

Reference points

• sm64 on 3DS. Hobbyists, one core developer plus contributors. Playable in months, 60 fps via 30 Hz logic and interpolation, refined over years.
• Ship of Harkinian console ports (Wii U, Switch). One or two skilled volunteers, several months each, on top of an existing PC runtime, with a shader-capable GPU and no stereo.
• Professional retro-port studios. Five to eight people for six to twelve months on ports with far more headroom than an 804 MHz ARM11 driving a PICA200.

Estimate to reach this state

Basis: $150–200 per hour contractor rates or $20–25k per month loaded per senior engineer; low-level port work lands at 100–200 shipped lines per engineer-day once debugging and hardware loops are counted, which puts 20,000 lines at six or more person-months before the performance grind. A human team would spend its time in the same places the agents did (the 32-bit address-space heuristics, the fog LUT, the HUD backstop, the audio driver, then months of profiler-driven grinding), would very likely have shipped 30 Hz plus interpolation first and deferred stereo, and might never have attempted the core-2 render thread.

Side by side

Cost per outcome

Some ways to read the bottom line, at published prices with 1-hour cache writes: