Kart Paddock

Completion times are kept here as run history, not a speed ranking. Human approval delays, quota interruptions, tool waits, and harness behavior were not controlled equally.

Recorded from the original builds, excluding later reviews and deployment. Total tokens include input, output and repeatedly read cached context. They are not a count of newly written code.

Astra and Fable API costs are estimates using September 7, 2026 standard token rates and recorded caching, not subscription charges. Output includes reasoning. Hosting, separate tool fees, taxes and fast-mode premiums are excluded. Tokenizers differ between models.

Harvest Rush · GLM
OpenCode 1.18.29, build agent, High effort, via OpenRouter.
71,489,610 total tokens: 10,789,778 input, 60,447,168 cache reads, 129,561 output and 123,103 separately recorded reasoning tokens. Session and per-message totals reconcile.

GLM timing: 5h 25m 44s adjusted elapsed across three prompt windows. The follow-ups were “I added credits” and “keep going”; neither added gameplay requirements. The windows totaled 16h 56m 30s, with an evidenced 11h 30m 46s permission wait removed. Low-battery sleep occurred within that wait and was not subtracted twice. Tools, testing and ordinary response latency remain included. This is not pure model compute time or an uninterrupted run.

GLM cost: $1.77910687 recorded by OpenCode, not independently reconciled against an OpenRouter invoice. Astra and Fable costs above are standard API-equivalent estimates; their accounting basis differs.

Cinder Peak Rally · Muse Spark 1.3
OpenCode 1.18.29, build agent, High effort, using OpenCode’s contributor-free provider (not OpenRouter). One initial prompt, 96 assistant messages, no recorded errors; 1h 4m 26s from September 8, 2026, 4:43:42 p.m. to 5:48:08 p.m. Pacific. 9,464,673 total tokens: 253,728 input, 9,101,664 cache reads, 73,962 output and 35,319 separately recorded reasoning tokens. Recorded session cost is $0 because of contributor-free provider access. At September 9, 2026 published standard rates ($1.25/M uncached input, $0.15/M cache reads, $4.25/M output including reasoning), the same token usage would cost approximately $2.15 ($2.14685385). At paid Contributor rates ($0.10/M input, $0.002/M cache reads, $0.20/M output), it would cost approximately $0.07 ($0.065432328). These are alternative token-price estimates, not bills; fees and taxes are excluded.

HyperKart: Nebula Drift · Gemini 3.8 Flash
Antigravity 2.12.2, High effort. 43,933,660 total tokens from 355 generation records: 2,033,490 uncached input, 41,707,441 cache reads, and 192,729 output. Output includes 64,511 thinking and 128,218 response tokens; these are not added twice. Mirrored retry-accounting records were counted only once.

Gemini timing: 1h 18m 36s across two prompt windows (48m 0s and 30m 36s). Initial prompt: September 8, 2026, 10:54:08 p.m. Pacific. Final response: September 9, 12:36:31 a.m. Total wall time: 1h 42m 23s, including a 23m 47s quota pause. The only continuation was “Continue from where you stopped”. Tools and permission waits within the two windows remain included; this is not pure model compute time.

Gemini cost: $5.375909325 estimated API-equivalent token cost at September 9, 2026 standard introductory rates: $0.75/M uncached input, $0.075/M cache reads, and $3.75/M output including thinking. Antigravity did not record a dollar charge in these records. This estimate is not a subscription bill and excludes cache storage, separate tool fees, hosting, and taxes. The original submission is published without gameplay fixes.

Muse pricing · Google pricing · OpenAI pricingAnthropic pricing