What the Taiwan campaign cost to run

Inference only, not a total. Move the sliders until the assumptions are where you want.

total inference cost across the four-day campaign

Model turns

Input tokens

Output tokens

Per day

Rate card & units

Assumptions

Files produced

Dream disclosed 1,395 files.

Model turns per file

Recon, retries, and dead ends.

Input tokens per turn

Agent loops resend accumulated context every turn.

Output tokens per turn

Flash, so tool calls and short reasoning.

Prompt cache hit rate

Set to 0 for the worst case where the operator wasted the discount.

Your assumptions against the presets

Table view

Scenario	Turns/file	Turns	Input	Output	Total

Excluded: infrastructure, proxies & residential IP pools, credential acquisition, the operator's own hours, and agent framework development. This is the inference line only.

Rates. DeepSeek official, current July 25 2026: $0.14 per 1M input on a cache miss, $0.0028 on a cache hit, $0.28 per 1M output. OpenRouter lists V4-Flash at $0.0679 and $0.168. The official higher rates are the default here on purpose.

Big Macs. One US Big Mac is $6.22, per The Economist's July 2026 index. Switch the dropdown above to price the campaign in burgers instead of dollars.

Timing. The campaign ran July 1 to 4, 2026. DeepSeek released V4-Flash weights on Hugging Face on July 31, 2026. Whoever ran this paid API prices. Anyone repeating it now doesn't have to.