What does your overnight autoresearch loop cost?

Compute budget by model tier. The GPU executes the experiment; the LLM suggests and repairs the patch. The improvements-per-100 figures below are hypothetical editable assumptions, not benchmark pass rates.

ExperimentsTraining min / experimentGPU rental $/hourLLM calls / experimentInput tokens / LLM callOutput tokens / LLM call

API model / assumed class	API USD/run	API USD total	GPU USD total	Combined USD	USD per reproducible win

Budget tier: wins per 100 trials

Decent tier: wins per 100 trials

Strong tier: wins per 100 trials

This calculator deliberately does not guess model quality. Set the reproducible wins from your frozen 30+ experiment trial. Rates are USD per 1M tokens, first-party list, verified October 17 2026. Introductory Gemini Argon price expires Nov 30; DeepSeek has peak hours; context, caching, tool fees, agent overhead and instance startup are excluded. If the GPU is not rented by the training minute, bill actual instance uptime instead.

Autoresearch source · OpenAI rates · Claude rates · Gemini rates · DeepSeek rates