What should you run at what cost?

Active Applied research, actively maintained.

Signal Room tracks a selection of high-impact models, provider-native reasoning configurations, speed evidence, agent harnesses, and self-hosting economics.

Model roster & configurations

Anthropic, OpenAI, Google, xAI, Meta, Mistral, Cursor, and major Chinese labs, with provider-native reasoning effort, thinking levels, and token budgets.

Read the data methodology

Speed & cost evidence

Time to first token, output throughput, task latency, and quota-burn cross matrices – vendor claims and unknowns are labelled, never blended together.

Open the Signal Room

Agent harnesses & self-hosting

Claude Code, OpenCode, Codex CLI, and more, alongside dated GPU-hosting prices, local Mac configurations, and open-weight model fit.

Read the research-data policy

Choosing a model, a reasoning configuration, an agent harness, or a self-hosting setup means comparing vendor claims that are rarely apples to apples. Signal Room is a non-exhaustive, inspectable report: every number is either measured, vendor-reported, or explicitly marked unknown, and every claim keeps a canonical source link.

Inspect the evidence, not just the conclusion

Every figure in the report traces back to an evidence class, a source, and an update date. Use it to reach your own conclusion, not to take ours on faith.