About ProChoice
Transparent, daily LLM and coding agent harness rankings built by Retrograde Labs.
Why ProChoice exists
Public AI leaderboards routinely scatter scores across dozens of incompatible scales while hiding real task cost, mixing reasoning effort tiers, or burying coding-agent harness effects inside model scores. ProChoice solves that by placing models on a shared composite scale per domain and plotting score directly against real dollar cost per task. Models that cost more while scoring lower are removed from the default frontier so teams can pick the strongest option for a real budget in seconds.
Methodology in brief
Each day, ProChoice ingests 18 mandatory public benchmark and first-party pricing feeds covering 9 model providers: Anthropic, OpenAI, Google, SpaceXAI, Meta, DeepSeek, Moonshot, Alibaba, and Zhipu. All 45 active benchmarks contribute directly to their domain composite with zero shadow candidates. Scores are aligned via equal-1.0 anchor standardization and pooled with heteroscedastic REML Empirical Bayes using cross-benchmark joint-jackknife linking uncertainty. Missing task costs for non-retired models with verified first-party API pricing are estimated on each domain's homogeneous cost axis with active-set monotone effort regression and propagated through bootstrap Pareto probabilities.
Controlled coding-agent harness evaluations live in their own Harnesses Composite Index, combining FrontierHarness v1.1 and Harness Leaderboard into one 25-harness board across 8 open control models so harness design and model choice are never conflated.
Open source, maintainer, and licenses
ProChoice is built and maintained by Retrograde Labs. The complete web app, econometric scoring math, and daily ingestion pipeline are open source under the MIT License at github.com/RetrogradeLabs/pro-choice. Compiled ProChoice composite indices and API outputs are published under Creative Commons Attribution 4.0 with full upstream attribution to each source benchmark. See the API documentation for direct JSON and Markdown access.