benchmark-models

Cross-model benchmarking runs a single gstack prompt through Claude, GPT, and Gemini to compare latency, token usage, and cost. It outputs a comparative table so you can evaluate which model best meets your performance targets.

Yes, measuring latency and token cost differences between models is the core function. The benchmark uses the gstack-model-benchmark binary to process your prompt and generates a comparative table detailing token usage and cost for each model.

Evaluating output quality during model comparison is handled via an optional dedicated judge. This quality assessment quantifies real-world output alongside latency and cost metrics, providing a comprehensive view of model performance.

To run a model benchmark, you can provide prompts from inline text, files, or on-disk sources. The Skill processes these inputs through the gstack-model-benchmark binary and returns a comparative table plus optional JSON results.

Yes, the gstack model benchmark supports exporting results to JSON. Alongside the standard comparative table, you can generate optional JSON results to integrate the latency, cost, and token data into downstream workflows.

Cross-model benchmarking should be used before choosing a model for customer-facing workflows. It provides data-driven evaluation of latency, cost, and quality across Claude, GPT, and Gemini rather than relying on subjective vibes.