How analysis works
Metergraph turns the traffic you already capture into evidence-backed ways to reduce LLM spend. It tests changes against your requests and your application’s answers, so each recommendation reflects your product rather than a generic benchmark.
An analysis can compare alternative models and search setups for a workload. Depending on your setup, results can also include batching and caching opportunities. You choose the traffic and changes worth considering. Metergraph runs the comparison and brings the strongest opportunities back to your dashboard.
From traffic to opportunity
Section titled “From traffic to opportunity”Each analysis starts with a workload and a representative sample of its captured traffic. Metergraph replays that sample against the alternatives you selected. It measures the quality, cost, and latency of each option against what your application produced. A model or search opportunity must cost less while meeting the quality bar.
Analysis runs in the background using traffic you already captured. Progress and results appear on the Opportunities page.
Quality comes first
Section titled “Quality comes first”The primary evaluation is task quality relative to your current setup. Your recorded production answer is the baseline, and each candidate is judged on how well it handles the same request.
Metergraph keeps the evidence behind the comparison, including the recorded and candidate answers, evaluation results, and explanations when available. The report also shows how many cases were evaluated, so you can review the result before making a change.
Read an opportunity
Section titled “Read an opportunity”An opportunity shows the proposed change and projected monthly savings. For graded model and search comparisons, it also shows the effect on quality and latency. Select View evidence to see what supports the recommendation, such as tested alternatives and examples or the eligibility and savings estimate.
What analysis costs
Section titled “What analysis costs”Analysis makes model calls for replay and evaluation. Metergraph provides those calls on the hosted cloud. A deployment you operate uses the provider credentials configured in your own infrastructure.