Model Arena & Benchmarking

Compare multiple foundation models concurrently on identical prompts with sub-millisecond telemetry tracking.

Time-to-First-Token (TTFT)

Measures the exact latency between request initiation and the first chunk streamed from the upstream provider.

Streaming Throughput (TPS)

Real-time tokens per second generated during streaming responses to evaluate generation velocity.

Real-Time Cost Delta

Calculates exact USD cost per run for each model and highlights percentage savings across alternatives.

Blind Battle Mode

Mask model identities (Model Alpha vs Model Beta) for unbiased qualitative human evaluation before revealing winners.

How to Run an Arena Battle

Navigate to /dashboard/arena.
Select 2 or 3 models from the catalog (e.g. anthropic/claude-sonnet-5 vs deepseek/deepseek-v4-pro).
Enter your test prompt or reasoning benchmark.
Click "Execute Arena Battle" to stream concurrent responses side-by-side.

Zero Data Retention Guarantee

Just like the main API Gateway, the Model Arena operates under strict Zero Data Retention (ZDR). Benchmark prompts and streamed completions are evaluated in your browser runtime and are never persisted to disk.