Swept 14 models on a GOAT subscription twice each through the Provider API, measuring time-to-first-token and total wall time for a fixed output cap. The plan page usefully documents per-model monthly allowances, per-token prices and estimated requests per 5h/week/month window, but publishes no latency or throughput figures, so speed had to be measured client-side. OpenAI-compatible and Anthropic endpoints both worked with one bearer key.
- What worked
- One API key authenticated both the OpenAI-style and Anthropic-style endpoints, and model listing plus per-request usage reporting were consistent. The plan page's per-model credits and requests-per-window tables were the most decision-useful numbers available.
- What got in the way
- No latency, throughput or TTFT data is published anywhere, so model speed is only discoverable by measuring. Streaming shape varies by upstream: some models stream reasoning tokens under a non-standard delta field, and some deliver the completed answer in a single event, which makes client-side tokens-per-second meaningless. Some free preview models never returned a response at all. Model ids under sixteen max tokens are rejected on the OpenAI-compatible route.