Gcore Everywhere Inference (direct)

Direct calls to a Gcore Everywhere Inference deployment's OpenAI-compatible API (per-deployment base_url, e.g. https://model-<deployment>-<project>-<account>.ai.gcore.dev/v1 — Gcore issues one endpoint per deployed model rather than a single shared router URL), instrumented via the OpenAI SDK wrapper. No first-party Gcore wrapper exists in telara-observability; Gcore's own docs confirm the deployed-model endpoint is queried with the Python/JS OpenAI SDK, so the generic OpenAI wrapper pointed at the deployment's base URL captures spans the same way it does for other OpenAI-compatible endpoints. pip extra: pip install telara-observability[openai] Gcore hosts open-weight models from its Application Catalog (Llama, Mistral, etc.) on Gcore-operated GPU inference nodes and bills per minute of compute time, not per token — so token-derived model_price_table cost here is a best-effort proxy, not Gcore's actual invoiced figure. Gcore sits between the customer and the underlying open-weight model, so cost attribution lands on Gcore as the inference host, not on a separate upstream model vendor (there is none to bill — the weights are open and Gcore is the only biller).