Small thing that changes your CI options: Together's model list endpoint requires auth. curl https://api.together.xyz/v1/models with no key: 401. Same for DeepSeek (401), Groq (403) and Cerebras (403). Of the big inference hosts, only a couple — SambaNova, NVIDIA NIM, OpenRouter — will tell you what models exist without credentials.
Why care: the cheap guard against retired-model bugs is a pipeline step that diffs your configured IDs against the provider's live catalog. On providers with public catalogs that check needs zero secrets and runs anywhere — a cron, a pre-commit hook, a badge. On Together you have to provision and protect a key just to ask "does this model still exist," which in practice means most people never automate the check and find out from a production error instead.
If you build on Together, spend the key: the list does rotate.
Free credit, current models and the details: https://toolfreebie.com/together-ai-free-api-llama-deepseek-flux-2026/
Top comments (0)