Chat completions

POST /v1/chat/completions mirrors the OpenAI Chat Completions API: the same model, messages, temperature, max_tokens, tools, tool_choice, and stream fields are accepted. Unknown fields are forwarded transparently upstream. The model must be a catalog id from GET /v1/models. An unknown model id is refused with 400 rather than passed through to an upstream. The simplest choice is the flagship id sable, which Sable keeps pointed at a strong frontier engine (currently disclosed on the Models page) without breaking callers when it improves.

Every response carries Sable-specific headers:

• x-sable-node: what served the request. On inference this is always gateway: one gateway process, named honestly, not a claim about separate hardware. On the confidential tier the enclave that actually ran the model is identified by its measurement and signer in the receipt's attestation block. GET /v1/nodes lists the same set: the gateway, one entry per configured attested backend, and any enrolled fleet nodes.
• x-sable-region: the region this deployment declares it runs in, or unspecified when the operator has declared none
• x-sable-privacy-tier: the tier it executed under
• x-sable-receipt, x-sable-receipt-sig, x-sable-receipt-signer: the signed, metadata-only receipt (streaming responses emit it as a trailing sable.receipt SSE event instead, since headers are already flushed)

Sable extensions

Three additional body fields are accepted:

• sable_privacy_tier: overrides the tier inherited from the API key for that single call. See Privacy tiers.
• sable_region: pins the request to a region. A pin is honored only by a deployment that declares it serves that region; any other region is a 400, never a silent reroute. See Region pinning.
• sable_scrub: set true to have the gateway redact secret/PII shapes (emails, API keys, EVM addresses, long hex) from the outbound prompt in memory before it reaches any vendor; the receipt records scrubbed: true. Also accepted on /v1/messages. See Outbound scrubbing.