Changelog

Changes that affect how you call Mortar, newest first.

October 30, 2026

API

• The prompt optimizer runs in shadow mode by default: it leaves every prompt as sent, and X-Mortar-Optimizer reports what it would have removed. A deployment can still turn it on, or off.

October 29, 2026

Documentation

• Published these docs: a quickstart, authentication, an API reference for every data-plane endpoint, and guides to streaming, cache-friendly prompts and choosing a deployment model.
• Added the model catalog, with sample input, output and cached-input prices.

API

• Added POST /v1/cost/analyze, which finds what a prompt wastes on prefill and how to fix it, and POST /v1/cost/estimate, which projects a monthly bill at a cache hit rate. Neither needs a key.
• Prompt caching is priced: a token read from the cache costs a tenth of the input price, and a token written to it 1.25 times. GET /v1/models lists each model's costPerMillionCacheRead and costPerMillionCacheWrite. cost.update and session.metrics report the tokens read and written and what each cost, name their baseline, and report a negative savedUsd when the cache cost more than it saved. POST /v1/cost/estimate prices Mortar's cache the same way, so at a 0% hit rate it costs more than the same model with no cache.
• Added GET /readyz for load balancers. It answers 503 while an instance shuts down.
• The telemetry stream is scoped to your organization. Send your API key to follow your own requests; without one, the stream opens but carries nothing.
• A chat completion without X-Mortar-Session or user starts a session of its own instead of sharing a default one, and every response names its session in X-Mortar-Session.
• Only the organization that started a generation can stop it.
• Chat completions report the prompt optimizer's work in X-Mortar-Optimizer, and X-Mortar-Optimize: off opts a request out.
• The prompt optimizer is off unless a deployment turns it on or sets it to shadow mode, and X-Mortar-Optimizer reports how long it took in elapsed_us.
• GET /health reports the build's version.
• The data plane now runs as a Go service. Endpoints, request fields and response formats are unchanged.
• Organizations have member roles. Creating, listing and revoking API keys needs the Admin or Owner role.

October 19, 2026

API

• Chat completions stream token by token as server-sent events. See Stream responses.
• Stopping a streamed reply keeps the text generated before the stop.