Changelog

Notable changes to the Ablatic API and these docs. Newest first.

2026-07-16

reasoning_effort now shapes the reasoning. low keeps the reasoning brief; medium (the default) is normal depth; high, xhigh, and max raise the reasoning-token budget so the model can reason longer. There are two reasoning modes — fast and thinking — and effort above medium lifts the budget rather than switching to a separate mode. max_tokens still bounds the reasoning, so a small max_tokens at high effort keeps the reasoning short.
reasoning_tokens in the usage object is always reported as a subset of completion_tokens, matching the OpenAI contract.

One model. GET /v1/models now returns a single model, talos-preview. You pick the reasoning mode per request with reasoning_effort instead of switching model ids.
reasoning_effort accepts the full OpenAI value set: none/minimal map to fast, low/medium to thinking, high/xhigh/max to extended. Unknown values fall back to thinking rather than returning a 422.
Prompt caching is billed. Cached input tokens are charged at a large discount and cache writes at the normal input rate — no write premium, like OpenAI's automatic caching. See Usage and billing.
The documented context window is 200000 tokens, the real served ceiling.

The public API documentation went live at docs.ablatic.ai.

Two compatible surfaces are documented: OpenAI style at POST /v1/chat/completions and Anthropic style at POST /v1/messages.
The four models are listed: talos, talos-fast, talos-thinking, and talos-extended. Default is talos.
Base URL, authentication, models, capabilities, errors, and rate limits are covered.

This is the first changelog entry. We will add a dated entry here when the API surface or these docs change in a way you need to know about.