LLM Providers

Overview

zoxaAI provides managed access to 4 LLM providers, each with a curated catalog of production-tested models. Each agent is configured with exactly one LLM provider and model.

Model selection is a curated catalog, not free text. You pick from the models listed on each provider below; arbitrary or newly-released model ids that aren't in the catalog are rejected at save time with a 422 pointing at GET /api/v1/user/configurations/providers. A provider only appears in the picker when it has at least one active model in the catalog.

Every model in the catalog carries metadata — latency (time-to-first-token / time-to-first-byte), throughput, an intelligence index, a tool-calling score, word error rate, context window, and per-unit cost — stored alongside its pricing. The catalog endpoint returns this metadata so you can compare models on speed, cost, and quality before choosing.

Every OpenAI model except gpt-5.4-nano and gpt-6-luna is dispatched with service_tier: priority (OpenAI's Fast tier) for lower and more consistent voice TTFT. OpenAI charges more for Priority calls; the cost calculator already reflects the Priority rates so the dashboard shows what OpenAI actually invoices. gpt-5.4-nano and gpt-6-luna bill — and run — at Standard.

The factory routes parameters per OpenAI's /v1/chat/completions constraints — picking the wrong shape returns 400:

• gpt-6 family (gpt-6-luna) — function tools work on chat-completions only with reasoning_effort: none, which the factory always sends (it is also the right setting for voice). Custom temperature + max_completion_tokens.

• gpt-5.x point releases (gpt-5.1 / 5.2 / 5.4 / 5.4-mini / 5.4-nano) — OpenAI moved the tools-plus-reasoning combo to /v1/responses. On chat-completions these models reject reasoning_effort whenever tools are present (which our pipeline always sends). They accept a custom temperature; the factory passes max_completion_tokens + temperature.

• gpt-4.1 family (gpt-4.1, gpt-4.1-mini) — classic shape: custom temperature + max_tokens.

Gemini 2.5 models run with thinking_budget: 0 (thinking off), and Gemini 3.x models run with thinking_level: minimal, so replies start without a reasoning phase.

enable_thinking is forced to false for every Qwen call. Every Qwen model in the catalog is a hybrid thinking model: with thinking on, it streams its reasoning before the actual reply, adding 1-2 seconds to TTFT and breaking streaming TTS — voice is hard-incompatible with that mode. With enable_thinking: false every catalog model answers directly. The flag is passed through OpenAI's extra_body channel; it is NOT a direct chat-completions kwarg.

Temperature range is 0.0 -- 1.0. Anthropic's Messages API rejects temperatures above 1.0. When you select an Anthropic model in the agent editor, the Temperature slider caps at 1.0 and any stored value above 1.0 is clamped. If you call the API directly, ensure temperature ≤ 1.0 for any Anthropic provider config.