v2.2.0 — gateway-addressed models keep their effort and their price

Model capability detection and pricing lookup now strip a gateway vendor prefix before matching.

A model addressed as openai/gpt-5.6-luna or z-ai/glm-5.2 previously failed every anchored rule, so it was classified as a standard model, reasoningEffort was silently deleted from the request, and the pricing lookup returned nothing. Setting an effort level had no effect and produced no warning.

Fixed

• One shared vendor-prefix strip now normalizes the model id for both capability detection and pricing, in TypeScript and Python alike. Pricing previously enumerated three vendor prefixes and lost every other vendor.

Changed

• Capability rules are read from model-capabilities.json instead of the hardcoded regexes that duplicated it. The file already shipped in both packages and described itself as the source of parity; nothing read it. Python now carries a byte-identical copy.
• fixedTemperature is no longer an alias for isReasoningModel. The temperature restriction belongs to the OpenAI families and has its own list, so a reasoning model from another vendor keeps its temperature.
• GLM is recognised as a reasoning model.

Added

• Pricing for gpt-5.6-luna, glm-5.2 and deepseek-v4-flash-0731; refreshed deepseek-v4-flash.

TypeScript and Python return byte-identical capability output for the same inputs.

LLMix 2.0.0

LLMix 2.0.0 is the production-ready rewrite of the LLM orchestration layer for AI agents, AI tools, and config-driven model workflows.

Highlights

• Config-driven model swaps with MDA presets, so provider, model, and runtime parameters can move out of application code.
• Keep the SDK you already use: OpenAI, Anthropic, Gemini, AI SDK v6, LiteLLM, OpenRouter, DeepInfra, Novita, Together, Sno GPU, or any async callable.
• Full call pipeline with cache lookup, circuit breaker, singleflight deduplication, adaptive concurrency, retries, key-pool rotation, thinking-token stripping, and telemetry.
• Two-tier response cache with L1 memory and optional Redis L2.
• Cross-runtime parity for Python, TypeScript, and Rust, including shared cache-key behavior and aligned retry semantics.
• Config Registry support for immutable, content-addressed preset snapshots and atomic runtime switches.

Runtime Support

• Python 3.14+
• TypeScript / Node 20+
• Rust 1.83+ via llmix-rs beta

Packages

Package publishing is handled separately from GitHub Releases:

• npm: @snoai/llmix
• PyPI: llmix
• crates.io: llmix-rs