Models

OpenAI's open-weight 120B reasoning model. Strong at coding, tool use and agentic tasks; the drop-in replacement Groq recommends for Llama 3.3 70B.

DeepSeek's fast, inexpensive V4 model with a 1M-token context. Strong general assistant and coding model at the lowest price per token here.

The newest DeepSeek Flash model: much quicker to answer than V4 Flash in our measurements, with a 1M-token context.

Z.ai's fast GLM 5.3 model with tool calling, JSON mode and a 1M-token context, at a low per-token price.

Meta's instruction-tuned 70B model. Still served here after its retirement on Groq — the same llama-3.3-70b-versatile id works unchanged.

Z.ai's full-size GLM 5.3 model: tool calling, JSON-schema output and a 1M-token context, for coding agents and long tasks.

DeepSeek's largest V4 model (0813 release) with a 1M-token context, hosted in the United States.

MiniMax's M3 model: tool calling and a 512K-token context at a price close to the Flash tier.

Moonshot AI's coding-tuned Kimi model with a 262K-token context, for coding agents and IDE assistants.

OpenAI's small open-weight reasoning model: the cheapest chat model here, for lightweight assistants, classification and extraction.

Alibaba's fast Qwen3.8 model with a one-million-token context window and cached-input pricing.

Alibaba's large vision-language model: send images in a chat request for OCR, document extraction and screenshot reasoning.

Fast, natural open-weight text-to-speech with 54 voices across 8 languages, on the OpenAI /v1/audio/speech endpoint.

Expressive Llama-based text-to-speech with emotive tags and eight English voices.

OpenAI's multilingual speech recognition model on the standard /v1/audio/transcriptions endpoint — same request shape as OpenAI and Groq.

NVIDIA's very fast English-first transcription model with accurate timestamps; ideal for long recordings and batch jobs.

Mistral's 24B speech model as a transcription endpoint: eight languages with automatic detection, strong on accented and noisy audio.

Mistral's small Voxtral model: the lowest price per minute here, eight languages with automatic detection.

BAAI's multilingual embedding model: 1,024-dimension vectors, 8K-token inputs, on the OpenAI /v1/embeddings endpoint.

Qwen's 8B embedding model: up to 4,096 dimensions with adjustable size, 32K-token inputs, multilingual.

Qwen's 8B reranker: scores passages against a query so RAG pipelines keep the right context.

Parse screenshots into structured, labeled UI elements for computer-use agents.