Skip to main content
MESH supports two ways to pick a model: router modes (let MESH choose) and explicit model IDs (pin a specific model). Pass any of the IDs below as the model field on /v1/chat/completions. Receipts identify which model was actually used.

Router modes

MESH house tier

MESH-curated models for everyday workloads. Pricing covers input / output per million tokens.

Anthropic

Meta Llama

Mistral

DeepSeek

OpenAI gpt-oss

OpenAI (chat + reasoning)

Google Gemma

Alibaba Qwen

AI21 Jamba

Moonshot Kimi

Z.AI GLM

Writer Palmyra

NVIDIA Nemotron

MiniMax

Perplexity Sonar (search-grounded)

Sonar models answer with real-time web search citations. Choose mesh/sonar for cheap factual lookups, mesh/sonar-pro for higher-quality grounded answers with tool calling, and mesh/sonar-reasoning when you want the model to reason over search results before answering.

xAI Grok

Picking a model

GET /v1/models returns the live machine-readable catalogue — IDs, context length, pricing, modalities, capabilities, supported parameters. GET /v1/models/{id} returns a single record. Treat the live API as the source of truth; this page is updated periodically.

Capability legend

  • streamingstream: true returns SSE deltas.
  • tools / tool calling — accepts tools[] + tool_choice; returns tool_calls with finish_reason: tool_calls.
  • structured output — accepts response_format: { type: "json_object" | "json_schema" }.
  • vision — accepts image_url parts in messages.
  • reasoning — model emits structured reasoning tokens.
  • web search — model answers from real-time web search and returns citations alongside the completion.
  • embeddings — addressed via /v1/embeddings.
Receipts include x-mesh-upstream-cost-usd, x-mesh-margin-usd, x-mesh-total-usd, x-mesh-receipt-id on every response.