OpenLuxTool calling & agents
Agent frameworks live and die by function calling. These OpenLux models accept the standard OpenAI tools parameter on /v1/chat/completions, so existing agent loops work unchanged — pick by budget and reasoning depth.
Every model in this collection was tagged Tools in the catalogue, meaning it can be given a JSON schema for functions and will answer with structured tool_calls instead of prose. That is the primitive every agent stack — LangChain, LlamaIndex, OpenClaw, custom loops — is built on.
Because tool calling is model-dependent, the practical question is which tier to run. Fast flash-tier models keep dozens of parallel sub-agents affordable; stronger reasoning models are worth their rate for planning, code edits and long multi-step trajectories. Both live behind the same endpoint, so agent code can switch without changes.
The picks
8 models from the live catalogue, with the rate each one is metered at.
Gemini 3 Flash (preview)
Cheapest tool-calling brain for sub-agents and parallel workers — $0.0368 per 1M tokens (-92% vs official rates).
Gemini 3.8 Flash
High-throughput tool use for production pipelines — $0.0551 per 1M tokens (-92% vs official rates).
Qwen3 Max
Budget tool calling with dependable coding and Chinese output — $0.0784 per 1M tokens (-88% vs official rates).
GPT-5.6 Sol
Balanced reasoning-to-price ratio for general agent loops — $0.1838 per 1M tokens (-96% vs official rates).
Claude Sonnet 4.5
The default choice for coding and long agent runs — $0.2647 per 1M tokens (-91% vs official rates).
GPT-6 Astra
Flagship planning and reasoning for the hardest trajectories — $0.3677 per 1M tokens (-96% vs official rates).
Claude Opus 5
Anthropic's deepest reasoning tier for critical decisions inside a workflow — $0.4412 per 1M tokens (-91% vs official rates).
Claude Fable 5.1
Long-form research and synthesis with tool support — $1.765 per 1M tokens (-82% vs official rates).
Tool calling & agents questions
Which OpenLux models support function calling?
Every model tagged Tools in the catalogue: GPT-6 Astra, GPT-5.6 Sol, Gemini 3.8 Flash, Gemini 3 Flash preview, Qwen3 Max, Claude Sonnet 4.5, Claude Opus 5 and Claude Fable 5.1.
How does tool calling work through the gateway?
Exactly as it does on the official APIs: pass a tools array with JSON schemas on /v1/chat/completions and read tool_calls from the response. No gateway-specific syntax and no extra SDK to install.
Which model is best for long agent runs?
Claude Sonnet 4.5 is the usual default for coding agents; GPT-6 Astra handles general planning and reasoning; flash-tier Gemini keeps cost down when an orchestrator spawns many short-lived sub-agents.
Can I combine tool calling with streaming or vision?
Streaming works with tool calls on all of these models. Vision and tool calling can be combined on models tagged both Tools and Vision — a common pattern for agents that read screenshots and call APIs in the same loop.
Compare the picks head-to-head
GPT-6 Astra vs Gemini 3 Flash (preview)
Rates, strengths and which model to pick.
GPT-5.6 Sol vs Gemini 3 Flash (preview)
Rates, strengths and which model to pick.
Gemini 3.8 Flash vs Gemini 3 Flash (preview)
Rates, strengths and which model to pick.
Gemini 3 Flash (preview) vs Claude Sonnet 4.5
Rates, strengths and which model to pick.
Gemini 3 Flash (preview) vs Claude Opus 5
Rates, strengths and which model to pick.
Gemini 3 Flash (preview) vs Claude Fable 5.1
Rates, strengths and which model to pick.
Put these picks behind one key
OpenLux meters every model in this collection on the same OpenAI-compatible endpoint, with pay-as-you-go billing and free trial credit to start.