Supported Providers

All model metadata lives in lm_deluge.models. At import time, each provider module registers its models with the global registry, so you can inspect them programmatically:

from lm_deluge.models import registry

print(len(registry), "models available")
print(next(iter(registry.values())))

registry maps model IDs (the values you pass to LLMClient) to APIModel objects containing API base URLs, env vars, pricing, and capability flags (supports_json, supports_logprobs, supports_responses, reasoning_model). To see the external name sent to the provider, inspect APIModel.name.

Provider Modules

Module	Provider	Required Environment Variables
`lm_deluge.models.openai`	OpenAI (GPT-4.1, GPT-5, o-series, Codex, computer-use preview)	`OPENAI_API_KEY`
`lm_deluge.models.openrouter`	OpenRouter-hosted models	`OPENROUTER_API_KEY`
`lm_deluge.models.anthropic` / `lm_deluge.models.bedrock`	Anthropic (direct or via AWS Bedrock)	`ANTHROPIC_API_KEY` (direct) or AWS credentials for Bedrock
`lm_deluge.models.google`	Google Gemini	`GEMINI_API_KEY`
`lm_deluge.models.cohere`	Cohere Command + Embed models	`COHERE_API_KEY`
`lm_deluge.models.mistral`	Mistral models	`MISTRAL_API_KEY`
`lm_deluge.models.meta`	Meta Llama models (direct API)	`META_API_KEY`
`lm_deluge.models.deepseek`	DeepSeek	`DEEPSEEK_API_KEY`
`lm_deluge.models.groq`	Groq-hosted Llama and Mixtral	`GROQ_API_KEY`
`lm_deluge.models.grok`	xAI Grok models	`GROK_API_KEY`
`lm_deluge.models.fireworks`	Fireworks-hosted models	`FIREWORKS_API_KEY`
`lm_deluge.models.together`	Together.ai models	`TOGETHER_API_KEY`
`lm_deluge.models.cerebras`	Cerebras inference	`CEREBRAS_API_KEY`
`lm_deluge.models.kimi`	Moonshot/Kimi	`KIMI_API_KEY`
`lm_deluge.models.minimax`	MiniMax	`MINIMAX_API_KEY`

Some providers (Anthropic, Meta) can also be accessed through AWS Bedrock. In that case the registry entry points at the Bedrock endpoint and lists api_key_env_var="AWS_ACCESS_KEY_ID" with implicit use of AWS_SECRET_ACCESS_KEY.

Discovering Models at Runtime

from lm_deluge.models import registry

anthropic = [cfg for cfg in registry.values() if cfg.api_spec == "anthropic"]
reasoning = [cfg for cfg in registry.values() if cfg.reasoning_model]

for cfg in reasoning:
    print(cfg.id, cfg.name, cfg.supports_responses)

Use this approach to build CLI selectors, validate configuration files, or dynamically choose models that support JSON mode, logprobs, background tasks, etc.

Cost Metadata

Each APIModel includes input_cost, cached_input_cost, cache_write_cost, and output_cost (all per million tokens). APIResponse.cost is calculated automatically when the provider returns token usage data.

For the latest provider-specific pricing and rate limits, consult the upstream provider documentation; LM Deluge stores the values that were current when the model definitions were last updated.