Providers

anyask integrates 18 LLM providers behind one interface — ask(), ask_async(), list_models(), list_models_async(), and get_provider() all take the same provider="..." value. Most providers just need an API key; three cloud-gateway providers (Bedrock, Vertex AI, OCI) use their own ambient credential chains instead and take extra constructor parameters in place of api_key.

There is no automatic routing or fallback anywhere in anyask — provider is always an explicit, required argument on every call.

Every provider defaults temperature=0.2 and max_tokens=2048 when the call doesn't specify them — the same defaults across all 18 providers, overridable per call via temperature=/max_tokens= kwargs. The one exception: when reasoning=True on OpenAI or an OpenAI-compatible vendor, both defaults are skipped entirely rather than sent, since reasoning models on some of these vendors reject a temperature override and use max_completion_tokens instead of max_tokens — see Reasoning support below.

All providers

Provider	provider value	Credential	Extra params	Default model
OpenAI	"openai"	OPENAI_API_KEY	—	gpt-4o-mini
Anthropic	"anthropic"	ANTHROPIC_API_KEY	—	claude-haiku-4-5-20251001
Google Gemini	"gemini"	GEMINI_API_KEY	—	gemini-2.5-flash
DeepSeek	"deepseek"	DEEPSEEK_API_KEY	—	deepseek-v4-flash
Groq	"groq"	GROQ_API_KEY	—	llama-3.1-8b-instant
xAI (Grok)	"xai"	XAI_API_KEY	—	grok-3-mini
Together AI	"together"	TOGETHER_API_KEY	—	meta-llama/Llama-3.1-8B-Instruct-Turbo
Fireworks AI	"fireworks"	FIREWORKS_API_KEY	—	accounts/fireworks/models/llama-v3p1-8b-instruct
Cerebras	"cerebras"	CEREBRAS_API_KEY	—	llama3.1-8b
Perplexity	"perplexity"	PERPLEXITY_API_KEY	—	sonar
OpenRouter	"openrouter"	OPENROUTER_API_KEY	—	google/gemini-2.5-flash-lite
Moonshot AI (Kimi)	"moonshot"	MOONSHOT_API_KEY	—	moonshot-v1-8k
Mistral AI	"mistral"	MISTRAL_API_KEY	—	mistral-small-latest
Azure OpenAI	"azure"	AZURE_API_KEY	endpoint (AZURE_ENDPOINT), api_version (AZURE_API_VERSION), deployment (AZURE_DEPLOYMENT) — model is passed as usual but the deployment name drives the actual call	(your deployment name)
Amazon Bedrock	"bedrock"	(none — AWS credential chain)	region (BEDROCK_REGION/BEDROCK_DEFAULT_REGION, required)	anthropic.claude-haiku-4-5-v1:0
Cohere	"cohere"	COHERE_API_KEY	—	command-light
Google Vertex AI	"vertexai"	(none — GCP ADC credential chain)	project (VERTEXAI_PROJECT, required), region (VERTEXAI_REGION, required)	gemini-2.5-flash
Oracle OCI	"oci"	(none — ~/.oci/config credential file)	compartment_id (OCI_COMPARTMENT_ID, required), region (OCI_REGION, optional — falls back to the config file)	meta.llama-3.3-70b-instruct

openai's extra also installs the SDK backing nine OpenAI-compatible vendors (deepseek, groq, xai, together, fireworks, cerebras, perplexity, openrouter, moonshot) — they share one HTTP client implementation (anyask/providers/openai_compat.py) and differ only in base URL, env var, and default model. Each still needs its own extra installed (e.g. pip install anyask[groq]) since credentials and default models are provider-specific, but no additional package beyond openai itself is pulled in.

Credentials resolve in this order everywhere: explicit function/keyword argument (api_key=..., region=..., ...) > provider-specific environment variable. There is no config-file fallback.

Two implementation shapes

Internally, providers fall into two groups — this only matters if you're extending anyask itself, not for calling it:

• Shape A — OpenAI-compatible chat-completions API (DeepSeek, Groq, xAI, Together AI, Fireworks AI, Cerebras, Perplexity, OpenRouter, Moonshot). These share one implementation (anyask/providers/openai_compat.py) and differ only in base URL, env var, and default model.
• Shape B — a native SDK with its own request/response shape (OpenAI, Anthropic, Gemini, Azure, Mistral, Bedrock, Cohere, Vertex AI, OCI). Bedrock, Vertex AI, and OCI additionally take extra constructor parameters (region/project/compartment_id) instead of api_key, since they authenticate via their cloud platform's own credential chain rather than a project-level API key.

Reasoning support

ask()/ask_async() take a reasoning: bool = False keyword. When True, each provider that supports it enables extended/deliberate reasoning using its own real mechanism — nothing is simulated, and support is not uniform:

Provider	reasoning=True behavior
Anthropic	Extended thinking (thinking={"type": "enabled", "budget_tokens": ...}); default budget 1024, override with thinking_budget=N. Temperature is left unset, since the API rejects an override while thinking is enabled.
OpenAI, and the 9 OpenAI-compatible vendors	reasoning_effort="medium" by default; override with reasoning_effort="low"/"high"/.... Support and accepted values depend on the specific model — an unsupported combination surfaces as a ProviderError from the vendor's own API.
Gemini, Vertex AI	thinking_config.thinking_budget=-1 (dynamic — model decides) by default; override with thinking_budget=N (or 0 to force off even when reasoning=True).
Amazon Bedrock	Same thinking field as native Anthropic, via additionalModelRequestFields — only applies to Claude models on Bedrock; other model families error from the API if it's not recognized.
Azure, Mistral, Cohere, OCI	Not supported — reasoning=True raises ProviderNotFoundError immediately rather than silently doing nothing.

reasoning=False (the default) leaves every provider's existing non-reasoning behavior completely unchanged.

A note on Vertex AI and OCI model IDs

Vertex AI serves the exact same Gemini model ID strings as the native gemini provider (e.g. gemini-2.5-pro means the same thing to both). OCI uses vendor.model-name IDs (meta.llama-3.3-70b-instruct) that share the meta. vendor prefix with Bedrock's own re-hosted model namespace. Since anyask has no auto-routing to begin with, this is never ambiguous in practice — you always select provider="vertexai" or provider="oci" explicitly — but it's worth knowing if you're cross-referencing model IDs between providers.