Providers

Smelt works with any LLM provider supported by LangChain's init_chat_model. This page covers setup, recommended models, and provider-specific tips.

Overview

Provider	provider value	Install extra	Env var
OpenAI	"openai"	pip install smelt-ai[openai]	OPENAI_API_KEY
Anthropic	"anthropic"	pip install smelt-ai[anthropic]	ANTHROPIC_API_KEY
Google Gemini	"google_genai"	pip install smelt-ai[google]	GOOGLE_API_KEY

OpenAI

Recommended models

Model	Best for	Notes
gpt-4.1-mini	Most tasks — fast, cheap, good quality	Best price/performance ratio
gpt-4.1	Complex schemas, nuanced classification	Higher quality, higher cost
gpt-4o	General purpose	Solid all-rounder
gpt-5.2	Maximum quality	Latest and most capable
o4-mini	Reasoning-heavy tasks	Good for complex logic

Setup

from smelt import Model

# Using environment variable (recommended)
# export OPENAI_API_KEY="sk-..."
model = Model(provider="openai", name="gpt-4.1-mini")

# Explicit API key
model = Model(provider="openai", name="gpt-4.1-mini", api_key="sk-...")

# With parameters
model = Model(
    provider="openai",
    name="gpt-4.1-mini",
    params={"temperature": 0, "max_tokens": 4096},
)

Tips

• Use gpt-4.1-mini as your default — it handles structured output very well and is cost-effective
• Set temperature=0 for classification and extraction tasks
• Use temperature=0.5–0.7 for creative tasks (summaries, marketing copy)
• Watch for rate limits — at high concurrency, you may hit tokens-per-minute limits. Smelt retries 429 errors automatically

Anthropic

Recommended models

Model	Best for	Notes
claude-sonnet-4-6	Most tasks — fast, great quality	Best balance for structured output
claude-opus-4-6	Complex reasoning, nuanced output	Highest quality, slower
claude-haiku-4-5-20251001	High volume, simple schemas	Fastest, most cost-effective

Setup

from smelt import Model

# Using environment variable (recommended)
# export ANTHROPIC_API_KEY="sk-ant-..."
model = Model(provider="anthropic", name="claude-sonnet-4-6")

# Explicit API key
model = Model(provider="anthropic", name="claude-sonnet-4-6", api_key="sk-ant-...")

# With parameters
model = Model(
    provider="anthropic",
    name="claude-sonnet-4-6",
    params={"temperature": 0, "max_tokens": 4096},
)

Tips

• Claude excels at following complex instructions — great for detailed prompts with many rules
• Anthropic models require max_tokens in some configurations — if you get truncated output, add params={"max_tokens": 4096}
• Claude tends to be conservative — if you want more creative output, increase temperature

Google Gemini

Recommended models

Model	Best for	Notes
gemini-3-flash-preview	Most tasks — very fast, good quality	Best speed/quality ratio
gemini-3-pro-preview	Complex tasks, highest quality	Slower but more capable
gemini-2.5-flash	Budget-friendly, simple tasks	Previous generation, still good

Setup

from smelt import Model

# Using environment variable (recommended)
# export GOOGLE_API_KEY="..."
model = Model(provider="google_genai", name="gemini-3-flash-preview")

# Explicit API key
model = Model(provider="google_genai", name="gemini-3-flash-preview", api_key="...")

# With parameters
model = Model(
    provider="google_genai",
    name="gemini-3-flash-preview",
    params={"temperature": 0},
)

Tips

• Gemini Flash is excellent for high-volume tasks — fast response times, generous rate limits
• Provider name is google_genai (not google or gemini)
• Gemini handles large batches well — you can often use batch_size=20-30 without issues

Other providers

Smelt works with any LangChain-supported chat model. Install the provider's LangChain package and use the appropriate provider value.

Azure OpenAI

pip install smelt-ai langchain-openai

model = Model(
    provider="azure_openai",
    name="my-gpt4-deployment",
    params={
        "azure_endpoint": "https://my-resource.openai.azure.com/",
        "api_version": "2024-02-15-preview",
        "api_key": "...",
    },
)

AWS Bedrock

pip install smelt-ai langchain-aws

model = Model(
    provider="bedrock",
    name="anthropic.claude-3-5-sonnet-20241022-v2:0",
    params={
        "region_name": "us-east-1",
    },
)

Ollama (local models)

pip install smelt-ai langchain-ollama

model = Model(
    provider="ollama",
    name="llama3.1:8b",
    params={"base_url": "http://localhost:11434"},
)

Local model quality

Smaller local models may struggle with structured output, especially complex schemas or large batch sizes. Use batch_size=1-3 and simpler schemas for best results.

Choosing a provider

By use case

Use case	Recommended	Why
General classification	gpt-4.1-mini or gemini-3-flash-preview	Fast, cheap, good quality
Complex extraction	claude-sonnet-4-6 or gpt-4.1	Better at following complex instructions
High volume (1000+ rows)	gemini-3-flash-preview	Fastest, generous rate limits
Maximum quality	gpt-5.2 or claude-opus-4-6	Best output quality
Budget-conscious	gpt-4.1-mini or gemini-3-flash-preview	Lowest cost per token
Privacy-sensitive	Ollama (local)	Data stays on your machine

Cost comparison

Approximate cost to process 1000 rows with batch_size=10 (100 LLM calls):

Model	Input cost	Output cost	Approx total
gpt-4.1-mini	Lowest	Lowest	~$0.10–0.50
gemini-3-flash-preview	Low	Low	~$0.10–0.50
claude-sonnet-4-6	Medium	Medium	~$0.50–2.00
gpt-4.1	Medium-high	Medium-high	~$1.00–5.00
claude-opus-4-6	Highest	Highest	~$5.00–20.00

Note

Actual costs depend on your schema complexity (affects output tokens), input data size (affects input tokens), and retry rate. Check result.metrics.input_tokens and result.metrics.output_tokens for exact usage.