Models

Lantern uses capability-based routing to map abstract capability names to concrete LLM models. This means your agent code never hardcodes a vendor model name -- it specifies what kind of intelligence it needs, and the model router picks the best option at runtime.

Supported providers

Lantern supports the following LLM providers out of the box:

• Anthropic -- Claude Opus, Sonnet, Haiku
• OpenAI -- GPT-5, GPT-4o, GPT-4o-mini
• Google -- Gemini Ultra, Gemini Pro, Gemini Flash

Capability routing

Instead of specifying "gpt-4" or "claude-3-opus", you specify a capability:

| Capability          | Description                              | Example models                    |
|---------------------|------------------------------------------|-----------------------------------|
| "auto"              | Best model for each step (cost+quality)  | Varies per step                   |
| "reasoning-large"   | Maximum reasoning capability             | Opus, GPT-5                       |
| "reasoning-small"   | Fast, cheap reasoning                    | Haiku, GPT-4o-mini, Gemini Flash  |
| "code"              | Optimized for code generation            | Sonnet, GPT-4o                    |
| "vision"            | Image understanding                      | Opus, GPT-4o, Gemini Ultra        |
| "embedding"         | Text embeddings                          | text-embedding-3, embed-v4        |

How "auto" works

When you set model: "auto", the smart router scores every available model and picks the best one. The scoring considers:

• Quality (40%) — how capable the model is for complex reasoning
• Speed (30%) — response latency
• Cost efficiency (30%) — inverse of token pricing

Model scoring table

Routing strategies

You can control the routing strategy via environment variable:

# Default — best balance of quality, speed, and cost
LANTERN_ROUTE_STRATEGY=balanced make run-api

# Cheapest available model (Haiku or GPT-4o-mini)
LANTERN_ROUTE_STRATEGY=cheap make run-api

# Highest quality regardless of cost (Opus 4 or GPT-4o)
LANTERN_ROUTE_STRATEGY=quality make run-api

# Fastest response time (Haiku or GPT-4o-mini)
LANTERN_ROUTE_STRATEGY=fast make run-api

Adding API keys

To use a provider, add your API key in the dashboard:

1. Navigate to Settings > Models
2. Click the provider you want to configure
3. Enter your API key
4. Click Save -- the key is encrypted at rest

[Screenshot: Model provider settings with API key input]

Multiple keys per provider

You can add multiple API keys for the same provider. The router will distribute requests across keys to avoid rate limits. You can also set a primary and fallback key.

Failover behavior

The model router handles provider failures automatically:

• Latency spike -- if a provider's response time exceeds the P95 threshold, subsequent requests route to an alternative
• 5xx errors -- immediate retry on an alternative provider
• Rate limits (429) -- exponential backoff with automatic rotation to another key or provider
• Timeout -- configurable per-step timeout with failover

Cost tracking

Every LLM call is metered and tracked. You can view per-agent and per-run cost breakdowns in the dashboard under Usage > Model costs. The router also provides recommendations for switching capabilities to save costs.

Overriding the router

If you need a specific model for a step (e.g., for compliance reasons), you can override the router in the SDK:

const result = await step("analyze", async () => {
  return ctx.llm.complete({
    messages: [...],
    capability: "reasoning-large",
    provider: "anthropic",  // Force Anthropic
    model: "claude-opus-4", // Force a specific model
  });
});