Update Model Pricing

How Pricing Works (Two-Tier System)

Pricing is resolved by CostMixin._calculate_cost() in this order:

1. •models.dev live rates (primary) — cached from https://models.dev/api.json
2. •Hardcoded MODEL_PRICING (fallback) — per-driver class variable
3. •Zero — if neither source has data

Most newer drivers (moonshot, zai, modelscope) have MODEL_PRICING = {} because their pricing comes entirely from models.dev. Older drivers (openai, claude, google, groq, grok, openrouter, azure) have hardcoded pricing as a fallback.

models.dev Integration

The mapping from prompture provider names to models.dev provider names lives in prompture/model_rates.py:

PROVIDER_MAP = {
    "openai": "openai",
    "claude": "anthropic",
    "google": "google",
    "groq": "groq",
    "grok": "xai",
    "azure": "azure",
    "openrouter": "openrouter",
    "moonshot": "moonshotai",
    "zai": "zai",
}

The cache is stored at ~/.prompture/cache/models_dev.json with a TTL configured by settings.model_rates_ttl_days (default 7 days).

When to Update What

Refreshing models.dev Cache

from prompture.model_rates import refresh_rates_cache
refresh_rates_cache(force=True)  # Fetch fresh data regardless of TTL

Or delete the cache file:

rm ~/.prompture/cache/models_dev.json

Updating Hardcoded MODEL_PRICING (Fallback)

Only needed for drivers where models.dev doesn't have pricing data, or as a safety fallback.

Files with hardcoded pricing:

Drivers with NO hardcoded pricing (models.dev only):

Free/local drivers (always $0):

ollama_driver.py, lmstudio_driver.py, local_http_driver.py, airllm_driver.py, hugging_driver.py

Steps for Hardcoded Updates

1. •Search the web for the provider's current pricing page
2. •Read the current MODEL_PRICING dict in the driver file
3. •Update prices, add new models, remove discontinued ones
4. •Preserve extra keys like "tokens_param" or "supports_temperature" — these control API behavior, not just pricing
5. •Check the unit: most drivers use per-1K tokens, but check _PRICING_UNIT on the class
6. •Run tests: pytest tests/ -x -q

Format

MODEL_PRICING = {
    "model-name": {
        "prompt": 0.005,        # cost per unit (see _PRICING_UNIT)
        "completion": 0.015,
        "tokens_param": "max_completion_tokens",  # API parameter name (optional)
        "supports_temperature": True,              # override for this model (optional)
    },
}

Both "prompt" and "completion" keys are required. Extra keys are optional and model-specific.

Side Effects

• •prompture/discovery.py reads MODEL_PRICING keys to list available models (static detection)
• •PROVIDER_MAP entries also feed discovery via get_all_provider_models() (models.dev detection)
• •Adding a model to MODEL_PRICING makes it appear in get_available_models() even without an API key configured
• •CostMixin._get_model_config() reads tokens_param and supports_temperature from MODEL_PRICING as fallback when models.dev data is unavailable

Verification

# Check live rates for a model
python -c "from prompture.model_rates import get_model_rates; print(get_model_rates('openai', 'gpt-4o'))"

# Check capabilities for a model
python -c "from prompture.model_rates import get_model_capabilities; print(get_model_capabilities('moonshot', 'kimi-k2.5'))"

# Check cache age
python -c "from prompture.model_rates import _META_FILE; import json; print(json.load(open(_META_FILE)))"

# Run tests
pytest tests/ -x -q