GIANT Model Registry¶
Single Source of Truth for LLM model identifiers used in GIANT
This diverges from the original paper to use current frontier models (Dec 2025).
Approved Models¶
These are the only model IDs allowed by src/giant/llm/model_registry.py.
| Provider | API Model ID | Runtime Support | Notes |
|---|---|---|---|
| OpenAI | gpt-5.2 |
Supported | Default OpenAI model |
| Anthropic | claude-sonnet-4-5-20250929 |
Supported | Default Anthropic model |
gemini-3-pro-preview |
Planned | Model ID is reserved; Google/Gemini provider is not implemented yet |
Pricing (USD per 1M tokens)¶
| Model | Input | Output | Image Cost |
|---|---|---|---|
claude-sonnet-4-5-20250929 |
$3.00 | $15.00 | Pixel-based (~$0.00048/1K px) |
gemini-3-pro-preview |
$2.00 | $12.00 | Included in token count |
gpt-5.2 |
$1.75 | $14.00 | Flat-rate per image ($0.00255/image) |
Notes:
- The Anthropic image cost in code is $0.00048 / 1K pixels (i.e., $0.48 / 1M pixels).
- Pricing and image-cost formulas are implemented in src/giant/llm/pricing.py.
Why These IDs?¶
GIANT pins model IDs for reproducibility and to ensure pricing is known. If you need to update pricing or supported models, update docs/models/model-registry.md and src/giant/llm/pricing.py, and keep src/giant/llm/model_registry.py aligned.
Code Usage¶
from giant.llm import create_provider
from giant.llm.model_registry import validate_model_id
# OpenAI
provider = create_provider("openai", model="gpt-5.2")
# Anthropic
provider = create_provider("anthropic", model="claude-sonnet-4-5-20250929")
# Google/Gemini (reserved model ID; provider not implemented)
validate_model_id("gemini-3-pro-preview", provider="google") # OK