What's New
🎯 Accurate Model Limits Detection
Fixed model context/output limits for all providers, especially for proxied models via Antigravity backend.
Key Fix: Now uses model ID pattern matching instead of owned_by field, so gemini-claude-* models correctly get Claude's limits instead of generic defaults.
Updated Limits (from models.dev):
| Provider | Model | Context | Output |
|---|---|---|---|
| Claude 4.5 | sonnet/opus/haiku | 200K | 64K |
| Claude 3.5 | haiku | 200K | 8K |
| Gemini 2.5 | pro/flash | 1M | 65K |
| GPT-4o | all variants | 128K | 16K |
| OpenAI o1/o3 | reasoning | 200K | 100K |
| Qwen3 | max | 262K | 65K |
| Qwen3 Coder | plus | 1M | 65K |
| DeepSeek | chat | 128K | 8K |
| DeepSeek | reasoner | 128K | 128K |
🧠 Dynamic Thinking Budget
For -thinking models, budgetTokens is now dynamically calculated as half of the output limit (e.g., 32K for Claude 4.5's 64K output), providing more room for extended reasoning.