Data Methodology
How we collect, verify, and display AI API and GPU pricing data
1. Data Types
2. Update Frequency
Automated scrapers run daily at 09:00 UTC, covering all live_api platforms. A full run takes approximately 10 minutes. Due to GitHub Actions scheduling variance, the actual start time may be delayed by 1–9 hours, but typically completes between 09:00–18:00 UTC.
Manually maintained data has no fixed schedule — it follows official announcements. We monitor official pricing pages and announcement channels for major models and typically update within 24 hours of an official announcement.
3. Price Normalization
All LLM prices are normalized to per-1M-token costs, split into input and output. Original units (e.g., per-1K tokens, per-character) are converted proportionally.
Multimodal pricing (image, video, audio) retains original billing units (per image, per second, per character) and is not merged with LLM token pricing, to avoid misleading comparisons.
4. Platform Coverage
Current coverage includes 70+ platforms: official AI API providers (OpenAI, Anthropic, Google, DeepSeek, etc.), global API aggregators (OpenRouter, Groq, Together AI, etc.), Chinese official model APIs (Zhipu, Moonshot, MiniMax, StepFun, etc.), CN relay stations (302.AI, SiliconFlow, YunWu AI, AiHubMix, etc.), and GPU cloud platforms (Vast.ai, RunPod, Lambda Labs, etc.).
Some platforms cannot be automatically scraped due to login requirements, anti-scraping measures, or non-public pricing. Where data is available, these are included as manually maintained entries.
5. Known Limitations
- Prices reflect the last scrape time, not real-time market rates. For large purchasing decisions, always verify on the provider's official page.
- Some providers offer volume discounts not reflected here. We show standard public pricing only.
- GPU spot prices fluctuate frequently; we show daily snapshots, not instantaneous availability.
- The USD/CNY rate is fixed at 7.25 and does not track live exchange rates.
- Model versioning (e.g., gpt-4o-2025-04-01 vs. gpt-4o) may cause apparent price inconsistencies. We display the most recently published pricing.