Frontier model sizes: Estimates through the price data lens (2024–2025)
LifeArchitect.ai
* Pricing via poe.com, inference credits per standard msg (Aug/2025). Parameter count estimates are total, rounded, based on public model sizes (green) and fitted to pricing using the simplified formula: Params (B) ≈ Credits × 1.2. Grounded on GPT-4 Classic, DeepSeek-V3, Llama 3.1 405B, GPT-4o. Model margins, architecture, reasoning, and inference optimizations introduce significant variability, widening the range of possible parameter count estimates.
† GPU inference estimate assumes total weights (params) loaded in FP4 on NVIDIA H100 94GB NVL GPUs (matching the hardware setup reported for ChatGPT’s inference fleet) with ~70GB of usable HBM each (≈140B params/GPU).
† GPU inference estimate assumes total weights (params) loaded in FP4 on NVIDIA H100 94GB NVL GPUs (matching the hardware setup reported for ChatGPT’s inference fleet) with ~70GB of usable HBM each (≈140B params/GPU).
Chart: Alan D. Thompson • Aug/2025
