We’ve been tracking inference pricing for 44 weeks now across 2,583 SKUs and 47 vendors. DeepSeek was the biggest single shock we recorded but the structural patterns that followed are more interesting than the initial drop.
Output tokens cost 3.84x more than input on average for text models. Most cost discussions anchor on input pricing because that’s the number on vendor landing pages. For agentic and generation-heavy workloads the real bill is much higher than people expect.
Open source runs 80% cheaper than proprietary equivalents on the same platforms. DeepSeek accelerated this gap but didn’t create it. It was already widening.
Caching saves 69.7% where it exists but only 20.3% of tracked models offer it. The savings are real but access is structurally uneven.
Model size spread is 4.8x between large and small tiers. Most teams don’t know this ratio before they commit to a model.
Price volatility was 0.27% on input this week. Quieter than usual but the structural gaps above have barely moved in months. We publish the full AIPI index every Monday at a7om.com. Happy to discuss methodology or share more granular data.