Prices are moving in both directions at once
OpenAI and Anthropic cutting, DeepSeek raising, Google discounting for four months with a doubling scheduled after. Anyone extrapolating a single direction for inference pricing in 2026 is choosing which evidence to ignore.
The comfortable narrative has been that inference costs fall monotonically, so build now and the economics improve underneath you. This week's evidence does not support a single direction, and planning on one is the mistake.
The four data points: Anthropic cancelled an announced Sonnet 5 increase, making introductory pricing permanent. Google shipped Flash at half price with a doubling scheduled for January. DeepSeek is raising. OpenAI is cutting while previewing a premium speed tier.
What actually explains the spread is position rather than cost. Labs fighting for a workload cut; labs that have won trust in one raise. DeepSeek is testing whether it is chosen for capability now that it is established, and Google is buying a coding position it does not yet hold. Neither move is about what inference costs to serve.
The practical consequence is that price should be treated as a competitive variable, not a technology curve — it can move against you. The defence is architectural: keep model choice a runtime decision, because the discount you budgeted against has an expiry date and the cheapest adequate model changes month to month.
VentureBeat — Google's Gemini 3.7 Flash targets coding and agents with a 50% introductory price cut → · Tech Startups — Top Tech News Today, August 14, 2026 → · Anthropic — Pricing — Claude Platform Docs →