Test-time compute has flipped AI economics: the smartest answers now cost the most, every time

The industry-wide shift toward 'test-time compute' or 'reasoning' models - which spend substantially more computation at the moment of answering a question, working through extended chains of intermediate reasoning steps before producing a final response, rather than relying purely on a single fast forward-pass through the model - has fundamentally changed AI inference economics in a way that has forced every major lab and enterprise customer to rethink cost-management strategies that were built around the older, cheaper, single-pass inference paradigm. The practical consequence has been the proliferation of tiered model offerings, where every major lab now offers a spectrum from fast, cheap, lower-reasoning-effort models suitable for simple queries up through slow, expensive, maximum-reasoning-effort models reserved for the hardest problems, with sophisticated routing systems - as seen in OpenAI's GPT-5 architecture - increasingly making the tier-selection decision automatically rather than leaving it to the end user, a design choice that optimises cost and quality simultaneously but reduces user visibility into and control over exactly how much compute, and therefore cost, any given query actually consumes. Enterprise customers managing AI costs at scale have had to build entirely new cost-forecasting and budgeting practices around this variable-cost reality, since a workload's actual inference cost can now vary substantially based on query complexity in ways that are much harder to predict in advance than the more predictable per-token pricing that characterised the pre-reasoning-model era, and several enterprise AI-cost-management startups have emerged specifically to help large customers monitor and control this newly variable cost structure. For Indian AI application developers operating with tighter margins than well-funded Western counterparts, the test-time-compute cost increase for the hardest reasoning tasks has been a genuine commercial challenge, pushing many toward more careful workload-routing strategies that reserve expensive reasoning-tier model calls only for queries that genuinely require them, while handling the bulk of simpler interactions with cheaper, faster models specifically to manage unit economics at India-appropriate price points. What to watch: whether reasoning-model costs decline meaningfully as efficiency research continues, how enterprise cost-management practices mature to handle the new variable-cost reality, and whether Indian AI application companies develop distinctive workload-routing expertise that becomes a competitive advantage in cost-sensitive market segments.

Original source: The Information