[{"data":1,"prerenderedAt":101},["ShallowReactive",2],{"profile":3,"writing-the-year-inference-got-cheap":14},{"name":4,"role":5,"tagline":6,"location":7,"contact":8},"Sankar Vema","AI Builder & Architect of Agentic Systems","I help enterprise leaders turn AI ambition into capability that actually ships.","India · open to global advisory engagements",{"email":9,"linkedin":10,"github":11,"blog":12,"twitter":13},"sankar.vema@gmail.com","https:\u002F\u002Fwww.linkedin.com\u002Fin\u002Fsankarvema\u002F","http:\u002F\u002Fsankarvema.github.io\u002F","http:\u002F\u002Fsankarvema.blogspot.com\u002F","https:\u002F\u002Ftwitter.com\u002Fsansvema",{"id":15,"title":16,"body":17,"date":86,"description":23,"draft":87,"excerpt":88,"extension":89,"meta":90,"navigation":91,"path":92,"seo":93,"slug":94,"status":88,"stem":95,"tags":96,"__hash__":100},"writing\u002Fwriting\u002Fthe-year-inference-got-cheap.md","The Year Inference Got Cheap",{"type":18,"value":19,"toc":78},"minimark",[20,24,29,32,35,39,42,45,49,72,75],[21,22,23],"p",{},"Quietly, the most important number in AI has been falling off a cliff. The cost\nof a unit of machine reasoning, a token, has dropped by more than an order of\nmagnitude in a short span, and it is still falling. Reasoning models that think\nbefore they answer arrived at the same time. Thinking, in other words, got both\nbetter and cheaper at once. Most enterprise AI strategies were written when it\nwas still expensive, and they have quietly gone out of date.",[25,26,28],"h2",{"id":27},"why-this-changes-the-plan","Why this changes the plan",[21,30,31],{},"When intelligence is expensive, you ration it. You use AI only on the few\nhigh-value calls, you keep prompts short, you avoid letting a model loop. Every\none of those instincts becomes wrong when the price collapses.",[21,33,34],{},"Cheap inference means you can afford to let a system think longer, check its own\nwork, try several approaches, and call a model many times inside a single task.\nThe behaviours that were too expensive last year, verification, self-correction,\nmulti-step reasoning, are now just line items. The question shifts from \"can we\nafford to use AI here\" to \"why are we not using it everywhere it removes toil.\"",[25,36,38],{"id":37},"the-trap-on-the-other-side","The trap on the other side",[21,40,41],{},"Cheap does not mean free, and cheap per call does not mean cheap at scale. When a\nsingle user action quietly triggers fifty model calls, your bill can grow faster\nthan your value. Falling prices reward volume, and volume without discipline is\nhow AI programs blow their budget while looking productive.",[21,43,44],{},"So the discipline is not rationing anymore. It is instrumentation. Know your cost\nper outcome, not your cost per token, and watch it like a unit economic.",[25,46,48],{"id":47},"what-to-do-now","What to do now",[50,51,52,60,66],"ol",{},[53,54,55,59],"li",{},[56,57,58],"strong",{},"Revisit the use cases you killed for being too expensive."," Many are now\nviable. Your no-list from last year is your opportunity list this year.",[53,61,62,65],{},[56,63,64],{},"Design for thinking, not just answering."," Let systems verify and retry.\nThe extra calls are cheap, the wrong answer is not.",[53,67,68,71],{},[56,69,70],{},"Track cost per outcome."," Falling unit prices hide rising totals. Measure\nthe thing the business actually pays for.",[21,73,74],{},"The teams that win are not the ones who spent the least on tokens. They are the\nones who noticed the price moved, and rebuilt the plan around it.",[21,76,77],{},"So before your next AI roadmap: which of your assumptions were priced at last\nyear's cost of thinking?",{"title":79,"searchDepth":80,"depth":80,"links":81},"",3,[82,84,85],{"id":27,"depth":83,"text":28},2,{"id":37,"depth":83,"text":38},{"id":47,"depth":83,"text":48},"2026-08-01",false,null,"md",{},true,"\u002Fwriting\u002Fthe-year-inference-got-cheap",{"title":16,"description":23},"the-year-inference-got-cheap","writing\u002Fthe-year-inference-got-cheap",[97,98,99],"ai","economics","strategy","eV4zupH4Q7_O-gzI6bXgUrK3WHJNV1CesdSMRatoQ9c",1790601541762]