Industry story
Reasoning models use 30x more energy than standard models, study finds
cloud-costs inference measurement model-pricing
A consortium of researchers including representatives from Cohere and Meta developed the 'AI Energy Score,' a star-rating system measuring model energy efficiency during inference (the process of running a model to answer queries, as opposed to training it). Their key finding: reasoning models — which work through problems step-by-step to mimic deliberate human thought, like OpenAI's o-series or Anthropic's extended thinking — consume on average 30 times more energy than standard models or models with reasoning disabled. This finding is significant because reasoning models are increasingly the default for high-quality AI outputs, meaning the industry's efficiency gains at the chip level are being more than offset by architectural shifts toward more compute-intensive inference patterns.
The research also highlights that inference, not training, now dominates AI's energy footprint. A 2022 Google paper found inference accounted for ~60% of its ML energy use; the UN University now estimates inference at 80–90% of total AI energy consumption. As AI chatbot usage has exploded since ChatGPT's 2022 launch, that share has likely grown further, meaning even a halt to new model training would leave a large and growing energy burden.
Full analysis
A group of researchers, including people from Cohere and Meta, built a star-rating system for how much energy a model burns when it answers you, not when it's trained. Their headline: reasoning models, the ones that think step by step before answering (OpenAI's o-series, Anthropic's extended thinking), eat about 30 times more energy per query than the same models with reasoning switched off. The bigger claim underneath is that answering queries, not training, is now 80 to 90% of AI's energy bill.
How hard is this to undo? Nothing to undo. This is a measurement, not a decision. What it changes is how you should price and route the reasoning traffic you already run.
What's actually being decided: Whether "just turn on reasoning for everything" is a cost mistake you're making right now, and whether the vendors setting your prices have any reason to fix it for you.
What sets the deadline: Nothing external. But if you turned reasoning on by default sometime in the last year, your 90-day cloud bill is the deadline that finds you.
The Skeptic. The 30x is real. The crisis is fan fiction. Nobody routes their whole load to o3. Reasoning gets invoked for the hard 5% where a standard model faceplants. The comparison the study runs, reasoning versus reasoning-off on the same prompt, is the wrong denominator. If o3 nails in one shot what took ten failed GPT-4o tries plus a human cleaning up, the energy math flips the other way. And the "inference is 80 to 90%" line leans on a UN University estimate stitched from thin disclosure and a 2022 Google number from before reasoning models existed. Direction, fine. Magnitude, hold it at arm's length.
The Builder. This one bites on Tuesday. If your cost model was calibrated on GPT-4o or Sonnet standard and you've since flipped traffic to o3 or Claude extended thinking, the model is wrong. OpenAI and Anthropic already price reasoning tokens punitively, so the bill was going to find you regardless. The study just names the physics. The trap is autoscaling: reasoning jobs have long, unpredictable token counts that wreck your reserved-capacity math. Run an eval loop on o3 with no token cap and you get bill shock at 90 days that no day-one demo shows you. Cap the reasoning budget per call. Route by difficulty, not by default.
The Compute Pragmatist. The inference story outran the hardware roadmap. NVIDIA's newest gear is tuned for training throughput. Reasoning inference is a different beast: it's bottlenecked on memory bandwidth and generates tokens one after another, which doesn't keep those chips busy. So 30x energy also means 30x more heat, power, and rack pressure in data centers that were laid out for dense training clusters. This is the clearest public evidence yet that the sequential-token bottleneck is real, and it hands Groq and Cerebras a genuine opening, since their chips chew through that specific bottleneck better. The counterweight: hyperscalers sank fortunes into H100s and H200s and have every reason to insist those are fine for reasoning.
The Safety Lens. Energy turns into a selection pressure, and it pushes the wrong way. When every reasoning token costs 30x, the cheapest move is to compress the model's step-by-step thinking into something shorter and more opaque. That's the visible chain-of-thought, the part that lets you audit why a model answered the way it did, getting squeezed to save money. Worse for anyone building around the EU AI Act: its compute thresholds count training FLOPs, the math done to build the model. This study says the energy, and the risk, now live at query time. The rules are measuring the wrong stage.
Where they part ways. The Skeptic and the Builder agree on the fact and split on what it costs you. The Skeptic says energy-per-correct-answer is the right unit and reasoning might win on it. The Builder says your invoice doesn't bill per correct answer, it bills per token, and the token meter is running now. Both are right, which tells you the fix is measurement: track cost per solved task, not cost per call, so you know whether reasoning is actually earning its 30x.
The Compute Pragmatist and Safety Lens point at the same squeeze from opposite ends. One sees it forcing new silicon; the other sees it forcing shorter, harder-to-read reasoning. Same pressure, and the cheap way out of it is the one that makes models less auditable.
What this hinges on. One belief: is reasoning selectively invoked or creeping toward default? If it stays surgical, the Skeptic wins and 30x is a line item. If it creeps, the Compute and Builder concerns compound fast. So the thing to verify before your next budget cycle is dead simple: what fraction of your production calls already hit a reasoning path, and is that fraction rising month over month? Then cap the token budget on the reasoning route and measure cost per solved task against the standard model on the same task set.
Prediction: Neither OpenAI nor Anthropic will ship a per-query energy or standardized efficiency disclosure for their reasoning models by 2026-12-31, and will keep pricing reasoning purely by token count.
Confidence: Medium. The incentive to stay quiet on efficiency is strong and unopposed.
Why: The AI Energy Score exists and now gives a public yardstick, so the labs could adopt it, and the Meta/Cohere involvement shows some players want the standard to spread. But OpenAI and Anthropic sell reasoning as the premium tier, and a 30x energy number reframes that tier as the wasteful one right as regulators hunt for a metric to attach rules to. Token pricing already lets them capture the cost without ever publishing an efficiency figure that invites scrutiny or a compliance threshold. The opposite outcome, either lab voluntarily rating its own flagship as an energy hog, only happens under a rule that doesn't exist yet, and the EU AI Act still counts training compute, not query-time energy.
Revisit by 2026-12-31: We're right if neither OpenAI nor Anthropic publishes a standardized per-query energy or efficiency rating for a reasoning model and both still price reasoning by tokens alone. We're wrong if either one publishes such a figure or adopts the AI Energy Score for a flagship reasoning model.
Comments