AI energy efficiency has a measurement problem hiding inside it: efficiency at the level of a single model release does not automatically translate into lower total energy use, a point LiveAIWire’s companion piece on the green AI myth covers in detail. What that piece does not do is tell you what to actually change about how your own organisation or product uses AI. This one does.
The good news is that the single highest-leverage AI energy efficiency decision available today requires no new hardware, no new research, and no waiting for the industry to fix its disclosure problem. It requires matching the model to the task, rather than defaulting to the biggest, most capable model available for every query.
The Model Selection Number Behind Real AI Energy Efficiency Gains
A systematic study published in October 2025 by researchers at Université Côte d’Azur and CNRS put a precise figure on what model selection alone could achieve. Analysing benchmarks across eleven common AI tasks, from text generation and translation to image classification and speech recognition, the researchers found that redirecting inference requests from the best-performing model to an appropriately sized, energy-efficient alternative could cut global AI energy consumption by 27.8 percent, saving an estimated 31.9 terawatt-hours worldwide in 2025 alone, roughly the annual output of five nuclear power reactors.
The utility cost of that saving was small. Across the tasks studied, switching from the largest available model to the identified energy-efficient one delivered 65.8 percent better energy efficiency on average, at a utility loss of just 3.9 percent. For some mature tasks the trade was even better than break-even: the researchers found that average utility actually increased by 4 percent after switching, because many of the models people default to out of habit or brand familiarity are not the best-performing ones available, just the most familiar.
What This Means for You
If your team is choosing which AI model to call for a given feature, AI energy efficiency should be one of the questions asked before a model is locked in, not an afterthought raised once a bill or a sustainability report arrives. The practical version of the question is simple: does this task need the most capable model available, or would a smaller, cheaper, faster model perform it just as well. For tasks like classification, extraction, translation, and routine summarisation, the answer is very often a smaller model than the one currently deployed.
Where the Savings Are Biggest, and Where They Are Not
The October 2025 study found the achievable savings varied enormously by task, from 1 percent to 98 percent depending on how mature the task is and how much the community has already invested in efficient alternatives. Time series forecasting showed the largest potential saving, at 92.8 percent, followed by speech recognition at 80.6 percent and image classification at 69.6 percent. Immature tasks fared worse: mathematical reasoning showed only a 36 percent saving and text-to-image generation just 18.3 percent, because efficient, well-benchmarked alternatives have not yet caught up with the best-performing models in those categories.
That unevenness matters for anyone setting an internal AI energy efficiency policy. A blanket instruction to “use smaller models” will save more in some parts of a product than others, and treating all AI tasks as equally amenable to downsizing will misallocate the effort. The tasks worth auditing first are the mature, high-volume ones: classification, translation, and routine text generation, where efficient alternatives already exist and are well documented.
The Reasoning Trap Nobody Is Pricing In
Hugging Face’s AI Energy Score project, run jointly with Salesforce, Cohere and Carnegie Mellon University, published a refreshed leaderboard in December 2025 that identifies a specific AI energy efficiency trap most teams are not yet accounting for: reasoning modes. Models with reasoning enabled use, on average, 30 times more energy than the same model with reasoning turned off. In the most extreme cases the project measured, a single model used between 150 and 700 times more energy with reasoning switched on, driven almost entirely by the far larger number of output tokens a reasoning model generates while working through a problem.
The same leaderboard update delivered an uncomfortable finding for anyone assuming AI energy efficiency simply improves on its own over time. Comparing a cohort of new models released in late 2025 against similarly sized models from February 2025, the majority, eight out of fourteen, used equal or greater energy than their predecessors. Some used almost twice as much.
That directly undercuts the assumption that newer necessarily means more efficient, and it is exactly the pattern the green AI myth piece describes at the industry level, now visible model by model. It also lines up with the task-level multipliers documented in the United Nations University’s 2026 environmental footprint report, where a single AI-generated image already used roughly 1,450 times the energy of basic text classification before reasoning was even part of the picture.
What Actually Works in Practice
Reasoning should be treated as a deliberate, task-specific decision rather than a default toggle left on for every query. A customer support classification task or a routine data extraction job does not need a model working through multi-step internal reasoning to produce an answer; a simpler model, or the same model with reasoning switched off, will typically do the job at a fraction of the energy cost. Reserve reasoning modes for the genuinely complex tasks, mathematical proofs, multi-step planning, ambiguous judgement calls, where the accuracy gain actually justifies the multiple-hundred-fold energy increase.
Model routing, automatically directing a query to the smallest model capable of handling it rather than a single default model for everything, is the practical mechanism that makes AI energy efficiency achievable at scale without asking every individual user to make a manual choice each time. Several routing tools have emerged specifically to pair Hugging Face’s energy benchmarks with performance data, letting a system pick the cheapest adequate model automatically.
This is the same logic LiveAIWire’s coverage of generative versus predictive AI makes for choosing the right category of AI for a business problem in the first place: matching tool to task, rather than defaulting to whichever tool is most capable, is the decision that determines both cost and energy footprint. Treating AI energy efficiency as a routing problem rather than a hardware problem is what makes it something a product team can act on this quarter, not a research agenda for next year.
Salesforce’s own adoption of the AI Energy Score framework is a useful example of what this looks like at company scale. The company has integrated the benchmark into its internal model evaluation process and now generates energy transparency data automatically as part of every model card, publishing that information for its production models going forward. None of that required new hardware. It required treating energy consumption as a metric worth measuring at the point a model is chosen, the same way performance and cost already are.
Measuring AI Energy Efficiency Before You Can Manage It
None of the model selection or reasoning discipline described above is possible without a baseline. Hugging Face’s AI Energy Score leaderboard, now covering 166 models across ten tasks and multiple modalities including text, image and audio, gives any team a public, standardised reference point for comparing AI energy efficiency across open models before committing to one in production. For proprietary models not on the public leaderboard, the project’s Docker-based submission process lets a company benchmark its own deployment under the same methodology, which is exactly the route Salesforce took.
The absence of a shared measurement standard has been one of the biggest obstacles to taking AI energy efficiency seriously at the procurement stage, because two vendors reporting efficiency figures using different methodologies cannot be meaningfully compared. Regulatory efforts including the EU AI Act’s Code of Practice, signed by major AI developers, now explicitly call for a standardised inference energy benchmark for exactly this reason. Until that becomes universal, asking any AI vendor which benchmark methodology backs their AI energy efficiency claims, and whether that methodology is publicly documented, is the single most useful procurement question available.
What This Looks Like as a Standing Practice
Treating AI energy efficiency as a standing practice rather than a one-time audit means revisiting model choice whenever a task’s requirements change, whenever a new model release claims better performance, and whenever usage volume for a given feature grows enough that the energy cost of the default model becomes material. A model that was the right choice at low volume is not automatically the right choice once that feature serves millions of queries a day, because the aggregate energy cost scales with volume in a way a single engineer’s initial choice rarely accounts for.
The organisations already ahead on this, Salesforce’s automatic model-card disclosure being the clearest public example, treat AI energy efficiency the same way they treat cost and latency: a metric tracked continuously, not a claim made once in a sustainability report and left unchecked for the following year.
Why This Matters More Than a Single Efficiency Statistic
None of these measures require the AI industry to solve its broader disclosure problem first, which is precisely why they are worth adopting now rather than waiting. LiveAIWire’s reporting on machine greenwashing in the AI industry found that vendor-level efficiency claims are routinely presented in ways that obscure total impact. Model selection and reasoning discipline sidestep that problem entirely, because the savings are measured at the level of the specific task an organisation actually controls, not at the level of a company-wide statistic nobody outside that company can audit.
The honest caveat is the same one that applies to AI energy efficiency at the industry level: none of this guarantees total AI energy consumption falls, because a genuinely efficient model that is cheaper to run tends to get used more. But model selection and reasoning discipline are different in kind from a headline efficiency statistic, because they reduce the energy cost of a fixed amount of existing work, rather than making a new capability cheaper to scale into new demand. That distinction is what makes them a more reliable lever for actual emissions reduction than waiting for the next model release to solve the problem on its own.
About the Author
Stuart Kerr is Technology Correspondent at LiveAIWire, covering artificial intelligence, emerging technology, and their impact on business, society, and everyday life. LiveAIWire publishes original AI journalism every weekday at liveaiwire.com.