Joules-to-GDP: Smaller Gemma 4 31B beats GLM 5.2 on value per unit of energy

The most valuable model is not the most efficient

The study compared five models on 220 GDPval tasks. Snorkel AI, which describes the findings on its blog, says its research team provided expert baselines for those tasks and helped develop GDPWorks, a set of 83 new tasks covering 36 occupations that GDPval does not include.

GLM 5.2 produced the most output value, $74.19 per task, using 0.430 MJ of energy, which Snorkel says is about the energy of four hours of active laptop use. Gemma 4 31B produced 44% less output value, but it returned about 2.5 times the net value per joule. Net value here means the dollar value of the AI's work minus what it cost to run. Gemma 4 31B led on both net value and expert time saved per joule, at 417.5 USD/MJ and 10.86 hours/MJ.

That does not mean the smaller model always wins. A model with a high return per joule can still finish too little of the work, which is why the paper reports total value next to the per-joule numbers. GLM 5.2 also ran on eight H200s, because its FP8 serving setup did not fit on the eight-H100 node used for the other four models, so its figures reflect both model and hardware.

Energy reveals what cost and speed hide

GPT-OSS 120B and GLM 5.2 had nearly the same cost and speed. GPT-OSS 120B cost about 3.1% more and finished about 2.1% faster, yet it used 68% less energy. GLM 5.2 delivered more net value per dollar of AI cost (37.69 vs 19.09) and per AI hour ($191.3 vs $102.1), while GPT-OSS 120B delivered more per megajoule (275.7 vs 168.1 USD/MJ). The choice of metric changes the winner.

Hardware also matters. Running Qwen3.6-27B on a single accelerator, the NVIDIA B200 returned $535.5 of net value per megajoule, against 391.3 for the H200, 402.2 for the H100 and 124.1 for the AMD MI350X. The paper points to differences in chip design, memory bandwidth and power draw. Because energy is measured across the whole run, including time spent waiting on tool calls, a chip's energy use depends on both its power draw and how long the work takes. The H200 and H100 came out about even, so the newer chip was not automatically better.

More GPUs and more retries do not pay for their energy

Moving Qwen3.6-27B from one H100 to eight made it only 22% faster, while energy use rose 3.24 times. Output value was about the same ($64.79 vs $67.81 per task), and net value per joule fell from 402.2 to 123.4 USD/MJ. Snorkel explains that the GPUs must pass data between each other and wait, and the agent's tool calls, such as running code or editing files, happen one at a time regardless of GPU count. All eight GPUs drew power for the whole run.

Retries showed a similar pattern. With four attempts per task, GLM 5.2's output value rose from $74.19 to $93.42, but energy went from 0.430 to 1.760 MJ and net value per joule fell from 168.1 to 48.6 USD/MJ. All four attempts ran and counted toward energy use, even when an earlier one had succeeded. A system that stops after a good answer would use less energy, and the paper did not test that.

Experts who use AI set a higher bar

Most results compare the AI with experts working alone. The paper also scored the same GLM 5.2 outputs against experts who worked with AI, which is closer to how many professionals work today. Net value per megajoule then fell 70%, from $168.1 to $50.0. The AI-assisted experts set a higher bar, and GLM 5.2 cleared it about half as often. They were also about 23% faster on average, leaving less time for the AI to save. Where GLM 5.2 did match them, it finished about 84% faster.

Energy also changes which kinds of work look most valuable. Architecture and engineering tasks produced about 10% more net value per task than software development, but software development produced about 27% more per joule. Each occupation has only five tasks, so the authors caution that these rankings rest on a small sample. When they removed one task per category at random across 10,000 trials, Computer and Mathematical stayed first on net value per joule in 76.5% of them.

How the measurement works and what it means

Net Value per Joule is the dollar value of an AI output, minus the cost of running the system, divided by energy used. Time Saving per Joule is expert time saved divided by the same energy. A judge model, Gemini 3.1 Flash-Lite, compares each AI output with an expert's deliverable without knowing which is which. A win or tie earns the task's dollar value, calculated from the expert's time and the median hourly wage for the occupation. A loss earns nothing, but cost and energy still count. Energy covers GPU, CPU and memory over the whole run, measured with Intelligence per Watt.

Snorkel says each of the 220 GDPval tasks was completed twice by different experts, once without AI and once with AI assistance, the latter using Claude through the Claude web chat interface. GDPWorks adds 36 occupations, such as bioinformatics scientists and purchasing managers, and its tasks are more demanding: a median expert-only time estimate of 4.0 hours versus 2.34 for GDPval.

For anyone choosing models or hardware, the findings suggest that cost and speed alone can mislead, and that the best pick depends on whether total output or return per unit of energy matters more.