Ethical Tech CoLab  /  AI Models Research

Cost, speed, and energy per accepted task

An interactive companion to a July 2026 comparative review of frontier, open-weight, and efficient AI model families. Every figure below carries its source, its date, and how far it has been independently verified.

Research cut-off 2026-07-22 A independent B institutional C provider-reported

There is no universally best AI model. Selection is a constrained optimisation across verified task quality, factual reliability, latency, token consumption, energy per accepted output, privacy, and total deployment cost.

The four instruments below let you interrogate that claim rather than take it on faith. The first is the one that matters most: advertised price per million tokens is not cost, because a model that has to be retried is not cheap.

What a request actually costs

Set the shape of your workload. The calculator applies the published input and output rates, adds reasoning tokens at the output rate, and then divides by the acceptance rate to give cost per accepted task. Defaults reproduce the document-analysis example in the review.

Billed at the output rate where a provider bills them at all.
Share of attempts that pass your quality check without manual correction.

Nominal cost per request

Read the spread, not the ranking. At the default settings the cheapest and most expensive options differ by roughly 26 times for the same request. Quality is not held constant across that spread, so the chart is a cost measurement and not a value judgement. Move the acceptance rate below 100% to see why a cheaper model can be the more expensive choice.

Output tokens carry the cost

Output is priced three to six times above input across every published schedule in the review, and it is generated sequentially. Verbosity is therefore a cost driver, a latency driver, and an energy driver at once.

Nominal API price per million tokens

Gemini 3.1 Pro is absent because the review did not extract a per-token rate for it; the provider publishes pricing only as a live schedule.

Stored parameters are not activated parameters

A mixture-of-experts model stores many parameters and activates a small subset per token. That lowers compute per token relative to a dense model of the same stored size. It does not remove the memory, routing, or communication cost, so active share is a compute proxy and not a serving-cost measure.

Share of stored parameters activated per token

Agents approach human performance without reaching it

Interactive agent benchmarks are where the remaining gap is clearest. Production systems fail through sequences of small errors rather than single wrong answers, which is why a few percentage points on a long-horizon task matter more than the same gap on a single-answer test.

Best reported model result against the human baseline

Vertical marks show human baselines where the AI Index reports one. Human scores are often measured under different time, tool, and incentive conditions than model scores, so the gap is indicative rather than exact.

Reasoning budget, not model size, drives energy per query

No single energy number describes a model. Prompt length, output length, batch size, precision, hardware generation, serving stack, and utilisation each change the result by large factors. The two scenarios below are the only two the source states, so the toggle switches between them rather than interpolating.

Estimated energy per frontier-scale query

Optimisation matters more than model choice. A peer-reviewed study of inference across hardware, serving frameworks, batching, and decoding strategies found that appropriate optimisations cut energy by up to 73% against an unoptimised baseline. A peer-reviewed

Do not divide sector totals by a query count. Data-centre electricity is projected to reach about 945 TWh in 2030 against roughly 415 TWh in 2024, but that covers every data-centre workload, not AI queries alone. B institutional

How to read any number on this page

The grade tells you how much independent scrutiny a figure has survived. It does not tell you whether the figure is true. A Grade A result can be wrong and a Grade C claim can be accurate.

A independent

Peer-reviewed studies, standardised academic benchmarks, and reproducible third-party evaluations. The only class admissible on its own for a comparative claim between providers.

B institutional

Technical reports, model cards, benchmark methodology pages, and official specifications. A provider is the definitive authority on its own commercial terms, so prices and context limits sit here.

C provider-reported

Launch benchmark tables and vendor performance claims. Recorded because excluding them would leave most commercial models undocumented. Never used alone to rank one provider above another.

! conditions unstated

Every provider benchmark figure in the review omits its harness, sampling policy, and tool permissions. Results whose conditions are unknown are recorded but are never placed in a ranked column beside results whose conditions are known.

Prices and rankings change without notice. Every value here is a record of what a named source stated on 2026-07-22, not a statement about the present. Verify against current provider documentation before any procurement, budgeting, or deployment decision. Nothing on this page is legal, financial, or procurement advice.

Sources

  1. Stanford Institute for Human-Centered Artificial Intelligence (2026). AI Index Report 2026. Grade B.
  2. Oviedo, F., Kazhamiaka, F., Choukse, E., Kim, A., Luers, A., Nakagawa, M., Bianchini, R., and Lavista Ferres, J. M. (2026). Energy use of AI inference: efficiency pathways and test-time compute. Joule. Grade A.
  3. Fernandez, J., Na, C., Tiwari, V., Bisk, Y., Luccioni, S., and Strubell, E. (2025). Energy considerations of large language model inference and efficiency optimizations. Proceedings of ACL 2025, 32556 to 32569. Grade A.
  4. International Energy Agency (2025, 2026). Energy and AI; Key questions on energy and AI. Grade B.
  5. Official provider documentation for OpenAI, Anthropic, Google DeepMind, xAI, DeepSeek, Mistral, and Meta, accessed 2026-07-22. Grade B for commercial terms, Grade C for performance claims.

The full source register, the extraction protocol, and the validation tooling live in the research repository. Figures reproduced here are drawn from that review and inherit its limitations, which are stated in full alongside it.