Could you get the same AI results for less?

At a glance

A single interactive web page by Jeffery Kline that asks whether you are overpaying for an AI model. You pick a task, the model you use and its reasoning setting. It then reports whether a cheaper single model, or a weighted blend of models, scores at least as well on every benchmark for that task, and shows the estimated API cost of each per 1,000 benchmark tasks.

Good fit

Someone paying for API calls who wants a quick, method-backed second opinion on model choice for one of ten listed jobs, such as long-document reasoning, terminal coding or workflow automation. It is most useful as a shortlist of cheaper candidates to test, with models from OpenAI, Anthropic and Google (GPT-6.1 Sol, Claude Opus 5.5, Gemini 4 Argon and others) in the comparison.

Consider something else when

Skip it if your work does not resemble one of the ten tasks, if you use a model outside its list, or if latency, rate limits or switching effort matter more to you than benchmark cost. The page says itself that performance on your own documents is unmeasured and that splitting real work across a blend is untested.

Start here

Open the page, choose your task, your current model and its reasoning setting, then read the headline. For long-document reasoning with GPT-6.1 Terra (Low), for example, it reports that a 50/50 blend of GPT-6.1 Terra (Medium) and GPT-6 Luna (Max) matches both benchmark scores at an estimated $75.18 versus $129.65, 42.0% lower. Treat any match as a candidate to test on your own tasks.

Cost and requirements

It is a public page you use in a browser. The dollar figures are the page's estimates of API spend per 1,000 benchmark tasks, not a quote for your workload. Each task uses its own benchmark, cost unit and data date. The long-document task, for example, draws on Artificial Analysis scores dated October 26, 2026 (version 1.0, published November 2, 2026).

Limits to know

Every saving is a benchmark estimate that leaves out latency, quotas, switching and other deployment costs. Blends are weighted averages of cost and score, and the author notes that splitting real work this way is untested. Some tasks lean on proxies: frontend development uses automation and terminal benchmarks as a stand-in, and HLE covers advanced academic questions rather than proof writing. A label like "High" can mean different amounts of compute at different providers. Single-slide creation compares single models only. The method is data envelopment analysis from Zhou et al. (2022), co-authored by the page's author.

Still unclear

• Whether the scores and prices will be refreshed after version 1.0, or the page stays a November 2026 snapshot.
• Which provider list prices the cost estimates assume, and whether they reflect batch or cached-token discounts.