Which AI model is best? The Zaiq AI Benchmark

• Capability is the Epoch Capabilities Index. Epoch AI combines results from dozens of benchmarks, such as maths, coding, science and reasoning tests, into one scale. Higher is better. A gap of a point or two is within the margin of error; ten points is a clear difference.
• Facts right is SimpleQA Verified: short factual questions with one correct answer, asked without web search. It shows how often a model states a fact correctly instead of guessing. Lower scores mean more made-up answers.
• Science is GPQA Diamond: graduate-level biology, physics and chemistry questions written by experts. The top models now score above 90%, so it no longer separates them much.
• Rand per million tokensis what the model costs through its API. A token is roughly three quarters of a word, so a million tokens is about 1,500 pages. “In” is what you send; “out” is what the model writes back.
• 1,000 messages is our worked example of a customer-service assistant: each message sends about 1 000 tokens (instructions, context and the question) and gets about 300back. Reasoning models can “think” before answering and bill that thinking as output, so hard tasks can cost several times more.
• Reads at once is the context window: how much text the model can take in at one time, in pages of about 500 words.

• The hardest work (complex analysis, long reports, difficult code): GPT-6 Astra or Claude Fable 5.1. They cost the most, so use them where quality pays for itself.
• A customer assistant or document Q&A: Gemini 3.8 Flash or GPT-5.6 Luna. They are capable enough for most business questions and cost a fraction of the top models, which matters when thousands of messages come in.
• Keeping data on your own servers: an open-weights model such as Kimi K3, or DeepSeek's V4 models, can run on infrastructure you control, which helps with POPIA's rules on sending personal information abroad. It takes engineering to run well.
• Everyday use for free: free ChatGPT runs GPT-5.6 Luna. See ChatGPT free vs paid in South Africa for what each plan costs in rand.

Picking the model is the easy part. The value comes from connecting it to your own data and processes, with a person checking anything that touches money or customers. That is what ZAIQ builds.

We do not run the tests ourselves. The capability, accuracy and science scores come from Epoch AI, an independent research group that publishes its benchmark results openly under a Creative Commons Attribution licence (Epoch AI, ‘Capabilities & benchmarking’, retrieved 28 September 2026). Prices and context windows come from OpenRouter's public model list, converted to rand at US$1 = R16.32. For models Epoch has not scored yet, we note where Artificial Analysis places them on its own Intelligence Index.

Our part is choosing the models South Africans actually use, matching each one across both sources, working out what it costs in rand, and explaining it in plain English. Where Epoch has not yet tested a model on a benchmark, the table says so rather than guessing. The table is rebuilt every month from the same sources, and the previous editions stay available as downloads.