How Wildlife Knowledge works
Each of the ten tasks asks a different kind of question. Choose a task to see what the models were asked, how the answers were scored, which reference database provided the answer, and the exact question every model received for one example, with each model's reply.
Tasks
How every question was asked
Each question was sent once to each model through OpenRouter, fixed to one provider, with only the instruction and the question: no conversation history, examples, web search or tools. Every model ran at the highest reasoning effort it offered, with room for 65,536 tokens of reasoning and answer. Answers were requested as a small JSON object so that they could be scored automatically, and every question let the model abstain by answering "null", apart from the real-or-invented question, which required a yes or no.
| Model | Developer | Weights | Provider route |
|---|---|---|---|
| OpenAI GPT-6.1 Sol | OpenAI | Closed | azure |
| Claude Haiku 5.5 | Anthropic | Closed | google-vertex/global |
| Gemini 4 Argon | Closed | google-ai-studio | |
| Grok 4.7 | xAI | Closed | xai/zdr |
| Qwen3.8 2.4T A95B | Alibaba | Open weights | alibaba |
| Mistral Large 4 | Mistral AI | Closed | mistral/zdr |
| Muse Spark 1.3 | Meta | Closed | meta |
About the examples
The examples are practice questions, used to check the wording before the scored run. They follow exactly the same wording as the scored questions, with different species, countries and name lists. The 185 scored questions and their reference answers are kept private so that future models cannot learn them.
The example answers come from that practice trial. It used the same models with two differences: the practice trial's output limit was 32,768 tokens (doubled to 65,536 for the scored run), and the common-name question then asked for the species' "common name in English" rather than its "common name in the English language". Open-weight Qwen joined the panel after the practice trial, so where it did not answer a practice question, the answer shown is from Qwen3.8 Max, the closed-weight model it replaced, and is labelled as such.
Independence
Model names and logos identify the models tested and remain the property of their owners. WildlifeBench is independent: it is not affiliated with, sponsored by or endorsed by any model developer. Results describe the named model versions, called through their APIs via OpenRouter in October–November 2026 under the conditions described here; they are not a verdict on any company's products as a whole.