Wildlife species facts recalled by AI models
Wildlife Knowledge measures how accurately general-purpose AI models, the models behind the chat assistants many people now ask about nature, recall checkable facts about wild species. It covers ten kinds of knowledge that ecologists and conservation practitioners rely on: taxonomy, whether a species exists at all, IUCN Red List status, native and introduced ranges, national species checklists, common names, naming authorities, invasive species and current scientific names. Seven models from seven developers each answered the same 185 questions once, without web search or other tools, so the answers show what each model knows on its own. Every answer was scored automatically against openly licensed reference databases, such as the Catalogue of Life, the IUCN Red List and the Global Register of Introduced and Invasive Species, with no human or AI judge.Model comparison
The combined score is the average of the ten task scores, with every task weighted equally.
Chart waiting to load. Exact values are in the table below.
Cost & performance
Combined score against the reported cost of 100 questions through OpenRouter, at October–November 2026 prices. Models that reason at greater length cost more per answer.
Values for each point
How to read this chart
Lower cost is to the right, on a linear scale from $0. Lines on each point show the 95% interval of the combined score. Costs are the per-answer charges OpenRouter reported for the scored run, so they include each model's reasoning; they are not list prices.
Correct, wrong or abstained answers per model
Abstaining is better than a wrong answer. Abstentions count as not correct in the score, so this chart shows them separately.
Inspection of individual benchmarking questions
Spider chart
Compare two models across the ten tasks. Each axis is a task score from 0 to 100%; the combined score is their average.
Results table
CI shows each score's 95% interval: the range the score would be expected to fall in for 95% of repeat samples of questions. Hover over a score to see its interval, or switch CI on to show every interval.| Model | Combined score | Taxonomy | Existence | Red List | Native range | Introduced range | Checklists | Common names | Authority | Invasive species | Nomenclature |
|---|---|---|---|---|---|---|---|---|---|---|---|
| Muse Spark 1.3 | 84.1% (80.6%–87.5%) | 100.0% (83.9%–100.0%) | 94.3% (87.7%–99.2%) | 67.6% (51.5%–80.4%) | 66.6% (47.5%–84.4%) | 58.4% (45.3%–70.2%) | 96.5% (93.8%–98.8%) | 95.0% (76.4%–99.1%) | 90.0% (69.9%–97.2%) | 93.0% (88.0%–97.0%) | 80.0% (65.0%–95.0%) |
| OpenAI GPT-6.1 Sol | 80.3% (76.5%–84.1%) | 95.0% (76.4%–99.1%) | 83.4% (72.6%–93.4%) | 62.2% (46.1%–75.9%) | 69.5% (50.7%–86.6%) | 49.6% (34.2%–64.6%) | 96.0% (93.5%–98.3%) | 95.0% (76.4%–99.1%) | 95.0% (76.4%–99.1%) | 92.0% (88.0%–95.5%) | 65.0% (45.0%–90.0%) |
| Gemini 4 Argon | 79.0% (75.6%–82.5%) | 95.0% (76.4%–99.1%) | 93.7% (89.7%–97.6%) | 62.2% (46.1%–75.9%) | 59.4% (39.2%–79.6%) | 23.9% (9.7%–39.0%) | 97.9% (96.1%–99.2%) | 100.0% (83.9%–100.0%) | 100.0% (83.9%–100.0%) | 92.5% (87.5%–96.5%) | 65.0% (45.0%–90.0%) |
| Grok 4.7 | 70.2% (65.3%–75.3%) | 90.0% (69.9%–97.2%) | 89.7% (84.9%–94.4%) | 43.2% (28.7%–59.1%) | 49.0% (29.5%–68.6%) | 26.0% (11.5%–41.9%) | 92.0% (88.6%–95.3%) | 95.0% (76.4%–99.1%) | 80.0% (58.4%–91.9%) | 87.5% (80.5%–93.0%) | 50.0% (20.0%–80.0%) |
| Qwen3.8 2.4T A95B (open weights) | 60.5% (55.9%–65.1%) | 75.0% (53.1%–88.8%) | 57.5% (50.0%–65.0%) | 54.1% (38.4%–69.0%) | 42.6% (24.8%–60.4%) | 19.8% (6.5%–35.0%) | 84.5% (78.2%–90.4%) | 80.0% (58.4%–91.9%) | 50.0% (29.9%–70.1%) | 81.5% (73.5%–88.5%) | 60.0% (45.0%–75.0%) |
| Mistral Large 4 | 59.0% (54.5%–63.5%) | 75.0% (53.1%–88.8%) | 68.4% (58.4%–79.2%) | 54.1% (38.4%–69.0%) | 42.8% (24.6%–61.5%) | 46.5% (34.2%–58.6%) | 87.6% (81.5%–93.1%) | 65.0% (43.3%–81.9%) | 25.0% (11.2%–46.9%) | 85.5% (79.0%–91.5%) | 40.0% (40.0%–40.0%) |
| Claude Haiku 5.5 | 55.7% (51.1%–60.3%) | 75.0% (53.1%–88.8%) | 74.3% (63.4%–85.1%) | 37.8% (24.1%–53.9%) | 31.8% (13.2%–51.8%) | 23.3% (13.9%–33.4%) | 84.5% (76.6%–91.3%) | 75.0% (53.1%–88.8%) | 25.0% (11.2%–46.9%) | 85.5% (78.5%–91.5%) | 45.0% (40.0%–55.0%) |
- Taxonomy Taxonomy knowledge · 20 questions
- Existence Species existence knowledge · 83 questions
- Red List IUCN Red List knowledge · 37 species
- Native range Native range knowledge · 20 species
- Introduced range Introduced range knowledge · 21 species
- Checklists Country checklist knowledge · 18 countries
- Common names Common name knowledge · 20 questions
- Authority Naming authority knowledge · 20 questions
- Invasive species Invasive species knowledge · 20 questions
- Nomenclature Nomenclature knowledge · 4 lists · indicative
Rank ranges
Point rank and the range of ranks across 10,000 bootstrap resamples (1 is best; tied models share an average rank).
| Model | Combined | Taxonomy | Existence | Red List | Native range | Introduced range | Checklists | Common names | Authority | Invasive species | Nomenclature |
|---|---|---|---|---|---|---|---|---|---|---|---|
| Muse Spark 1.3 | 1 (1–1) | 1 (1–2.5) | 1 (1–3) | 1 (1–3.5) | 2 (1–4) | 1 (1–2) | 2 (1–3) | 3 (2.5–3.5) | 3 (1.5–4) | 1 (1–3) | 1 (1–2.5) |
| OpenAI GPT-6.1 Sol | 2 (2–3) | 2.5 (2–3.5) | 4 (2–5) | 2.5 (1–5) | 1 (1–3) | 2 (1–3) | 3 (1–3) | 3 (2.5–3.5) | 2 (1.5–3) | 3 (1–4) | 2.5 (2.5–3.5) |
| Gemini 4 Argon | 3 (2–3) | 2.5 (2–3.5) | 2 (1–3) | 2.5 (1–5) | 3 (2–6) | 5 (4–7) | 1 (1–2) | 1 (1–2.5) | 1 (1–2) | 2 (1–4) | 2.5 (2.5–3.5) |
| Grok 4.7 | 4 (4–4) | 4 (2.5–5) | 3 (2–4) | 6 (4–7) | 4 (3–7) | 4 (4–7) | 4 (4–5) | 3 (2.5–3.5) | 4 (3–4) | 4 (3–6) | 5 (2.5–7) |
| Qwen3.8 2.4T A95B | 5 (5–6) | 6 (3–7) | 7 (7–7) | 4.5 (1–6) | 6 (3–7) | 7 (4–7) | 7 (5–7) | 5 (3.5–7) | 5 (5–6) | 7 (6–7) | 4 (2–5.5) |
| Mistral Large 4 | 6 (5–7) | 6 (4.5–7) | 6 (5–6) | 4.5 (1–6) | 5 (3–7) | 3 (2–3) | 5 (4–7) | 7 (6–7) | 6.5 (5.5–7) | 6 (3.5–7) | 7 (5.5–7) |
| Claude Haiku 5.5 | 7 (6–7) | 6 (4.5–7) | 5 (4–6) | 7 (6–7) | 7 (5–7) | 6 (4–7) | 6 (5–7) | 6 (5–6.5) | 6.5 (6–7) | 5 (4–7) | 6 (5–6.5) |
How to read these results
This is an exploratory pilot. Each model answered each question once, so the results do not show how much a model's answers vary between runs. With about 20 questions per task, small differences between models are within the uncertainty, which the interval lines show. Nomenclature knowledge rests on four lists of five names per model and is indicative only; it will be expanded in a later version.
Scores measure agreement with versioned reference databases. Where a database was out of date, a published source was used instead. The 185 scored questions and their reference answers are kept private, so that future models cannot learn them; the Methodology page shows the exact wording of every question type with example answers.
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.