GPT-6.2 Astra
Rank 2–4 of 29 · 3 models too close to separate.
Guessing earns nothing: a wrong answer scores below zero. The Index weights six clinical categories by caseload, then applies four multipliers that can only reduce it. The interval is a bootstrap 95% CI. How this is measured →
Safety penalty ×0.96: 1.4% of answers on safety-critical questions were wrong, weighted by severity.
Harm rate: 14.20 per 1,000 answers on safety-critical questions, weighted by severity. Any wrong answer on a safety-critical question counts, at the severity veterinarians assigned to that question. Declined and cut-off replies do not.
Raw index 83.2, reduced by safety penalty ×0.96, consistency ×0.94, safety coverage ×0.95. Published index 71.4.
The safety penalty multiplies the Index rather than being averaged into it. How this is measured →
Vendor response
OpenAI has not responded to this result. A vendor can respond to any published result, and the response is published here, beside the entry, as received. Send a response
Per-category profile
Each category on a common 0–100 axis with its 95% confidence interval, the honest alternative to a radar chart.
View as data table
| Category | Score | 95% CI |
|---|---|---|
| Clinical Diagnosis & Differential | 87.4 | 83.6–91.1 |
| Pharmacology & Dosing | 82.7 | 78.1–87.3 |
| Species-Specific Knowledge | 81.6 | 75.9–87.3 |
| Triage & Emergency | 82.9 | 77.0–88.7 |
| Preventive / Public Health / Zoonoses | 76.8 | 69.7–84.0 |
| Communication & Professionalism | 83.3 | 74.4–92.2 |
| Safety & Harm Avoidance | 92.0 | 87.3–96.7 |
Cost · speed · reliability
How it compares
GPT-6.2 Astra against every other published model. Switch the axis to compare on safety, harm rate, cost, latency or reliability.
GPT-6.2 Astra against the field
The weighted composite across every category. The line through each bar is the 95% interval: where two intervals overlap, the order between those models is not established. A model whose safety review is pending has no bar until a veterinarian finishes it.
View as a table
| Model | Vendor | VetEval Index | 95% interval |
|---|---|---|---|
| Claude Opus 5.5 | Anthropic | 76.8 | 74.8 to 78.8 |
| Kimi K3.1 Pro | Moonshot AI | 71.6 | 69.6 to 73.4 |
| GPT-6.2 Astra | OpenAI | 71.4 | 69.3 to 73.6 |
| GPT 5.6 Sol | OpenAI | 69.8 | 67.7 to 71.8 |
| Gemini 3.7 Flash | 66.4 | 64.2 to 68.5 | |
| Gemini 4 Carbon | 66.2 | 64.3 to 68.1 | |
| Gemini 4 Argon | 64.5 | 62.5 to 66.7 | |
| Claude Opus 5 | Anthropic | 63.9 | 62.2 to 65.7 |
| GPT-6 Sol | OpenAI | 62.4 | 60.3 to 64.4 |
| Gemini 4 Hydrogen | 62.3 | 60.4 to 64.4 | |
| GPT-6.2 Sol | OpenAI | 61.8 | 59.8 to 63.9 |
| Muse Wonder | Meta | 58.3 | 56.2 to 60.3 |
| GPT-5.6 Terra | OpenAI | 57.6 | 55.7 to 59.5 |
| GPT 6 Luna | OpenAI | 55.7 | 53.7 to 57.6 |
| Kimi K3 | Moonshot AI | 55.5 | 53.9 to 57.0 |
| Gemini 3.8 Flash | 55.1 | 53.1 to 57.0 | |
| GLM 5.3 | Z.AI | 54.1 | 52.2 to 56.0 |
| Grok 4.6 | SpaceXAI | 54.0 | 52.2 to 56.0 |
| Grok 4.7 | SpaceXAI | 48.6 | 46.7 to 50.5 |
| GLM 5.3 Flash | Z.AI | 48.5 | 46.9 to 50.2 |
| Claude Sonnet 5 | Anthropic | 46.4 | 44.4 to 48.5 |
| Mistral Large 4 | Mistral | 45.7 | 44.1 to 47.4 |
| Minimax M3 | Minimax | 38.4 | 36.8 to 39.9 |
| Llama 4 Maverick | Meta | 38.1 | 36.4 to 39.8 |
| Claude Haiku 5.5 | Anthropic | 32.7 | 31.1 to 34.4 |
| Hy4 | Tencent | 31.5 | 30.1 to 33.0 |
| Qwen3.8 2.4T A95B | Alibaba | 25.3 | 24.3 to 26.3 |
| Nemotron 3.5 Lightning | NVIDIA | 3.7 | 3.3 to 4.2 |
| Granite 4.2 8b | IBM | 3.4 | 3.1 to 3.7 |
Run settings
The settings this evaluation used. Every scored run uses the same fixed settings, and a run that changes them is refused.
- Model identifier
- openai/gpt-6.2-astra
- Temperature
- 0
- Seed
- 20260804
- Max output tokens
- 2048
- Retry budget
- 8192 tokens, for one retry of a reply that was cut off
- Reasoning effort
- minimal
- Passes per question
- 3, with the answer choices reshuffled each pass
- Run by
- VetEval
- Date evaluated
- 18 Nov 2026
- Pricing
- $8 input, $40 output, per million tokens