# MedQA / white > Source: Benchize (https://benchize.com) — https://benchize.com/benchmarks/vals-ai%3Amedqa%3Awhite. When citing or reusing this content, credit "Benchize" and link to https://benchize.com/benchmarks/vals-ai%3Amedqa%3Awhite. Evaluating language model bias in medical questions. Canonical URL: https://benchize.com/benchmarks/vals-ai%3Amedqa%3Awhite Last updated: Wed, 21 Oct 2026 18:52:20 GMT Reproduction and quotation must name Benchize, link to the canonical HTML URL, and retain individual source citations. Evaluating language model bias in medical questions. Professional work · Models tested: 12 · Score publishers: Vals AI ## Model results | # | Model | Score | ± | Cost | Latency | | | --- | --- | --- | --- | --- | --- | --- | | 1 | [Gemini 3 Flash Preview](https://benchize.com/models/gemini-3-flash-preview-high.md)High | 96.3% | SE ±0.42 pp | $0.01 | 27.09source duration | | | ### Original source values \| Metric \| Value \| Unit \| Scope \| Direction \| \| --- \| --- \| --- \| --- \| --- \| \| `accuracy` \| 96.3 \| percent \| reported \| Higher is better \| \| `cost_per_test` \| 0.013138 \| USD \| per_test \| Lower is better \| \| `latency` \| 27.09 \| source_duration \| reported_duration \| Lower is better \| ### Configuration computeEffort Not reported harness Not reported maxOutputTokens 65536 provider Google reasoning Not reported Reasoning effort high Temperature 1 topP Not reported verbosity Not reported ### Reported statistics Uncertainty Standard error: ±0.42 pp sourcePopulation 95 stderr 0.422 tokenTotals Not reported ### Source [Original source](https://www.vals.ai/_astro/benchmark_view_medqa.Cco1kvog.json) Source Vals AI Original model label google/gemini-3-flash-preview Source version 1 Observed 2026-10-21T16:46:00.691Z Evidence SHA-256 `d67905978d932f64698ebe9e230d9519e5bae1c9d7176dda8379dc3874b6f90f` Source locator `/tasks/white/google~1gemini-3-flash-preview` | | | | | | | | 2 | [Gemini 3.1 Pro Preview](https://benchize.com/models/gemini-3-1-pro-preview-high.md)High | 96.25% | SE ±0.43 pp | $0.02 | 37.72source duration | | | ### Original source values \| Metric \| Value \| Unit \| Scope \| Direction \| \| --- \| --- \| --- \| --- \| --- \| \| `accuracy` \| 96.25 \| percent \| reported \| Higher is better \| \| `cost_per_test` \| 0.017949 \| USD \| per_test \| Lower is better \| \| `latency` \| 37.716 \| source_duration \| reported_duration \| Lower is better \| ### Configuration computeEffort Not reported harness Not reported maxOutputTokens 65536 provider Google reasoning Not reported Reasoning effort high Temperature 1 topP Not reported verbosity Not reported ### Reported statistics Uncertainty Standard error: ±0.43 pp sourcePopulation 95 stderr 0.425 tokenTotals Not reported ### Source [Original source](https://www.vals.ai/_astro/benchmark_view_medqa.Cco1kvog.json) Source Vals AI Original model label google/gemini-3.1-pro-preview Source version 1 Observed 2026-10-21T16:46:00.691Z Evidence SHA-256 `d67905978d932f64698ebe9e230d9519e5bae1c9d7176dda8379dc3874b6f90f` Source locator `/tasks/white/google~1gemini-3.1-pro-preview` | | | | | | | | 3 | [MiniMax M2.5](https://benchize.com/models/minimax-m2-5.md) | 92.35% | SE ±0.59 pp | $0.00295 | 33.53source duration | | | ### Original source values \| Metric \| Value \| Unit \| Scope \| Direction \| \| --- \| --- \| --- \| --- \| --- \| \| `accuracy` \| 92.35 \| percent \| reported \| Higher is better \| \| `cost_per_test` \| 0.002954 \| USD \| per_test \| Lower is better \| \| `latency` \| 33.532 \| source_duration \| reported_duration \| Lower is better \| ### Configuration computeEffort Not reported harness Not reported maxOutputTokens 196608 provider MiniMax reasoning Not reported Reasoning effort unknown Temperature 1 topP 0.95 verbosity Not reported ### Reported statistics Uncertainty Standard error: ±0.59 pp sourcePopulation 95 stderr 0.594 tokenTotals Not reported ### Source [Original source](https://www.vals.ai/_astro/benchmark_view_medqa.Cco1kvog.json) Source Vals AI Original model label minimax/MiniMax-M2.5 Source version 1 Observed 2026-10-21T16:46:00.691Z Evidence SHA-256 `d67905978d932f64698ebe9e230d9519e5bae1c9d7176dda8379dc3874b6f90f` Source locator `/tasks/white/minimax~1MiniMax-M2.5` | | | | | | | | 4 | [Claude Sonnet 4.6](https://benchize.com/models/claude-sonnet-4-6-thinking.md)Thinking | 92% | SE ±0.61 pp | $0.04 | 34.9source duration | | | ### Original source values \| Metric \| Value \| Unit \| Scope \| Direction \| \| --- \| --- \| --- \| --- \| --- \| \| `accuracy` \| 92 \| percent \| reported \| Higher is better \| \| `cost_per_test` \| 0.035394 \| USD \| per_test \| Lower is better \| \| `latency` \| 34.904 \| source_duration \| reported_duration \| Lower is better \| ### Configuration computeEffort Not reported harness Not reported maxOutputTokens 128000 provider Anthropic reasoning true Reasoning effort unknown Temperature 1 topP Not reported verbosity Not reported ### Reported statistics Uncertainty Standard error: ±0.61 pp sourcePopulation 95 stderr 0.607 tokenTotals Not reported ### Source [Original source](https://www.vals.ai/_astro/benchmark_view_medqa.Cco1kvog.json) Source Vals AI Original model label anthropic/claude-sonnet-4-6 Source version 1 Observed 2026-10-21T16:46:00.691Z Evidence SHA-256 `d67905978d932f64698ebe9e230d9519e5bae1c9d7176dda8379dc3874b6f90f` Source locator `/tasks/white/anthropic~1claude-sonnet-4-6` | | | | | | | | 5 | [MiniMax M2.1](https://benchize.com/models/minimax-m2-1.md) | 90.8% | SE ±0.65 pp | $0.00232 | 34.69source duration | | | ### Original source values \| Metric \| Value \| Unit \| Scope \| Direction \| \| --- \| --- \| --- \| --- \| --- \| \| `accuracy` \| 90.8 \| percent \| reported \| Higher is better \| \| `cost_per_test` \| 0.00232 \| USD \| per_test \| Lower is better \| \| `latency` \| 34.686 \| source_duration \| reported_duration \| Lower is better \| ### Configuration computeEffort Not reported harness Not reported maxOutputTokens 131000 provider MiniMax reasoning Not reported Reasoning effort unknown Temperature 1 topP 0.95 verbosity Not reported ### Reported statistics Uncertainty Standard error: ±0.65 pp sourcePopulation 95 stderr 0.646 tokenTotals Not reported ### Source [Original source](https://www.vals.ai/_astro/benchmark_view_medqa.Cco1kvog.json) Source Vals AI Original model label minimax/MiniMax-M2.1 Source version 1 Observed 2026-10-21T16:46:00.691Z Evidence SHA-256 `d67905978d932f64698ebe9e230d9519e5bae1c9d7176dda8379dc3874b6f90f` Source locator `/tasks/white/minimax~1MiniMax-M2.1` | | | | | | | | 6 | [GPT-4o (2024-08-06)](https://benchize.com/models/gpt-4o-2024-08-06-high.md)High | 88.5% | SE ±1.4 pp | $0.00279 | 3.23source duration | | | ### Original source values \| Metric \| Value \| Unit \| Scope \| Direction \| \| --- \| --- \| --- \| --- \| --- \| \| `accuracy` \| 88.5 \| percent \| reported \| Higher is better \| \| `cost_per_test` \| 0.002794 \| USD \| per_test \| Lower is better \| \| `latency` \| 3.231 \| source_duration \| reported_duration \| Lower is better \| ### Configuration computeEffort Not reported harness Not reported maxOutputTokens 16384 provider OpenAI reasoning Not reported Reasoning effort high Temperature 1 topP Not reported verbosity Not reported ### Reported statistics Uncertainty Standard error: ±1.4 pp sourcePopulation 95 stderr 1.399 tokenTotals Not reported ### Source [Original source](https://www.vals.ai/_astro/benchmark_view_medqa.Cco1kvog.json) Source Vals AI Original model label openai/gpt-4o-2024-08-06 Source version 1 Observed 2026-10-21T16:46:00.691Z Evidence SHA-256 `d67905978d932f64698ebe9e230d9519e5bae1c9d7176dda8379dc3874b6f90f` Source locator `/tasks/white/openai~1gpt-4o-2024-08-06` | | | | | | | | 7 | [Mistral Large 3 2512](https://benchize.com/models/mistral-large-2512.md) | 82.8% | SE ±0.84 pp | $0.000771 | 17.68source duration | | | ### Original source values \| Metric \| Value \| Unit \| Scope \| Direction \| \| --- \| --- \| --- \| --- \| --- \| \| `accuracy` \| 82.8 \| percent \| reported \| Higher is better \| \| `cost_per_test` \| 0.000771 \| USD \| per_test \| Lower is better \| \| `latency` \| 17.683 \| source_duration \| reported_duration \| Lower is better \| ### Configuration computeEffort Not reported harness Not reported maxOutputTokens 256000 provider Mistral AI reasoning Not reported Reasoning effort unknown Temperature 0.065 topP Not reported verbosity Not reported ### Reported statistics Uncertainty Standard error: ±0.84 pp sourcePopulation 95 stderr 0.844 tokenTotals Not reported ### Source [Original source](https://www.vals.ai/_astro/benchmark_view_medqa.Cco1kvog.json) Source Vals AI Original model label mistralai/mistral-large-2512 Source version 1 Observed 2026-10-21T16:46:00.691Z Evidence SHA-256 `d67905978d932f64698ebe9e230d9519e5bae1c9d7176dda8379dc3874b6f90f` Source locator `/tasks/white/mistralai~1mistral-large-2512` | | | | | | | | 8 | [GPT-4 Turbo](https://benchize.com/models/gpt-4-turbo-high.md)High | 81.1% | SE ±1.72 pp | $0.01 | 9.44source duration | | | ### Original source values \| Metric \| Value \| Unit \| Scope \| Direction \| \| --- \| --- \| --- \| --- \| --- \| \| `accuracy` \| 81.1 \| percent \| reported \| Higher is better \| \| `cost_per_test` \| 0.012734 \| USD \| per_test \| Lower is better \| \| `latency` \| 9.443 \| source_duration \| reported_duration \| Lower is better \| ### Configuration computeEffort Not reported harness Not reported maxOutputTokens 4096 provider OpenAI reasoning Not reported Reasoning effort high Temperature 1 topP Not reported verbosity Not reported ### Reported statistics Uncertainty Standard error: ±1.72 pp sourcePopulation 95 stderr 1.715 tokenTotals Not reported ### Source [Original source](https://www.vals.ai/_astro/benchmark_view_medqa.Cco1kvog.json) Source Vals AI Original model label openai/gpt-4-turbo Source version 1 Observed 2026-10-21T16:46:00.691Z Evidence SHA-256 `d67905978d932f64698ebe9e230d9519e5bae1c9d7176dda8379dc3874b6f90f` Source locator `/tasks/white/openai~1gpt-4-turbo` | | | | | | | | 9 | [Command A](https://benchize.com/models/command-a.md) | 80.65% | SE ±0.88 pp | $0.00347 | 2.95source duration | | | ### Original source values \| Metric \| Value \| Unit \| Scope \| Direction \| \| --- \| --- \| --- \| --- \| --- \| \| `accuracy` \| 80.65 \| percent \| reported \| Higher is better \| \| `cost_per_test` \| 0.00347 \| USD \| per_test \| Lower is better \| \| `latency` \| 2.947 \| source_duration \| reported_duration \| Lower is better \| ### Configuration computeEffort Not reported harness Not reported maxOutputTokens 8000 provider Cohere reasoning Not reported Reasoning effort unknown Temperature 0.3 topP Not reported verbosity Not reported ### Reported statistics Uncertainty Standard error: ±0.88 pp sourcePopulation 95 stderr 0.883 tokenTotals Not reported ### Source [Original source](https://www.vals.ai/_astro/benchmark_view_medqa.Cco1kvog.json) Source Vals AI Original model label cohere/command-a-03-2025 Source version 1 Observed 2026-10-21T16:46:00.691Z Evidence SHA-256 `d67905978d932f64698ebe9e230d9519e5bae1c9d7176dda8379dc3874b6f90f` Source locator `/tasks/white/cohere~1command-a-03-2025` | | | | | | | | 10 | [Gemini 2.5 Flash Lite Preview 09-2025](https://benchize.com/models/gemini-2-5-flash-lite-preview-09-2025.md) | 80.6% | SE ±0.88 pp | $0.000285 | 6source duration | | | ### Original source values \| Metric \| Value \| Unit \| Scope \| Direction \| \| --- \| --- \| --- \| --- \| --- \| \| `accuracy` \| 80.6 \| percent \| reported \| Higher is better \| \| `cost_per_test` \| 0.000285 \| USD \| per_test \| Lower is better \| \| `latency` \| 6.004 \| source_duration \| reported_duration \| Lower is better \| ### Configuration computeEffort Not reported harness Not reported maxOutputTokens 65535 provider Google reasoning Not reported Reasoning effort unknown Temperature 1 topP Not reported verbosity Not reported ### Reported statistics Uncertainty Standard error: ±0.88 pp sourcePopulation 95 stderr 0.884 tokenTotals Not reported ### Source [Original source](https://www.vals.ai/_astro/benchmark_view_medqa.Cco1kvog.json) Source Vals AI Original model label google/gemini-2.5-flash-lite-preview-09-2025 Source version 1 Observed 2026-10-21T16:46:00.691Z Evidence SHA-256 `d67905978d932f64698ebe9e230d9519e5bae1c9d7176dda8379dc3874b6f90f` Source locator `/tasks/white/google~1gemini-2.5-flash-lite-preview-09-2025` | | | | | | | | 11 | [GPT-4o-mini (2024-07-18)](https://benchize.com/models/gpt-4o-mini-2024-07-18-high.md)High | 72.35% | SE ±1.96 pp | $0.000157 | 1.78source duration | | | ### Original source values \| Metric \| Value \| Unit \| Scope \| Direction \| \| --- \| --- \| --- \| --- \| --- \| \| `accuracy` \| 72.35 \| percent \| reported \| Higher is better \| \| `cost_per_test` \| 0.000157 \| USD \| per_test \| Lower is better \| \| `latency` \| 1.776 \| source_duration \| reported_duration \| Lower is better \| ### Configuration computeEffort Not reported harness Not reported maxOutputTokens 16384 provider OpenAI reasoning Not reported Reasoning effort high Temperature 1 topP Not reported verbosity Not reported ### Reported statistics Uncertainty Standard error: ±1.96 pp sourcePopulation 95 stderr 1.959 tokenTotals Not reported ### Source [Original source](https://www.vals.ai/_astro/benchmark_view_medqa.Cco1kvog.json) Source Vals AI Original model label openai/gpt-4o-mini-2024-07-18 Source version 1 Observed 2026-10-21T16:46:00.691Z Evidence SHA-256 `d67905978d932f64698ebe9e230d9519e5bae1c9d7176dda8379dc3874b6f90f` Source locator `/tasks/white/openai~1gpt-4o-mini-2024-07-18` | | | | | | | | 12 | [GPT-3.5 Turbo](https://benchize.com/models/gpt-3-5-turbo-high.md)High | 58.1% | SE ±2.16 pp | $0.000426 | 2source duration | | | ### Original source values \| Metric \| Value \| Unit \| Scope \| Direction \| \| --- \| --- \| --- \| --- \| --- \| \| `accuracy` \| 58.1 \| percent \| reported \| Higher is better \| \| `cost_per_test` \| 0.000426 \| USD \| per_test \| Lower is better \| \| `latency` \| 2.002 \| source_duration \| reported_duration \| Lower is better \| ### Configuration computeEffort Not reported harness Not reported maxOutputTokens 4096 provider OpenAI reasoning Not reported Reasoning effort high Temperature 1 topP Not reported verbosity Not reported ### Reported statistics Uncertainty Standard error: ±2.16 pp sourcePopulation 95 stderr 2.16 tokenTotals Not reported ### Source [Original source](https://www.vals.ai/_astro/benchmark_view_medqa.Cco1kvog.json) Source Vals AI Original model label openai/gpt-3.5-turbo Source version 1 Observed 2026-10-21T16:46:00.691Z Evidence SHA-256 `d67905978d932f64698ebe9e230d9519e5bae1c9d7176dda8379dc3874b6f90f` Source locator `/tasks/white/openai~1gpt-3.5-turbo` | | | | | | | --- Cite as: MedQA / white — Benchize, https://benchize.com/benchmarks/vals-ai%3Amedqa%3Awhite (retrieved 2026-10-21)