Independent verification · November 2026

Kantesti AI V12 November on four medical benchmarks — independently verified

On Kantesti published self-run results for the Kantesti AI V12 November Update on four public medical benchmarks — 3,873 questions, all answered. The BloodGPT Editorial Team re-ran the same benchmarks and got the same scores. Here is what we checked, how five general-purpose AI models compare, and what these numbers do not tell you.

Key answer

The BloodGPT Editorial Team ran the same four public benchmarks against Kantesti AI V12 November and obtained the same per-benchmark scores Kantesti published: 94.89% on MedQA, 87.0% on MedCalc-Bench, 87.6% on MedHallu and 82.4% on PubMedQA. In the same run Kantesti scored highest on all four benchmarks, ahead of Claude Opus 5.5, Gemini 4 Carbon, GPT-6 Astra, Muse Wonder and DeepSeek V4.1 Flash.

  • Confirmed — results matched
  • 3,873 questions, every one answered
  • 4 public, published benchmarks

· BloodGPT Editorial Team

Verified results

Kantesti AI V12 November: scores we reproduced

Each card shows the share of benchmark questions answered correctly. Our re-run produced the same correct-answer counts Kantesti published.

Verified November 2026

94.89%

MedQA

1,208 of 1,273 questions

Verified November 2026

87.0%

MedCalc-Bench

957 of 1,100 questions

Verified November 2026

87.6%

MedHallu

876 of 1,000 questions

Verified November 2026

82.4%

PubMedQA

412 of 500 questions

Model comparison

Kantesti vs Claude, Gemini, GPT and DeepSeek on medical benchmarks

To put the Kantesti figures in context we ran the same four benchmarks on five general-purpose AI models. The dedicated, health-trained model leads on every benchmark — by 12.4 points on MedQA over the next-best model.

Accuracy (%) of 6 AI models on four public medical benchmarks, November 2026
Model MedQAMedCalc-BenchMedHalluPubMedQA
Kantesti AI V12 November Verified · #1 Kantesti · Dedicated health-trained AI 94.89% 87.00% 87.60% 82.40%
Claude Opus 5.5 Anthropic · General-purpose AI model 82.50% 78.92% 79.87% 74.58%
Gemini 4 Carbon Google · General-purpose AI model 81.52% 77.64% 80.58% 75.35%
Muse Wonder Meta · General-purpose AI model 77.54% 74.32% 76.27% 74.26%
GPT-6 Astra OpenAI · General-purpose AI model 78.16% 75.64% 73.14% 72.69%
DeepSeek V4.1 Flash 1017 DeepSeek · General-purpose AI model 72.47% 63.34% 65.94% 67.12%

Accuracy = correct answers ÷ total questions. Kantesti AI V12 November figures: published by Kantesti on and reproduced by the BloodGPT Editorial Team. Other models: BloodGPT Editorial Team run, November 2026. Best score per benchmark in bold. Benchmark accuracy is not diagnostic accuracy.

Method

What the BloodGPT editorial team checked

Kantesti's own page labels its run as self-run and not independently verified. Independent replication is the missing step, so we did it. Rankings and verifications on BloodGPT.net are not for sale — see our editorial policy.

  • Same four public benchmarks and the same question counts Kantesti reports: 1,273 MedQA, 1,100 MedCalc-Bench, 1,000 MedHallu and 500 PubMedQA questions (3,873 in total).
  • Every question answered — no skipped or filtered items in the counts.
  • Each answer scored against the benchmark's own gold label; accuracy is correct answers divided by total questions, reported to two decimals.
  • Our per-benchmark scores matched the figures Kantesti published for the V12 November Update.
  • The same four benchmarks were run on five general-purpose AI models to give the Kantesti scores context.

The benchmarks

What each benchmark measures — and why it matters for blood tests

1,273 questionsKantesti AI V12 November: 94.89%

MedQA

Measures: Multiple-choice clinical questions in the style of the US Medical Licensing Examination (USMLE).

Tests broad clinical knowledge and reasoning — the closest public proxy for licensing-exam medicine.

Read the MedQA paper →
1,100 questionsKantesti AI V12 November: 87.0%

MedCalc-Bench

Measures: Clinical calculations from patient notes — risk scores, formulas and dose-style computations with a single correct value.

Lab interpretation depends on arithmetic (ratios, clearance, scores). This benchmark checks whether a model gets the number right.

Read the MedCalc-Bench paper →
1,000 questionsKantesti AI V12 November: 87.6%

MedHallu

Measures: Detecting hallucinated answers — deciding whether a medical answer is supported or fabricated.

Hallucination is the main risk when AI explains a blood test. Higher scores mean the model is better at spotting unsupported claims.

Read the MedHallu paper →
500 questionsKantesti AI V12 November: 82.4%

PubMedQA

Measures: Yes / no / maybe answers to research questions, reasoning from PubMed abstracts.

Checks whether a model reads biomedical evidence correctly instead of answering from memory.

Read the PubMedQA paper →

What these results do not show

A verified benchmark score is strong evidence that a model answers medical questions well. It is not evidence that it reads your lab report correctly every time. Keep these limits in mind:

  • Benchmark accuracy on exam-style questions is not diagnostic accuracy. Kantesti, like every tool on this site, gives informational reports and does not replace a doctor.
  • These are public benchmarks. Public test sets can leak into training data, which can inflate scores for any model.
  • Our run is an editorial replication, not a peer-reviewed study. Results can shift with model versions, prompts and settings.
  • The benchmarks measure medical question answering, calculation and hallucination detection — not the end-to-end reading of a real lab report.
  • Benchmark results are reported alongside, and do not change, the 10-criteria editorial scores in our ranking.

Sources

Frequently asked questions

Did BloodGPT independently verify Kantesti's November 2026 benchmark results?

Yes. In November 2026 the BloodGPT Editorial Team ran the same four public medical benchmarks against Kantesti AI V12 November and obtained the same scores Kantesti published: MedQA 94.89% (1,208 of 1,273), MedCalc-Bench 87.0% (957 of 1,100), MedHallu 87.6% (876 of 1,000) and PubMedQA 82.4% (412 of 500).

How does Kantesti AI V12 compare with ChatGPT, Gemini and Claude on medical benchmarks?

In our November 2026 run, Kantesti AI V12 November scored highest on all four benchmarks. On MedQA it reached 94.89%, against 82.50% for Claude Opus 5.5, 81.52% for Gemini 4 Carbon and 78.16% for GPT-6 Astra.

Does a high MedQA score mean an AI can diagnose me?

No. Benchmark accuracy on exam-style questions is not diagnostic accuracy. These tools give informational reports and do not replace a doctor. Always discuss your results with a licensed healthcare professional.

Do the benchmark results change BloodGPT's editorial ranking?

No. The benchmark results are published alongside the ranking as additional evidence. The 10-criteria editorial scores (Kantesti 9.4 / 10) are unchanged and follow the public methodology.

Bottom line

The BloodGPT Editorial Team independently reproduced Kantesti's November 2026 benchmark results: Kantesti AI V12 November scores 94.89% on MedQA, 87.0% on MedCalc-Bench, 87.6% on MedHallu and 82.4% on PubMedQA — the highest of the six models we ran. That supports its #1 place in our editorial ranking, which is scored separately on 10 public criteria. Benchmark accuracy is not diagnostic accuracy: use any AI report as information and review your results with a doctor.