188 Net Points: Bianca Ciobanu Breaks the MIR Record — But AI Already Reached 200

On 1 March 2026, the Ministry of Health published the provisional results of the MIR 2026. While 15,283 candidates checked their scores, at MedBench we had known for weeks a figure that puts everything in perspective: three AI models had already answered all 200 valid exam questions without a single mistake.^[1]

The human record goes to Bianca Ciobanu Selaru with 188 net points — the highest score in MIR history. An extraordinary milestone. But one that arrives at the very moment AI operates in a region no human has ever reached: fifteen models surpass 194 net points, and three achieve absolute perfection at 200/200.

This article connects both worlds: the exceptional human effort and the artificial performance that no longer has a ceiling in this format.

1. The Record: Bianca Ciobanu and 188 Net Points

Bianca Ciobanu Selaru is 41 years old, of Romanian origin, and started medical school at 34. She combined her studies with work. On 1 February 2026, she sat the MIR exam and scored 188 net points, equivalent to 119.3784 points on the official scale (exam score plus academic record).^[2]

The previous record was held by Noelia García with 186.67 net points in the MIR 2024 (187 correct, 1 error, 12 blank).^[3] Bianca surpasses that record by more than one net point — a notable leap in an exam where every tenth of a point at the top counts.

Bianca's story is, in itself, an argument against stereotypes: age is not a limit, origin is not a limit, life experience adds up. What she did on 1 February required years of sustained preparation and a perfect day of execution.

2. Overview of the Examination

Key figures from the MIR 2026, according to the provisional listings from the Ministry of Health:^[2]

3. The Gap: Three Years of AI vs. Humans

The trend across three editions is clear:

• MIR 2024: AI surpasses the best human by 7 net points. It is the benchmark's first edition, and there is already an artificial advantage from the start.
• MIR 2025: The gap soars to 25 net points. The best human result was lower (165.67 netas), while AI rose (190.67). The exam was considered "harder", and AI was less affected by the difficulty than humans.
• MIR 2026: The gap narrows to 10.67 net points, but for a nuanced reason: the MIR 2026 was perceived as "more accessible" (the human median rose to 102), which allowed Bianca to reach 188. AI, for its part, rose 8 absolute net points to 198.67.

The conclusion is not that the gap "is closing" — it is that AI always wins, and the distance depends more on exam difficulty (which affects humans more) than on model advances.

For full context on AI performance in the MIR, see "199 out of 200: AI Only Misses One on the MIR 2026" and "Two Weeks Later: 22 New Models and Triple 200/200".

4. The Bell Curve and the Machines

This chart places the 15,283 MIR 2026 candidates on their estimated distribution (Gaussian with mean 97.35 and σ≈30). What it reveals is the magnitude of Bianca's achievement — and the magnitude of the gap with AI:

• The human mean (97.35 net points) and the median (102 net points) sit in the central zone of the bell curve, where most candidates cluster.
• Bianca Ciobanu (188 net points) is more than 3 standard deviations above the mean. She is an exceptional outlier: statistically, only 1 in roughly 700 candidates would reach that zone in a normal distribution. In practice, only one person has ever achieved it in MIR history.
• AI operates outside the human bell curve. The purple line (198.67 net points, standard top) and the teal line (200 net points, ALMA/MIRI) are in a region where human density is essentially zero.

Put another way: Bianca's 188 net points are a result we probably will not see repeated for years. And yet they would not even make the top 20 of AI models evaluated in MedBench.

5. The Extended Ranking: 15 Models vs. the Best Human

The visual impact is immediate: every AI bar surpasses the red line of Bianca Ciobanu (188 net points). And the human median (102 net points, amber line) sits far to the left, giving real scale to the comparison.

Key highlights:

• Three models with 200/200: ALMA, MIRI and Gemini 3.1 Pro Preview — perfect score, zero errors.
• Three tied at 198.67: Gemini 3 Flash, o3 and GPT-5 — one single mistake each.
• Provider diversity: Google, OpenAI, Anthropic and custom models. It is not the dominance of a single laboratory; it is a systemic capability of current technology.
• The cheapest model in the top 15 (Gemini 3 Flash, €0.34) surpasses the all-time human record by more than 10 net points.

The full cost and efficiency analysis is in "The Swiss Army Knife and the Scalpel". For the open-weights vs proprietary comparison, see "The Cathedral and the Bazaar".

6. Controversies of the Examination

No MIR is free from controversy, and the 2026 edition has accumulated several worth mentioning:

Academic record and appeals: More than 2,900 candidates appear in the provisional listings with a score of 5 points (the minimum default), an unusually high figure. According to specialised sources, between 1,000 and 1,500 candidates had filed formal appeals to the Ministry of Health over discrepancies in the academic record calculation. An error in the record can displace a candidate hundreds of places in the ranking, making this far more than an administrative issue.^[4]

Listing transparency: Unlike previous editions, the provisional MIR 2026 listings do not include the breakdown of correct answers, errors and blank questions per candidate — only the final net score. This has drawn criticism in candidate forums such as MIRentrelazados, where it is noted that without this breakdown it is impossible to individually verify that the score is correct.^[5]

Exam difficulty: The median of 102 net points in 2026 (compared to 102.15 in 2025) suggests a similar or slightly lower difficulty level than the previous year. The human record of 188 net points points in the same direction: an exam that allowed the best-prepared candidates to reach higher scores. Some candidates note that the topic distribution favoured those with broad preparation, although this is hard to confirm without granular official data.

Candidate-to-place ratio: With 15,283 candidates for 9,276 places, approximately 6,000 candidates are left without a place. The addition of the new Emergency Medicine specialty (86 places) slightly expands the offering, but does not offset the growth in candidates.

7. What the MIR Measures — and What It Does Not

The MIR is an exam of 200 multiple-choice questions on factual medical knowledge. It is a selection instrument designed to rank thousands of candidates, and it fulfils that function. But it is a narrow slice of what practising medicine means.

What the MIR does not measure:

• Clinical empathy and patient communication
• Physical examination and procedural skills
• Managing uncertainty — the MIR has one correct answer; real clinical practice often does not
• Multidisciplinary teamwork
• Ethical judgement in real dilemmas
• Life experience — precisely what distinguishes Bianca's trajectory

AI has demonstrated that, for knowledge retrieval and pattern matching at scale — exactly what a multiple-choice exam measures —, artificial performance surpasses human performance consistently and increasingly. This is not surprising: it is the type of task for which language models are optimised.

But Bianca Ciobanu's story — 41 years old, work experience, second country, career started at 34 — is a reminder that medicine is far more than answering questions correctly. The MIR opens the door to specialist training; what comes after is what defines the doctor.

8. Conclusions

1. 188 net points = all-time human milestone. Bianca Ciobanu Selaru's result is the best ever recorded in the MIR. It is an achievement of exceptional preparation that probably will not be repeated in the short term.

2. AI has already reached 200/200. Three models (ALMA, MIRI, Gemini 3.1 Pro) solved the complete exam without errors. The ceiling of the MIR format has been reached for artificial intelligence.

3. The gap is consistent across three years. In all three benchmark editions (MIR 2024, 2025, 2026), AI has surpassed the best human. It is not a one-off peak; it is a consolidated trend.

4. The process controversies deserve attention. The problems with the academic record (thousands of appeals) and the lack of transparency in the listings (no breakdown of correct/incorrect answers) are issues that affect trust in the system and merit institutional reform.

5. The comparison has limits. The MIR is a multiple-choice exam. It does not capture empathy, physical examination, uncertainty management or clinical judgement. AI dominates the format; medicine is far more than the format.

For the full technical analysis of AI in the MIR 2026, here is the benchmark article series:

• "MIR 2026: The Perfect Storm" — Context and benchmark design
• "199 out of 200: AI Only Misses One on the MIR 2026" — First results analysis
• "The Swiss Army Knife and the Scalpel" — Costs, efficiency and specialisation
• "The Cathedral and the Bazaar" — Open-weights vs proprietary
• "ALMA and MIRI: Agentic RAG" — Custom models with 200/200
• "Two Weeks Later" — Update with 22 new models