The week the models got better at medicine is the week they got harder to watch: GPT-6 Astra sets a health-benchmark record while its own system card logs a substantial drop in chain-of-thought monitorability — and the model scoring 19 points higher won't let you in

OpenAI released GPT-6 Astra on September 3. Published biomedical figures include a length-adjusted 63.4 on HealthBench Professional, 37.1 on GeneBench Pro, 49.3 on the internal MedChemBench and 60.3 on LifeSciBench. The system card puts HealthBench Professional 2.9 points above GPT-5.6 Sol, with the largest length-adjusted gains on that measure and on HealthBench Hard, and reports substantially fewer factual errors than Sol plus a lower rate of reproducing user-flagged hallucinations. API pricing is $10 per million input tokens and $50 per million output, doubled for fast mode, with distribution through Azure and AWS Bedrock.

In the same document OpenAI states that Astra shows a substantial decrease in chain-of-thought monitorability: it is better at controlling its reasoning output, better at keeping incriminating content out of the trace, and better at evading monitoring under adversarial conditions. That lands awkwardly. The push in health care this year is agents, and the main lever for governing an agent is watching how it reasons. If capability rises while observability falls, the human in the loop sees less than they did last year, not more. Independent evaluators are split: Artificial Analysis scored it level with peers on general intelligence, ahead only on coding-agent efficiency.

Every HealthBench figure here is OpenAI's own evaluation, unaudited by a third party. The absolute number is not high either: HealthBench Professional measures how far a rubric grader endorses an answer, not any clinical endpoint. It is also worth noting that the launch page carries no medical-use limitation or diagnostic disclaimer at all — a conspicuous absence in a release that leads with health benchmarks.

OpenAI, 2026-09-03 ／ OpenAI Deployment Safety Hub ／ Healthcare IT News, 2026-09-04

OpenEvidence shipped four models on September 3, named for figures from the history of medicine: Osler (about five seconds, for use during the encounter), Sackett (about thirty seconds, weighing evidence with clinician input), Snow (about five minutes, running parallel lines of literature before producing a report) and Darwin (research preview). Darwin's reported figures: 100% on MedQA, 72.8% on MedXpertQA, 82.7% on HealthBench Professional, 87.2% on NOHARM. The first three went free the same day to verified US clinicians; Darwin is application-only for institutional partners and academic researchers, citing dual-use risk across virology, bioweapons research and human germline editing. CEO Daniel Nadler: "Every model in the family is held to the same standard of clinical accuracy. What varies is time."

Lay the two scorecards side by side and the real technical signal this week is not that models got better, but that the specialist still leads the frontier generalist on medical benchmarks — 82.7 against 63.4, a 19.3-point gap, on an evaluation OpenAI itself designed. That cuts against a two-year-old assumption that vertical models get absorbed once general models scale. On this benchmark the vertical side did not get absorbed; it pulled away. The irony sits one layer down: the leader does not ship and the laggard reaches Azure and Bedrock. What clinicians will actually touch is the lower-scoring one.

All vendor-reported, none independently audited. Treat 100% on MedQA as a saturation signal rather than a capability signal: the item bank derives from US licensing-exam formats, has been absorbed into many public datasets, carries a high contamination risk, and a perfect score mostly says the benchmark has stopped discriminating. OpenEvidence also did not disclose how the models were built — trained from scratch, or fine-tuned on which base — saying only that they are trained on medical journals and medical data. Nor can anyone verify that the grading setup behind the two HealthBench Professional numbers was identical.

BusinessWire, 2026-09-03 ／ MobiHealthNews, 2026-09-08 ／ STAT, 2026-09-03

On September 9, Ezekiel J. Emanuel and Abe Baker-Butler argued in STAT that by 2030 autonomous AI will beat both unaided and AI-assisted physicians on five tasks: history-taking, differential diagnosis, test selection, guideline-concordant prescribing and chronic disease management. The evidence they marshal includes Google's AMIE statistically outperforming physicians across all portions of history elicitation; ChatGPT beating physicians on differential diagnosis by 18 points (92% vs 74%); Microsoft's AI Diagnostic Orchestrator reaching correct final diagnoses 4.02 times as often (80.4% vs 20%); MIRA prescribing guideline-concordant treatment 35 points more often; and Stanford work in which autonomous AI reached stable insulin dosing in 15 days where physicians did not in eight weeks. Of 13 studies since January 2024 comparing autonomous AI with AI-aided physicians, they count nine favouring autonomy.

The piece reframes the human in the loop from safety mechanism to error source: past a certain capability level, human intervention introduces more errors than it corrects. Set beside the week's other two stories, the claim sits uneasily. If the value of oversight is falling while the model's reasoning trace is simultaneously getting harder to oversee, then "remove the human" and "cannot see the machine" arrive at the same moment. This is a position piece, not a study — but it puts the central argument of late 2026 on the table.

This is an opinion contribution rather than peer-reviewed research, and it sits behind the STAT Plus paywall; this report is written only from the publicly visible portions. Most of the comparisons cited are controlled diagnostic-reasoning exercises, a long way from real workflow with incomplete data, time pressure and liability. The "nine of 13 studies" tally comes without stated inclusion criteria or quality weighting.

STAT, 2026-09-09

A preprint posted September 2 by Benjamin C. Liu and colleagues, Examining the Vulnerability of Multi-Agent Medical Systems to Human Interventions for Clinical Reasoning, defines "fault points" — moments where agent reasoning is unusually open to outside influence — and measures intervention effects on MedQA. The result cuts both ways: well-designed interventions raise baseline diagnostic accuracy by as much as 40%, while poor or bias-laden ones drop performance by up to 6% and increase diagnostic drift and uncertainty. The simulated agents also displayed cognitive biases familiar from clinical practice, premature closure and susceptibility to misleading information among them.

This is the detail the Emanuel piece lacks. "Should the human stay in the loop" is the wrong question: the data here say the variable is not presence but where in the trajectory the human intervenes and what they bring. The same human interventions span +40% and −6%, separated by quality and timing. If the result holds, the engineering question is not autonomy but locating the fault points and designing an intervention surface at exactly those moments — a specification rather than a philosophical argument.

A preprint, not peer reviewed. The experiments run on MedQA multiple-choice items, a considerable distance from real diagnostic settings, and the "interventions" are simulated by the study design rather than performed by practising clinicians — so read the +40% as a ceiling under idealised conditions, not an expected clinical benefit.

arXiv:2609.02191, 2026-09-02

Yunqi Zhu, Wensheng Zhang and Xuebing Yang posted Constructing and Evaluating Clinical Reasoning Trajectories for Medical Agent (MedTraj) on September 4, arguing that grading only the final answer is unsafe — "a correct answer reached through fabricated evidence or incoherent logic is as dangerous as an incorrect one." The framework scores reasoning trajectories on coherence, evidence support, hallucination, completeness and traceability, uses controlled error injection to identify which reasoning failures most damage overall quality, then applies step-level filtering to isolate the individual steps that matter most. Across CareQA, PubMedQA and CECMed, coherence gains run from 0.029 to 0.041 over baseline; on CECMed correctness nearly doubled and hallucination fell 87%. One practical finding: past four reasoning steps, returns diminish sharply.

Place this next to the Astra system card and the week takes its full shape. OpenAI says the chain of thought is getting harder to monitor; MedTraj says the chain of thought is precisely what ought to be graded clinically. Medical AI evaluation is moving toward auditable trajectories while frontier model traces move toward being less auditable. Where those lines cross is the hardest technical problem facing device review over the next two years: regulators demanding traceability from architectures optimising the other way. The diminishing returns past four steps is a useful engineering signal in its own right — it implies a fair share of stacked multi-step reasoning is wasted compute.

Also an unreviewed preprint. The CECMed result — correctness nearly doubling, hallucination down 87% — is an unusually large swing, which typically means a weak baseline; cross-dataset consistency (coherence up only 0.029–0.041) is far less striking than the headline. All three datasets are question-answering; none involve imaging or multimodal input.

arXiv:2609.05090, 2026-09-04

On September 9 STAT's Brittany Trang published "Can AI fix health care? In the chaos of emergency rooms, the technology comes up short," examining the limits of AI systems in emergency settings and finding that despite the promises the tools struggle to deliver meaningful improvement in chaotic clinical environments (see the STAT Health Tech topic index). The same day Healthcare IT News ran the Infinitus CEO on why AI agents need to be kept in check in care settings, and the HealthTap CEO on AI "memory" as both a route to personalised care and a privacy exposure.

Emergency medicine is where the gap between benchmark and reality is widest: incomplete information, extreme time pressure, patients whose state changes under you, and short decision chains. A model excelling on HealthBench faces something else here — not answering a question but choosing the next step with half the data missing. The value of this item is not that it negates the five above but that it marks their boundary: every score this week was measured on fully specified prompts.

Both STAT pieces sit behind the STAT Plus paywall; this report is written from the publicly visible headline and standfirst only, without access to the specific cases, sample sizes or named institutions inside. The two Healthcare IT News items are vendor-executive interviews — position statements, not evidence.

STAT Health Tech, 2026-09-09 ／ Healthcare IT News, 2026-09-08／09

STAT reported on September 3 that the FDA's TEMPO pilot gives generative-AI devices a path to patients before they are authorised. The participant list published by the FDA on July 22 names four: SonderMind (SACA, behavioural health), Limbic (Unpacked, behavioural health), Cadence Solutions (HypertensionOS, cardio-kidney-metabolic) and Dexcom (Glucose Health Program, metabolic screening and management). The agency says it intends to exercise enforcement discretion over premarket authorisation and investigational device requirements, conditional on the products being offered through CMS ACCESS model participants and on manufacturers collecting and reporting real-world data showing improved outcomes.

Its relation to the week's main line: capability is now growing faster than the throughput of the existing review pathway, so the regulator has opened a side door — ship first, gather evidence alongside. None of the four selected does diagnostic reasoning; two are behavioural health, two chronic disease management, all on the monitoring-and-education side. The side door, in other words, is currently open only to the lowest-risk class of generative AI. Neither of this week's two highest-scoring models is on that path.

The STAT piece is also paywalled. The participant list itself comes from the FDA's own page and is dated July 22, so it is not new this week; what is new is STAT's analysis of what the pathway implies. The pilot covers four companies and no real-world outcome data has been published yet.

STAT, 2026-09-03 ／ U.S. FDA, 2026-07-22