A million-token context window makes comprehensive chart review technically more plausible than it has ever been. It does not, on its own, make chart review safe, because an entire medical record is not merely a large document, it is temporally inconsistent, duplicated, partially structured and clinically incomplete, and no amount of context capacity resolves those properties by itself.
Three separate things that should not be confused
GPT-6 Astra: a foundation model, the underlying reasoning and context capability. ChatGPT for Healthcare: a governed enterprise product built for hospital deployment, with its own compliance framework, business associate agreements and institutional controls. And the Epic plugin: a read-only connection to authorised EHR information, retrieving what a clinician already has permission to see. These are three distinct layers, and conflating them, assuming Astra's model capability automatically means the Epic connection now works differently, or that ChatGPT for Healthcare's enterprise status implies compliance for any use of the underlying model, is exactly the error this article exists to prevent.
What a one-million-token context window changes
More encounters, letters, results and medication histories can fit within a single task than any previous model generation could hold at once. Less reliance on aggressive pre-summarisation, since earlier context limits forced systems to compress or discard material before a task even began, a compression step that itself introduces error. And better ability to revisit earlier material within the same task, checking a detail against something read much earlier without losing it to a context window that has since moved on.
A practical pre-consultation example
Consider a patient with genuine multimorbidity, seen across several specialties, with conflicting medication lists between primary and secondary care, outstanding investigations ordered weeks ago and never chased, and a recent admission with follow-up requirements nobody has yet actioned. A system capable of holding this entire picture within one task, rather than working from a necessarily incomplete summary, can in principle surface exactly the kind of cross-cutting pattern, the medication discrepancy, the unchased result, the unmet follow-up requirement, that a time-pressured clinician reviewing fragments in sequence might reasonably miss.
What the Epic connection currently does
Retrieves information the clinician already holds authorised permission to access. Preserves existing Epic permissions exactly as they stand, rather than expanding what any individual account can see. Points back to chart evidence, so a generated summary's claims can be checked against the specific source entry. And supports preparation and synthesis, the pre-consultation and evidence-gathering work this cluster's broader coverage treats throughout.
What it does not currently establish
Automatic record writing, the connection is explicitly read-only. Medication or test ordering, no order-entry capability exists through this connection. Autonomous care-plan changes, nothing in the current integration authorises the system to alter a treatment plan. Universal access for individual ChatGPT users, the Epic connection belongs to the enterprise ChatGPT for Healthcare product specifically, not to individual consumer or clinician accounts. And, stated as its own important editorial point, proof that every workflow within ChatGPT for Healthcare or the Epic connection is already Astra-powered, which OpenAI has not confirmed.
Important editorial box
OpenAI has not stated that GPT-6 Astra powers the recently announced Epic integration. The two announcements may be discussed together as a direction of travel, since both represent OpenAI's broader healthcare ambition, and they should not be represented as a confirmed technical architecture until OpenAI states so directly.
Failure modes of enormous context
Relevant signal buried within duplication, where the genuinely important entry sits among dozens of near-identical repeated notes. Incorrect chronology, where the sequence of events is misread despite every individual date being technically present in the context. Superseded diagnoses treated as current, an earlier working diagnosis persisting in a summary after later entries have revised or excluded it. Copied-forward errors, where an inaccurate entry, propagated across multiple notes through routine copy-forward documentation practice, gets amplified rather than caught, the specific risk this cluster's dedicated coverage of copy-forward amplification treats in depth. Missing external records, since even a million tokens of available context cannot include information the system was never given access to in the first place. Confusion between intended and administered medicines, a genuinely dangerous specific failure mode where a prescribed but not yet dispensed, or discontinued but still-listed, medicine is treated as currently being taken. And misleading instructions embedded within retrieved documents, where content inside a retrieved file could itself contain confusing or contradictory guidance the system needs to recognise as such rather than follow uncritically.
Why provenance matters
Every summary statement should point to dates and source documents, so a clinician can verify rather than simply trust. Contradictions should be surfaced rather than silently resolved, since a system that quietly picks one version of a conflicting record has made a clinical judgement it was not authorised to make. Uncertainty must be explicit, stated plainly rather than smoothed over by fluent prose. And high-impact conclusions require confirmation, meaning any summary point that would materially change a clinical decision deserves independent verification against the primary source before it is acted upon.
The economics of long-context chart review
Larger contexts increase cost and latency, a genuinely practical consideration for any system processing entire records routinely rather than selectively. Retrieval and model routing, pulling only the relevant portion of a record for a specific task rather than processing the entire chart every time, may remain more efficient than brute-force full-context review for many real workflows, an architectural choice worth understanding as a genuine trade-off rather than assuming maximum context use is always the right design.
What this means for the NHS
No equivalent EMIS or SystmOne integration has been announced, meaning the majority of NHS primary care, which runs on these two systems rather than Epic, is entirely outside the scope of anything described in this article. UK deployment of any comparable capability would require local information-governance and clinical-safety evaluation, DCB 0129 among the frameworks this cluster's broader coverage treats throughout. And US public-data connections do not solve UK guideline localisation, a genuinely separate problem no amount of American regulatory and payer data resolves.
Verdict
Entire-record review is becoming technically plausible. Safe entire-record understanding remains a system-design challenge, provenance, contradiction handling, uncertainty communication, human confirmation of high-impact conclusions, rather than merely a context-window problem that a bigger number alone resolves.
Frequently asked questions
Can GPT-6 Astra write to a patient's Epic record?
No: the current Epic connection is explicitly read-only, unable to amend records, place orders or send patient messages, regardless of which underlying model powers the connection.
Does a bigger context window mean fewer chart-review errors?
Not automatically: a larger context window changes what is technically possible to include, and it does not by itself solve the duplication, chronology and copy-forward-error problems that make medical records genuinely difficult to synthesise safely.
Is this technology relevant to NHS primary care today?
Not directly: no EMIS or SystmOne integration has been announced, meaning the majority of NHS general practice sits entirely outside what this specific Epic-based capability currently addresses.