GPT-6 for medical research: document workflows with Astra, Sol and Luna

29 September: GPT-6.1 Sol and complex PDF reading

OpenAI introduced GPT-6.1 Sol on 29 September 2026, reporting improvements in complex PDF understanding and scientific workflows. Those provider evaluations motivate a focused reading exercise; they do not establish clinical superiority or confirm GPT-6.1 Sol access in MedSafeAI. Check your current model selector.

Try a task where document layout changes the meaning: extract a results-table row together with its column headings, denominator and footnote. The worked example below shows what to check. It is an invented teaching example, not a published study or a result from testing a model.

Keep the denominator and footnote attached to the number

Imagine a table where 30 of 100 participants in group A and 27 of 90 in group B each had at least one event. A summary that only says “group B had 10% fewer events” compares counts and hides that both observed proportions are 30%. The same counts can mean something different if the table reports recurrent events or different follow-up periods. Read the heading and footnote before interpreting the row.

Use this prompt with one paper: “Transcribe the selected row and its column headings. For each arm, keep the event count, denominator, time point, analysis population and linked footnotes. Give the page and table number. Separate copied values from calculations. If a denominator, definition or footnote is unreadable, say so; do not infer it. Then list what must be checked before comparing the groups.”

Open the source page beside the answer. Confirm each value and follow every footnote marker; retain “not reported” where appropriate. Only calculate a proportion when the numerator and denominator describe the same population and time frame. A descriptive proportion alone does not show a treatment benefit.

In MedSafeAI, start with one table in a paper you already have. Ask All AI can expose different readings of that row; use Consensus to locate the disagreement, then resolve it from the PDF. Keep the checked row and its source location for your next literature discussion.

22 September update: Sol and Luna join the GPT-6 family

OpenAI introduced GPT-6 Sol and Luna on 22 September as faster, lower-cost companions to Astra. Its API changelog confirms text and image inputs and text output. These are provider announcements, not evidence of clinical superiority.

The new release does not by itself confirm Sol or Luna access inside MedSafeAI. Check the current model selector. The exercises below can use the models you already have; our proposed division of work is a starting point for comparison, not a measured ranking.

Split the document work into tasks you can judge

Keep the same documents and question when comparing models. Instead of asking each model to do everything at once, try these bounded tasks and record what you had to correct. The model choices in this table are trial candidates, subject to availability.

For the first task, try: “List the documents you can read. For each, give its title, year, source type and the section relevant to [question]. Mark missing information. Do not summarize findings yet.” Then use one of the three deeper document exercises below.

Record waiting time and your own verification/editing time separately. Keep a count of missing fields and substantive corrections. A quickly generated draft may still require more work overall. This is a small workflow comparison, not a clinical validation study.

Start with what OpenAI actually announced

OpenAI announced GPT-6 Astra on 3 September 2026. Its announcement emphasizes progress in browsing, scientific work and tasks that require several steps. Its model documentation lists research and document creation among intended uses, with text and image inputs. Those are provider descriptions. They suggest useful things to investigate, but do not establish how accurately the model will handle your particular medical question.

MedSafeAI includes GPT-6 Astra in its model offering. Use your workspace’s model selector to check current access. Start with one of the three exercises below, check the output against the original material, and use the comparison tools when you want another model’s perspective.

1. Turn a paper into an inspectable results table

Choose a paper you know well and make its main text and supplement available. Ask Astra to extract the population, comparison, main outcome, time point, reported estimate and limitations. Request the location of each entry. The useful output is a table you can check quickly against the paper, not a confident paragraph that hides where its details came from.

Try: ‘Using this paper and supplement, build a results table for [research question]. Preserve the reported units, denominators and uncertainty. Locate each factual entry by page, section or table. Separate author conclusions from your interpretation. Mark missing information. Finish with three questions for journal club; do not create a treatment recommendation.’

2. Compare two guideline versions without losing the qualifiers

Select two dated versions from the issuing organization and keep them together in AI Library. Limit the comparison to one population and topic. Ask for old wording, new wording, source locations and any stated change in strength or certainty. Give the model an explicit ‘not located’ category so absent text is not automatically treated as a removed recommendation.

Open the passages behind the largest changes. Check exceptions, settings and eligibility conditions before writing your own summary. A different heading or page number can make an unchanged recommendation appear new; source reading resolves that ambiguity. Keep local implementation decisions outside this document-comparison task.

3. Build a short research brief with an unresolved-questions section

Use Evidence Search to gather material for a bounded question, then select the sources that belong in the brief. Ask Astra to organize them by the question they address, the population studied and the limitation that matters most. With Strict Mode, keep the workflow focused on citations and inspect the sources behind important statements. Request a separate list of gaps rather than a conclusion that fills them in.

For a useful comparison, send the same question to the other models through Ask All AI. Consensus can bring together agreements and critical differences. Investigate a disagreement by opening the relevant individual responses and source passages. Several models reaching the same conclusion is a reason to inspect the reasoning, not a clinical validation result.

Judge the corrected output, not the launch headline

Record the model, date, prompt, source set and corrections you made. Did it retain the important conditions? Were the cited passages actually relevant? How much editing remained? Repeat the same task with unchanged inputs when the difference matters to your choice. Keep OpenAI’s official announcement and safety overview alongside your notes, and distinguish their evaluation claims from your own observations. Start with one document you can verify before expanding the work.