Claude Opus 4.7 is a real upgrade for med school. The chat session is not the system.

Four things. First, vision resolution rose from 1.15 to 3.75 megapixels, which is the difference between a labeled anatomy figure being unreadable and being parseable, and is the single biggest unlock for slide decks built around microscopy and histology images. Second, the new xhigh effort level produces sharper mechanism explanations on the kind of multi-step pathway questions you see in biochemistry and physiology. Third, the model is more honest about uncertainty in clinical-scenario items, which means fewer confidently wrong distractors. Fourth, the new tokenizer uses roughly 1.0 to 1.35 times as many tokens per page, so a long lecture deck that fit in 4.6 may need to be chunked. The pricing is unchanged at $5 per million input tokens and $25 per million output tokens.

Yes for born-digital PDFs, conditionally yes for scans. A born-digital deck (exported from PowerPoint or Keynote) parses cleanly: the model reads the text layer and the slide images at the new 3.75MP cap. Image-only scans go through Anthropic's vision pipeline as well, which now resolves small caption text on histology and microscopy slides that 4.6 routinely missed. The remaining failure mode is photocopied-three-times handout scans where the OCR is hostile and the labels are ghosted; for those, run them through a separate OCR pass first, or upload the original PowerPoint file if you have access to it.

Three reasons, all about the system around the model rather than the model itself. First, no rubric is enforced at output time, so factual correctness, distractor quality, type coverage, and clarity vary across the 200 questions in a single response. Second, no stem rewording on revisit; by the time you have answered the same MCQ twice, you are memorizing the wording, not the slide. Third, no spaced repetition, so the questions you got wrong are not surfaced more often than the questions you got right. On the held-out three-document eval, Studyly scored 81.3 versus Turbolearn 57.8 and the field average 67.9, where Turbolearn and the field are roughly the level of a generic LLM-quiz wrapper. The model is the generator. The retention is the system.

Three honest options. First, the Claude Pro consumer subscription at $20 a month gives you generous usage in claude.ai, which is enough for a few decks a week if you avoid long agent sessions. Second, the API at $5 input / $25 output per million tokens is metered; a 90-slide PDF at the new tokenizer rate costs maybe $0.20 to $0.60 input and $0.30 to $1.00 output for a full 200-question generation, so roughly $0.50 to $1.50 per deck end to end. Third, an agentic wrapper like Studyly's free tier on app.jungleai.com lets you drop in real lecture decks at no cost (no credit card) to test the workflow before you decide whether the rubric and the spaced repetition are worth the paid tier.

Studyly does not publish the model behind the question pipeline, and the eval on /quality is model-agnostic by design: the 81.3 score is a property of the generation system and the rubric, not of any one frontier model. Practically, this means the eval result holds even when the underlying model rotates between releases (4.6, 4.7, and whatever ships next), because the rubric is the gate. If you want to run Claude Opus 4.7 directly on your own decks with full transparency, the prompt template above is the starting point. If you want the rubric, the auto-rephrasing, the slide-number citations, and the spaced repetition to come built in, that is what app.jungleai.com adds.

Build the slide-number citation into the prompt and do not accept any question that does not include it. The prompt template above adds an explanation_quote and explanation_quote_slide field; any output without those gets thrown out before you study it. Spot-check the first 20 questions against the deck. If the citation slide actually contains the explanation_quote string, the question is grounded. If it does not, the model invented the fact, and you discard it. In practice, with the rubric encoded in the prompt and xhigh effort, the rate of ungrounded questions is low single digits on a clean PDF deck, and is higher (roughly 8 to 12 percent) on heavily abbreviated bullet decks where the model has to expand the bullets into stems.

You can, but it is not the highest-leverage use of Claude Opus 4.7. The Step exams are item-writer-vetted by NBME, and UWorld and AMBOSS already publish first-exposure question banks at a level no LLM consistently matches on clinical-vignette stems. Where 4.7 actually wins is class exams, where the source material is your professor's slide deck and there is no pre-existing question bank. Use UWorld for breadth on Step content. Use 4.7 (or a system wrapped around 4.7) for depth on class material. The mistake students make is trying to replace UWorld with a chat session; the right framing is to use the chat session for the part UWorld doesn't cover.

Sort of, and not in a useful way. If you paste the same MCQ back into a chat and ask for a reworded version, you get a reworded version. The problem is that a chat session has no memory of which underlying fact you have already mastered and which you are still missing, so you cannot drive the reword loop from your own miss data. The reword loop only produces retention when it is gated by spaced repetition: hard misses come back sooner with a different surface form, easy hits drift further out. That gating happens in the system around the model, not in any single Claude turn, which is why the same model produces different retention outcomes inside vs outside a wrapped study app.