This is not a leaderboard. It is not a benchmark review. It is the stack I actually reach for right now, in May 2026, when I am trying to study, write, build, present, or solve some annoying workflow problem quickly.
The broad rule has not changed: do not marry one platform. The best workflow is still a set of tools with different jobs. The change since the last version is that the tools are starting to connect to each other in useful ways. The winning products are not just better models. They are better work surfaces.
The May 2026 Shift
The center of gravity has moved from single chatbot sessions to connected workspaces: Notion pages with AI, NotebookLM notebooks inside Gemini, OpenEvidence with clinical workflows, and design/code tools that turn rough ideas into usable prototypes.
Notion AI Is Becoming My Study Desk
Notion has become more than a second brain for me. I am using Notion AI much more for studying, reviewing information, and turning scattered material into something I can actually work with.
The workflow is simple: dump the source material into a page, add my own messy notes, and then use Notion AI to structure it. It can turn a pile of reading into a study outline, extract key questions, reorganize meeting notes, or build a usable database from a rough thought dump. That sounds mundane, but it is exactly the kind of mundane work that drains time.
The reason this works is context. Notion already knows the project, the surrounding pages, and the structure of the workspace. You are not starting every prompt from zero. You are asking questions inside the system where the work already lives.
Notion Meetings Has Earned a Spot
Notion AI Meeting Notes has also been great. I use it often because it records the meeting, transcribes it, and produces a summary with action items directly inside the workspace where the follow-up needs to happen.
That matters. A meeting note app that creates a nice transcript somewhere else is useful, but it still creates another place to check. Notion Meetings lands in the same system where my projects and tasks live. That makes the output much easier to act on.
The usual caveat applies: get consent, know your institutional rules, and do not record clinical encounters unless your institution has explicitly approved the workflow. Convenience is not a compliance policy.
NotebookLM Plus Gemini Is Finally a Loop
I have been enjoying the connection between NotebookLM and Gemini. Google now lets notebooks sync between Gemini and NotebookLM, which means the same project knowledge can move between two different modes of work.
NotebookLM remains the better grounded study environment. It is where I put guidelines, PDFs, papers, lecture notes, and disease-site material. It is still excellent for asking source-grounded questions, making study guides, generating audio overviews, and reviewing material without relying on the model's memory.
Gemini is the better general reasoning and web-connected interface. Being able to pull notebook context into Gemini changes the workflow. I can use NotebookLM to organize and ground the material, then use Gemini to reason across it, draft from it, or connect it to current web information.
For studying, this is genuinely useful. It feels less like "upload a PDF and ask a question" and more like building a living knowledge base that can move with you.
OpenEvidence Is Becoming a Platform
OpenEvidence is still one of the clinical pillars. For medical questions, I use it because it is designed around evidence and citations rather than generic chatbot fluency.
The latest updates make it feel less like a single search box and more like a clinical platform. The answers increasingly include figures, graphs, and visual summaries when the question benefits from them. That is a big deal. Sometimes a Kaplan-Meier curve, forest plot, or dosing table tells you more than another paragraph of prose.
OpenEvidence has also been putting out a tremendous number of tools: deeper consult-style research, clinical workflow features, coding support, and more specialty-specific evidence integrations. We need to update the OpenEvidence entry soon. At this point, the entry is basically sprinting behind the product with a clipboard.
My rule is still the same: OpenEvidence is a tool, not an attending. Read the cited sources when the answer matters. But as a fast, evidence-grounded first pass, it is hard to beat.
Nano Banana Pro for Figures and Workflows
Nano Banana Pro, especially through Google AI Studio, has become my default for visual generation: workflows, figures, conceptual diagrams, icons, and visual scaffolds for talks.
The key is not to ask for a final perfect figure in one shot. That usually fails. I use it to generate components:
- a workflow diagram skeleton
- a clean icon set
- a patient education visual
- a conceptual figure for a talk
- a draft schematic that I can rebuild in Canva, BioRender, PowerPoint, or Claude Design
The best workflow is still describe, refine, generate, assemble. I dictate the figure idea, use Gemini to tighten the prompt, generate with Nano Banana Pro, then bring the output into a tool where I can control labels and layout.
This is the closest AI has come to removing the blank-slide problem.
Claude Design Is the Surprise Winner
Claude Design has been phenomenal. I have been using it to create quick HTML websites, front-end mockups, and interface concepts that I can then move into Claude Code or Codex for real implementation.
I have not seen an equivalent ChatGPT workflow yet that feels as natural for front-end design. ChatGPT can write code. It can make decent layouts. But Claude Design feels different because it holds the visual system together while you iterate. The typography, spacing, colors, and layout do not fall apart after two rounds of changes.
That matters for scientific and medical work. Most of us are not trying to win a design award. We are trying to make a patient education page, a grant dashboard, a research tool, a meeting site, or a quick internal prototype that does not look like it was assembled during a fire drill.
The pattern I like:
- Use Claude Design to create the front end.
- Tighten the interaction and layout.
- Move the implementation into Claude Code or Codex.
- Turn the prototype into something maintainable.
That is a very different workflow from asking a chatbot to "make me an app."
Remote Coding From the Phone
The other big change is that I am using remote coding sessions more often from my phone. Codex and remote app workflows make it possible to solve small coding problems while I am between things in clinic.
I am not talking about building a production app from the hallway. I mean quick, bounded work:
- fix a broken script
- draft a small data transformation
- review an error message
- scaffold a quick utility
- nudge a local project forward when I have five minutes
This is where AI coding agents are quietly useful. They let me keep momentum on small technical problems without needing to sit down at the full workstation. The best use case is not heroic. It is reducing friction.
The Stack Right Now
What I Am Watching
The next version of this stack will probably be less about model names and more about connected workflows.
The interesting question is no longer "which chatbot is smartest?" The interesting question is: where does the work live, how much context does the AI have, and can the output move into the next step without friction?
That is why Notion, NotebookLM, Gemini, OpenEvidence, Claude Design, and Codex are all interesting right now. They are not just answer machines. They are becoming work surfaces.
And in medicine and research, that is what we need: not a smarter toy, but a stack that helps us read, think, write, build, and verify faster.
Product references: Google on Gemini notebooks and NotebookLM, NotebookLM chat upgrades, Notion AI Meeting Notes, Anthropic Claude Design, and OpenEvidence DeepConsult.