Asha vs Google Med-PaLM (MedLM): which medical AI can you actually use?
Med-PaLM 2 was Google's flagship medical research model, famous for hitting 86.5% on USMLE-style questions. Google productized it as MedLM on Vertex AI in late 2023, then deprecated MedLM on September 29, 2025. Asha runs on Google's current best medical-grade AI (Vertex AI Gemini 3.x) under physician guardrails, and it is live, free, and citation-grounded today.
TL;DR
Med-PaLM 2 was a landmark research result (86.5% on MedQA / USMLE-style questions in 2023) and Google productized it as MedLM for Google Cloud Vertex AI customers in healthcare and life sciences. Per Google's official Vertex AI release notes, MedLM access ended September 29, 2025. As of mid-2026, MedLM is deprecated and no successor product carries the Med-PaLM brand for direct medical-question answering. Google's healthcare AI strategy has shifted toward MedLM-successor offerings on Vertex AI plus general-purpose Gemini models. Asha is built on Vertex AI Gemini 3.1 Pro (primary) and Claude Sonnet (fallback) with a physician-designed retrieval-augmented architecture, BAA-covered through Google Cloud's HIPAA-eligible Vertex AI partner endpoint. Asha is live today, free at the entry tier, and you can sign up in seconds.
Side-by-side comparison
| Feature | Asha|AI | Google Med-PaLM 2 / MedLM |
|---|---|---|
| Availability today | Live at app.askasha.org. Sign up in seconds. | MedLM deprecated September 29, 2025 per Google Cloud release notes |
| Access path | Open web signup, free tier | Was enterprise-only via Vertex AI for allow-listed healthcare/life-sciences customers |
| Benchmark performance | RAG-grounded answers; benchmarks not the headline (architecture is) | 86.5% MedQA / USMLE-style (2023 published result for Med-PaLM 2) |
| Citation grounding | Every answer cites the corpus (PubMed, FDA, StatPearls, clinical guidelines) | Benchmark scores ≠ citation-grounded answers; MedLM was generative, not retrieval-grounded |
| Backbone LLM | Vertex AI Gemini 3.1 Pro (primary), Claude Sonnet (fallback) | Med-PaLM 2 (deprecated) |
| Sacred medical refusals | Architecturally enforced (prescription, dosing, diagnosis) | Google's product disclaimer posture (research-quality, not for clinical use) |
| HIPAA posture | PHI anonymization, BAA-covered via Vertex AI HIPAA-eligible endpoint | Was BAA-eligible on Vertex AI for healthcare customers (now deprecated) |
| Fleet breadth | Asha, Arohi, Lyra, Sage (4 healthcare agents) | Single research model; no patient/clinician/practice-mgmt fleet |
| API access for developers | A2A protocol, self-serve signup | Was Google Cloud Vertex AI only, allow-listed |
| Built by | Two board-certified physicians (DNAi Systems) | Google Research + Google DeepMind |
| Pricing | Free (100/mo); Pro $14.99/mo or $9.99/mo annual; A2A API tiers | Was Vertex AI enterprise pricing, allow-listed customers only |
When to choose Med-PaLM / MedLM
As of mid-2026, this is not a current option. MedLM was retired September 29, 2025. If you are reading this because your organization has historic MedLM integration, the relevant decision is: which successor product do you migrate to? Google now points customers toward Vertex AI Gemini for healthcare workloads. Asha is one of several products built on that stack with a physician-designed application layer on top.
When to choose Asha
Asha is live, citation-grounded, free at the entry tier, and built on the successor to the technology that powered Med-PaLM. Choose Asha when:
- You want a medical AI you can actually use today, without a Google Cloud enterprise contract or allow-list.
- You want benchmark-driven research to translate into a citation-grounded product, not stay locked in a paper.
- You want Google's best AI (Vertex Gemini 3.1 Pro) running under physician-authored guardrails and a curated medical corpus.
- You want sacred refusals on prescription, dosing, and definitive diagnosis built into the architecture.
- You want a four-agent fleet (Asha, Arohi, Lyra, Sage) that covers patient Q&A, practice management, research synthesis, and nutrition.
- You are a developer who wants self-serve API access without waiting for a Google Cloud sales conversation.
Specific use cases
Patient or clinician asking a clinical question right now
Med-PaLM / MedLM: Not accessible (deprecated September 2025).
Asha: Live at app.askasha.org. Free, citation-grounded, sacred refusals enforced.
USMLE-style board-question benchmarking
Med-PaLM 2 (paper): 86.5% on MedQA in the 2023 published result. Important research milestone.
Asha: Built for citation-grounded clinical Q&A, not for benchmark optimization. The architectures address different goals: Med-PaLM 2 maximized correctness on a closed test set; Asha maximizes citation-grounded correctness on real clinical questions.
Hospital system migrating off MedLM
Google's path: Vertex AI Gemini for healthcare workloads (developer build-your-own).
Asha A2A API: Pre-built medical AI on Vertex Gemini with physician guardrails; $999/month enterprise tier, BAA-covered, no Vertex AI engineering work required.
Research lab needing a medical LLM
Med-PaLM 2: Was never released for general research access; access was allow-listed and is now deprecated.
Asha: Free tier sufficient for most research probing. Lyra (research-specific agent in the fleet) is the right fit for systematic-review and meta-analysis workflows.
Pricing breakdown
Med-PaLM 2 / MedLM was enterprise Google Cloud Vertex AI pricing for allow-listed healthcare and life-sciences customers. Public per-token pricing was never published for the MedLM tier. The product is deprecated as of September 29, 2025. Asha is free for 100 queries/month, and Asha Pro is $14.99/month ($9.99/month billed annually) for consumers, and A2A API at $49 Developer / $199 Pro / $999 Enterprise. The upshot: if you wanted Med-PaLM access in 2024 and could not get on the allow list, Asha is the closest equivalent that is actually available today, running on the successor stack.
Frequently asked questions
Is Med-PaLM 2 still available?
No. Per Google Cloud's official Vertex AI release notes, MedLM (powered by Med-PaLM 2) was deprecated and access ended September 29, 2025. As of mid-2026, there is no public successor under the Med-PaLM brand.
What did Med-PaLM 2 actually do well?
Med-PaLM 2 scored 86.5% on MedQA (USMLE-style multiple-choice medical exam questions) in the 2023 paper, a 19-point jump over the prior Med-PaLM. Physician raters preferred Med-PaLM 2's long-form answers over physician-generated answers on 8 of 9 clinical evaluation dimensions per the published evaluation. It was an important demonstration that LLMs could pass medical licensing-style exams.
Does Asha beat the Med-PaLM benchmark?
Asha is built for a different objective: citation-grounded clinical answers with sacred refusals, not maximizing scores on closed multiple-choice benchmarks. Asha is RAG-grounded; the LLM (Vertex AI Gemini 3.1 Pro) is one component, and the retrieval layer over PubMed, OpenAlex, StatPearls, DailyMed, FDA, and clinical guidelines does most of the accuracy work. Direct benchmark comparison is apples-to-oranges.
Why did Google retire MedLM?
Google has not published a detailed rationale. The visible signal is that Gemini family models (which Asha uses today via Vertex AI) caught up to and surpassed the specialized Med-PaLM line on Google's own internal evaluations, and Google's healthcare AI strategy consolidated around Vertex AI Gemini with BAA coverage for healthcare customers. Asha is downstream of that consolidation.
Does Asha run on Google's AI?
Yes. Asha's primary cognition and verbalization runs on Vertex AI Gemini 3.1 Pro (and Gemini 3 Flash for sub-threshold tasks), via the Google Cloud HIPAA-eligible Vertex AI endpoint. Anthropic Claude Sonnet runs as a fallback. The physician-authored prompt, retrieval, and refusal architecture sit on top.