AI Therapy App The Path Hits 95 on Vera-MH Safety Benchmark
Founded by Tony Robbins and Calm alumni, The Path launched with $14.3 million in seed funding to build specialized foundation models for clinical AI therapy.
On May 21, 2026, AI therapy platform The Path launched out of stealth with $14.3 million in seed funding. Co-founded by Tony Robbins and former Calm engineering leaders Anson Whitmer and Tyler Sheaffer, the company replaces engagement-optimized consumer chatbots with specialized models tuned specifically for therapeutic outcomes. The initial capital round was led by Prime Movers Lab, alongside strategic investors including Apolo Anton Ohno, Deontay Wilder, and Designer Fund.
Clinical Safety Benchmarks
Evaluating clinical safety requires different metrics than general reasoning tasks. The Path relies on the open-source Vera-MH (Validation of Ethical and Responsible AI in Mental Health) standard, co-developed by Spring Health and clinical researchers. Vera-MH simulates multi-turn conversations using 100 clinician-designed personas to measure risk detection and appropriate clinical response.
General-purpose models struggle with this domain, often prioritizing user compliance over therapeutic necessity. The Path achieved a score of 95 out of 100 on the Vera-MH benchmark, significantly outpacing standard frontier models.
| Model | Vera-MH Score |
|---|---|
| The Path | 95 |
| Claude Opus 4.5 | 65 |
| GPT-5.2 | 65 |
| Gemini 3 Pro | 37 |
Technical Implementation and Safeguards
Consumer LLMs are post-trained to agree with users and provide immediate answers. The Path alters this behavior at the foundation level, training its models to challenge cognitive distortions and guide users through structured psychological frameworks. If you build clinical applications, determining how to evaluate AI output requires distinguishing between helpful assistance and necessary clinical pushback.
The system implements persistent historical memory across sessions, allowing the AI to track psychological progress toward long-term user goals. For developers exploring how to add memory to AI agents, The Path demonstrates how continuous context management shifts a product from a transient chat tool into a stateful tracking system. Users can select from 11 distinct AI therapist personalities and tune parameters like directness.
Safety guardrails include a deterministic crisis protocol. The platform detects severe psychological distress and instantly routes the user to human clinical teams and national crisis hotlines. This dual-layer approach ensures the system does not rely entirely on generative output during critical medical emergencies.
Availability and Pricing
The company previously operated under the name Mental, focusing specifically on men’s mental health. During that beta phase, the platform processed over 3.5 million messages across an initial base of 50,000 members. The rebranded application is now available on the iOS App Store and Google Play.
The Path currently operates on a free tier during its launch window. The company plans to transition to a subscription model priced at approximately $40 per month. Entering a digital healthcare infrastructure where the U.S. average delay between symptom emergence and treatment exceeds 10 years, the platform targets massive scale. However, it must also navigate stringent privacy expectations following data-sharing controversies from early telehealth competitors.
If you design AI systems for sensitive domains, general-purpose intelligence is insufficient for production deployment. Specialized benchmarks and domain-specific post-training are mandatory requirements for safely handling high-stakes interactions.
Get Insanely Good at AI
The book for developers who want to understand how AI actually works. LLMs, prompt engineering, RAG, AI agents, and production systems.
Keep Reading
How to Fine-Tune Cosmos Predict 2.5 for Robotics With LoRA
Learn how to adapt NVIDIA's 2B and 14B Cosmos Predict 2.5 world foundation models using parameter-efficient fine-tuning methods like LoRA and DoRA.
$12M Seed Backs NanoClaw's Containerized AI Agent Framework
NanoCo has raised $12 million to scale NanoClaw, a minimalist, security-focused alternative to OpenClaw that isolates AI agents inside Docker containers.
NeoCognition raises $40M seed for self-learning AI agents
NeoCognition emerged from stealth with a $40M seed round led by Cambium Capital and Walden Catalyst to build agents that learn through trial and error.
Untrained Tasks Now Possible via π0.7 Robotic Brain
Physical Intelligence unveils π0.7, a foundation model enabling robots to solve novel, complex problems through compositional generalization.
IBM MAMMAL Foundation Model Unifies Gene and Protein Analysis
IBM Research released MAMMAL, a unified 458-million parameter foundation model that processes genes, proteins, and molecules in a single shared framework.