Hey - It’s Nico.

Welcome to another Failory edition. This issue takes 5 minutes to read.

If you only have one, here are the 3 most important things:

• JiviAI, a startup building an AI Whatsapp doctor, has shut down — learn why below

• How to steal a customer base

• Anthropic just released Claude Science, a new AI workbench for scientists — learn why this matters below

This Week In Startups

🔗 Resources

How to steal a customer base

AI's Optimisation Era

Frontier no more? The five forces challenging Frontier Models

If your product hasn't taken off

📰 News

Google introduces a faster, cheaper image generator with Nano Banana 2 Lite

Anthropic launches Claude Sonnet 5 as a cheaper way to run agents

OpenAI releases GPT 5.6 only to trusted organizations

Anthropic restores Claude Fable 5 worldwide as US lifts export controls

💸 Fundraising

Patronus AI raises $50 million to develop digital environments for testing AI agents.⁠

⁠Assort Health scores $120M series C to scale voice AI agent platform for healthcare

Attention Raises $30 Million Series B To Build AI Platform For Revenue Teams

Coval raises $28M as enterprises push voice agents into production

Fail(St)ory

The Cost of Being Everyone’s AI Doctor

JiviAI tried to put a doctor-like assistant inside a phone, a WhatsApp chat, and a clinical AI platform. They shut down in June 2026 after less than three years in market.

The tension was clear: healthcare AI can attract millions of curious users very quickly, while running medical-grade models, agents, voice systems, and multilingual support can burn cash faster than those users create a business.

What Was JiviAI:

JiviAI sold the idea of instant health guidance. Its consumer product was positioned as a personal AI doctor that could answer symptom questions, explain reports, track health signals, and help users make everyday decisions around stress, sleep, diet, and care.

The obvious use case was the moment before a doctor visit. A user has a fever, a confusing blood report, chest discomfort, or anxiety about a symptom, and wants a plain-language answer now. Jivi gave them a low-friction first step.

The app promised symptom checking based on medical cases and clinical data. Users could describe what they felt and get possible explanations, guidance on when to rest, and advice on when to see a doctor.

Report explanation was another strong wedge. Medical reports are full of ranges, acronyms, and scary-looking flags. Jivi translated that into simpler language, which is useful even for people who already have access to doctors.

The product also moved into mental health and lifestyle coaching. It offered stress, sleep, and mood self-assessments, plus AI chat using CBT and DBT-style methods. 

Jivi’s reach depended on accessibility. Public app copy claimed 100+ languages, and the company pushed WhatsApp as an entry point. 

That wide surface area made Jivi more ambitious than a simple symptom checker. It was trying to be a consumer health companion, a multilingual access layer, and eventually a clinical AI platform. The user experience looked lightweight, but the company behind it had to carry medical AI infrastructure, language coverage, and trust expectations on every interaction.

The Numbers:

• 🏥 Founded: 2023

• 📍 HQ: Gurugram, India

• 💸 Funding: $2.99M across one seed round

• 👥 Team: 11–50 employees publicly listed

• 📱 Users: 1.5M+ claimed in app-store copy

• 🌍 Reach: 3.5M users across 200 countries claimed later

• 🛑 Shutdown: Reported on June 26, 2026

Reasons for Failure: 

• AI health had expensive usage before it had durable revenue: Jivi claimed millions of users across many countries, but health queries are long, sensitive, multilingual, and costly to serve. Free or cheap guidance only works if it converts into paid care, provider workflows, insurance deals, or enterprise contracts.

• Proprietary medical AI put Jivi in a costly race: Jivi was promoting MedX, medical LLMs, voice AI, multimodal systems, agents, and open-source model work. That is a wide technical surface for a seed-stage company. Larger model labs keep improving the baseline and lowering prices, which makes proprietary differentiation expensive to defend.

• The product straddled consumer and clinical markets: Consumer health apps need trust, habit, and low acquisition cost. Clinical platforms need procurement, compliance comfort, workflow integration, and clear accountability when advice affects care. Jivi’s public materials spoke to both sides, which made the vision bigger and the execution burden heavier. A user may like instant report explanation, while a clinic needs reliability, auditability, escalation rules, and a clear operational owner.

• The financing path closed before the model matured: A planned funding round did not close, and acquisition talks failed to materialize. That left Jivi trying to fund a heavy medical AI platform before usage had become a repeatable revenue engine. Once a company has committed to proprietary models and broad platform claims, it has limited room to shrink into a small, cheap experiment.

Why It Matters: 

• Usage in AI healthcare can be misleading when sessions are easy to start but hard to monetize.

• Building proprietary models changes the funding requirement from app-company money to infrastructure-company money.

• Medical AI gets cleaner when one buyer has the budget, workflow pain, and accountability to pay early.

Trend

AI Scientists

This week, Anthropic added a new Claude to its growing family: Claude Science.

Its announcement did not generate the same splash as Claude Code or Claude Cowork, but this product is worth watching. 

Claude Science is Anthropic’s attempt to move Claude into the actual workflow of scientists: papers, datasets, code, figures, compute jobs, citations, and all the annoying glue work that sits between them.

Why it Matters

• Accuracy becomes inspectable: Claude Science attaches the code, environment, explanation, and history behind an output. That pattern should spread to any AI product where users need to trust the result before acting on it.

• Chat is turning into a workbench: The product sits inside the workflow: reading, coding, running jobs, making figures, revising them, and keeping the path back. The interface matters less than the job history.

• Domain context becomes the moat: Claude Science is useful because it knows the tools, databases, file types, and compute setup scientists already use. In vertical AI, the generic model is only the starting point.

Claude Science

Claude Science puts Claude inside the scientific workflow: reading papers, working with datasets, running code, making figures, checking citations, and helping prepare research outputs.

The product combines tools scientists usually use separately. Instead of jumping between PubMed, Jupyter, R, terminals, scientific databases, file viewers, and custom scripts, a researcher can coordinate more of that work from one place.

Claude Science is preconfigured for fields like genomics, single-cell biology, proteomics, structural biology, and cheminformatics. 

It can connect to scientific databases, journals, preprints, and domain-specific models, so researchers do not have to manually pull context from a dozen separate places.

It can produce research artifacts, not just text. Claude Science can run analyses, create figures, render 3D protein structures, show genome browser tracks, and work with chemical structures. 

Claude Science is also trying to make the work easier to trust. When it creates a figure or artifact, it keeps the code, environment, explanation, and message history behind it. A scientist can inspect the steps, rerun the work, catch mistakes, and challenge the result instead of treating the output like a black box.

Other Signals

• OpenAI GPT-Rosalind: OpenAI’s life-sciences model is built to help researchers reason through real research tasks, not just answer biology questions. It can pull evidence, analyze biological data, help design experiments, troubleshoot lab protocols, and use tools to produce reviewable outputs. It was launched in April.

• Google DeepMind Co-Scientist: DeepMind published Co-Scientist in Nature in May. It is a multi-agent system that generates, debates, ranks, and refines scientific hypotheses. The “tournament of ideas” setup is the interesting bit. It treats hypothesis generation as something agents can compete over, improve, and review before a human decides what deserves attention.

• Google Gemini for Science: Google also packaged Co-Scientist into a broader Gemini for Science push. The bundle includes Hypothesis Generation, Computational Discovery, and Literature Insights. That gives Google coverage across three common jobs: coming up with ideas, testing many computational approaches, and turning literature into something researchers can actually use.

• FutureHouse Robin: Robin was published in Nature in May and is focused on biological discovery. It combines literature search, hypothesis generation, experimental planning, data analysis, and follow-up insight generation. 

• Amazon Bio Discovery: AWS launched Amazon Bio Discovery in April. It gives scientists access to 40+ biology models, helps them choose models and inputs, and connects computational candidates to lab testing. In one MSK collaboration, AWS said the platform generated nearly 300,000 antibody molecules and narrowed them to 100,000 candidates for testing.

The Trend

AI-for-science is turning into its own product category. 

Claude Science is the cleanest example this week because it feels like a real workspace, not a demo pretending to be a lab. The bigger point is that OpenAI, Google, Microsoft, AWS, FutureHouse, and Anthropic are all circling the same opportunity.

Science has huge workflows full of expensive people, fragmented tools, private data, slow review loops, and repetitive analysis. AI products that can live inside those workflows, leave a trail, and earn trust will be much more valuable than another smart sidebar.

Help Me Improve Failory

That's all for today’s edition.

Cheers,

Nico