Anthropic recently released Fable 5, the most capable frontier model available: a tier above the Opus line of models, with a million-token context window and noticeably better reasoning on exactly the kind of dense, ambiguous, judgment-heavy material that clinical trials are full of. It is genuinely impressive, and yet it is not running in HumanTrue, and it will not be until something specific changes. The reason is not capability, it's data. This post is about why that distinction is the whole game for us.
How we run AI: privately, by design
First, a bit of background on how we actually use AI. When you upload a protocol to HumanTrue, your data is not sent to a model maker's API. There is no OpenAI or Anthropic API key in our backend quietly shipping protocol content to a third party. Instead, we run frontier models as components inside our own architecture, through enterprise cloud services: Microsoft Azure (Azure OpenAI and Foundry) and AWS Bedrock. Inference happens in an isolated, contractually bounded environment. Prompts, agents, and the clinical trial data behind them stay inside that boundary. The model maker never sees the data and therefore can never train on it.
We have written about this before. How do AI Applications Work? explains the architecture in plain terms: how a vendor can use the most advanced models available while keeping customer data private. And The Uncomfortable Conversation Has a Missing Chapter makes the case that this architectural separation, not contractual assurance alone, is what sponsors should demand from every AI vendor.
The one requirement that is non-negotiable: no training on customer data
Clinical trial data is among the most sensitive data there is. A protocol encodes years of proprietary trial design. Operational documents carry patient-adjacent information. Timelines, endpoints, and enrollment details reveal a sponsor's competitive position. This is not data you paste into whatever tool is newest.
So we hold a hard line: customer data must never be used to train a model, and it must never persist with the model maker. Not "retained briefly," not "reviewed to improve quality."
This is not paranoia; it is the precondition for adopting AI at all. When Salesforce launched in 1999, the hard question was not whether cloud software was useful; it was "can I trust you with my data?" The SaaS industry spent two decades building the standards, certifications, and contracts that let businesses answer yes. AI resets that question, because models improve by seeing data, and model makers have every incentive to see more of it. Sponsors and CROs know this, which is why the no-training, no-persistence guarantee is the new normal. As we put it last time: right to access is not consent to use.
Why Fable 5 doesn't clear that bar (yet)
Here is the specific, current reason. As of this writing (June 2026), Fable 5 is not available under zero data retention terms. Anthropic retains prompts and outputs for 30 days for trust and safety purposes, and flagged content can be subject to human review. The model is capable enough that Anthropic wants to watch for patterns of misuse that only become visible across many requests.
Crucially, the requirement follows the model into the private cloud paths we rely on. To run Fable 5 on AWS Bedrock, you must explicitly opt in to sharing data with Anthropic, and as AWS's own announcement puts it: "once you opt into data retention, your data will leave AWS's data and security boundary." On Azure, the model likewise requires a subscription with retention enabled. The isolated environment, the thing our whole answer rests on, is exactly what the current terms ask you to open up.
To be fair and precise: this is not training. Anthropic states the retained data is used for safety monitoring, and taking a conservative posture toward the most capable model ever released is arguably the responsible move on their part. Retention is a different thing from training. But under our standard, the distinction does not rescue it. A 30-day window in which clinical trial tokens persist outside our cloud boundary, with the possibility of human review, is a deal-breaker.
So we made a deliberate, conservative choice: Fable 5 is not in HumanTrue. That is a statement about today's terms, not a permanent judgment of the model or its maker. We would rather wait than relax the standard.
"Newest model" is not the same as "right to deploy"
The industry reflex is to reach for the newest, most capable model the moment it ships, and to advertise that you did. Plenty of tools, including ones your employees may already have at work, will happily run the latest model against whatever gets pasted into them. That is fine for drafting a memo. It is not fine for analyzing a clinical trial.
The differentiator is not which model a vendor can call; any developer with a credit card can call the newest model within an hour of its release. The differentiator is the guarantees around the data you feed it. Capability without those guarantees is not a feature. It is a liability.
This is why every model we consider goes through the same gate, no matter how impressive the benchmarks: private deployment inside our architecture, no token sharing with the model maker, no training on customer data. Fable 5 does not pass that gate today and neither would any other model offered under similar terms.
What changes our answer
We expect this to be temporary. Retention requirements like this one tend to relax as providers build confidence in their safeguards, and the moment a zero-retention, no-training path for Fable 5 (or its successors) exists inside Azure or AWS Bedrock, we will move quickly to adopt it, the same way we adopt every model that clears the gate.
Until then, HumanTrue runs on frontier models we can deploy under our full data standard, and customers get frontier-grade capability without giving up control of their data. That has been the deal from day one, and we will not trade it for a benchmark score.
If you are evaluating AI vendors, this post suggests a useful question to add to the list: "Which models do you run, under what retention terms, and can you show me?" We listed three things sponsors should require from every AI vendor; this post is what living by them looks like in practice. If you are wrestling with the same trade-offs, we would like to hear from you.