To The Moon

FT recently dropped an article showing the spend split between different Anthropic models (indexed back to June). What was interesting about this chart (below) was that the recent growth across July wasn’t driven by Fable, but by Opus 5.

Many people asked questions about this. Why isn’t Fable 5 driving more adoption?

The overwhelming response to the question of “why has Fable 5 seen more lackluster adoption” was zero data retention (or ZDR). This topic has been front and center this summer. OpenAI wrote a blog post about it a few weeks ago. In my opinion, it’s becoming increasingly clear that users have a preference when it comes to data retention policies, and if you believe the majority of responses to Martin’s tweet, the labs have some decisions to make!

So what is Zero Data Retention? Typically, the standard API terms of service / arrangements has been the same for the last few years. The prompts / responses are held for some period of time (~30 days) to monitor for abuse. Abuse could be something like is the model used for hacking, or some other harmful intent. The labs want to know if the models were used this way so they could either do something about it, or use it as forensic evidence.

Of course, there are some models / customer contracts where nothing persists after the response is generated. No logs, no traces, no review queue. Nothing to leak, nothing for an employee to spy on, nothing for someone who steals employees credentials to see, nothing to subpoena, etc. Quick flag- zero doesn’t literally mean zero. Safety flags still exist even in this case (ie data that is flagged as harmful can be stored for much longer, but unflagged deleted immediately).

One distinction on this entire topic is policy vs architecture. A 30 day deletion policy in many ways is a “promise” (ie a policy). ZDR is more of an architecture. Policies / promises can be amended, terms of service can be updated, breaches can render it all irrelevant, etc. Architecture, on the other hand, has none of that. You can’t retroactively amend it. And we’ve seen a practical example demonstrating the policy vs architecture issue. In May 2025 a federal judge ordered OpenAI to preserve a lot of logs (I believe output logs). This included chats users had deleted. This was part of the New York Times copyright case. All of a sudden, the policy of “we delete after 30 days” became “we delete after 30 days unless someone says otherwise” (in this case the someone was a federal judge). For ZDR customers? There was nothing this judge’s order could do. There was nothing for OpenAI to preserve (for ZDR customers) because that data was deleted (architecturally) from the get go.

Ok circling back to Fable 5. When Anthropic launched it this summer (in June), they launched it with a 30 day retention policy - basically a 30 day retention of every prompt and output. AND everywhere the model is offered (their API, Bedrock, Foundry, Copilot, coding tools, etc). No ZDR exceptions, and existing ZDR agreements specifically did not apply to Fable 5 traffic. It’s easy to read that and think “huh, why, that seems weird.” But it’s actually reflects a very consistent viewpoint Anthropic has (which I don’t necessarily think is wrong!). The Fable series of models is by far the most capable model class ever created. And because it’s so powerful, it can be used in quite negative ways (cyber attacks, bio, etc). Anthropic would say the most serious misuse patterns only become visible (and thus truly block-able) when you can correlate across many interactions. On top of that, you can’t investigate something you never stored in the first place. If you take catastrophic risk seriously (which Anthropic clearly does), this data retention policy makes sense.

However, the rest of the market quickly reacted to this. Microsoft reportedly restricted its employees from using Fable 5. GitHub disabled it by default in CoPilot. And back to the chart (and tweet) I mentioned at the beginning of this post, the ramp data suggests the rest of the market did the same thing - routed traffic away from Fable 5. Of course, the rationale for the lower Fable 5 adoption is just a hypothesis. The cost of Fable 5 (relative to their other models or competing models) is clearly also a consideration for customers. Fable 5 is expensive, at $10 per million input tokens, roughly double GPT-5.6. Most workloads don't necessarily need the frontier (I've written before about the token bifurcation - the majority of tokens may flow to cheap, good-enough models while a thin slice of frontier tokens drives a majority the revenue). Opus 5 at half the price of Fable 5 always had the potential to cannibalize Fable.

However, the cost-only explanation (for the “lackluster” Fable 5 adoption) has some holes in it. Frontier models often complete the task in fewer tokens and fewer attempts. They’re more token efficient - expensive per token doesn’t mean the same thing as expensive per task (or per output). I do think (and hear very similarly) that the lower than expected usage of Fable 5 was more about the data retention policy (and remember, policy is different from architecture!). Ramp's own economist told the FT the retention requirement was a drag on adoption (in addition to price).

In the OpenAI piece I mentioned earlier,

Offering Zero Data Retention for Frontier Models

, they previewed something called Private Safety Processing. Customers can keep their data on infrastructure that they control (or sits encrypted under keys only the customer holds), and OpenAI runs automated misuse classifiers that scans for misuse patterns. OpenAI only receives an alert category and severity (not the actual data / prompt / response). I do think there’s some signal here. It’s not common to see a big launch-style blog post about retention policies…unless those policies are driving (or pushing away!) revenue. This is definitely a important topic, that will only become more important as the models get more capable.

So what does this all mean? Couple things come to mind. For one, the data flywheel inverted a bit. Some of the highest-value customers are now saying they’ll pay extra specifically to have their data “forgotten.” (so labs can’t learn from this to improve their products). The new improvement loop gets built from purpose-built data (RL environments, evals, opt-in partnerships), not from the data the customers spit out. Now the asset is that RL factory, not just the data traces.

One interesting thing for founders - your retention posture is now (in some ways) inherited from your vendor’s retention policies. “What does your model provider retain, and for how long?” is something I’d expect to show up on just about every security questionnaire you fill out. Companies using these models will want to get ZDR passthrough into your own contracts so you can pass that on. If not, architect for model portability, OR you’ll have to build that trust infrastructure yourself. There will definitely be demand for confidential inference, customer-held keys, audit tooling for AI traffic, etc.

For the labs: unilateral retention is appearing like a tax customers just aren’t willing to pay. Safety-through-retention either is something that starts to become an industry-wide standard, or it gets rebuilt cryptographically so retention isn’t required at all. My bet is the latter...

I’ll end with this - the CISO is starting to become a lot more important constituent in the buying process. At the same time, for many years a strong pitch in software was “your data makes our product better.” The strongest pitch in AI might turn out to be something entirely different: “we never saw your data at all.”