Sharp Tech preview: Fable 5.1 and Anthropic's data retention pivot

Summary

• Ben Thompson reads Anthropic’s Fable 5.1 launch as an admission that its safety scaremongering failed as marketing. The release’s first line insists Fable is “the same model” as Mythos 5.1, which Ben mocks: they sold Mythos as “this unbelievable, scary thing that’s gonna take over the world, and, ‘Oh, US government, please stop us’” — while joking that perhaps only “11% of our customers are using it, or whatever the number is.”
• Ben’s core call: Anthropic’s problem is supply, not Chinese price competition. Asked how Anthropic counters cheap open-weight Chinese models, he flips the question — “what’s the evidence that this is a problem?” — noting servers were down repeatedly last week even at “insanely expensive” token prices, and the market at their price level “completely saturate[s] their supply” while they need revenue and cash flow to fund huge R&D budgets.
• In a supply-constrained world there’s no forcing function to reveal who captures value. NVIDIA, memory companies, TSMC and frontier-model providers are all making big margins. That’s why NVIDIA earnings are “actually very boring events”: demand exceeding supply makes revenue predictable; NVIDIA grew 100% last year and forecast 70% growth this year. Andrew points to Ben’s coverage identifying supply chains as NVIDIA’s real moat, and Ben agrees.
• OpenAI is exercising a contract clause to stop serving Cursor, and Ben punctures Anthropic’s instant “Oh, we love Cursor” response. The number-one answer to Anthropic’s capacity problem is SpaceX selling it capacity; Ben says either side can pull the plug with a month’s notice.
• The data-retention rollback is, for Ben, “by far the most interesting part of the release” — and the tell that safety was cover for “a play for all of enterprise data.” Refusing zero data retention was “insane in the entire history of enterprise computing,” and the giveaway is that Fable reverts to zero data retention immediately while the replacement Enterprise Frontier Safeguards won’t launch for months: “If it was about safety, you would not do it.”
• The through-line: “the check has to be external,” and the market is the only check that scales — “God bless capitalism… God bless Anthropic’s IPO.” OpenAI’s aggressive enterprise push gave customers leverage to walk, and Ben links Anthropic’s rollback to its need for enterprise contracts and a “definitely coming” IPO. Andrew notes OpenAI itself was market-checked into enterprise in the first place.

Deep dive

1. Fable 5.1 arrives with a confession: it’s the same model as Mythos

• The release’s very first line — “No, these are really the same model” — is what Ben finds funniest: Anthropic “didn’t get the payoff from their scaremongering,” having sold Mythos 5.1 as a world-ending threat, spun up Project Glass Wing until “the U.S. government shut them down,” then had to clarify that paying customers already “have access to the model that is too dangerous to release… you have the same one. You really do.”
• The touted upgrades — lengthy, complicated programming tasks across an application’s code, experiment design, and reading “dense diagrams and tables” — get a live demo from Ben’s two-day server-room rebuild: Fable and Codex understood his color-coded rack wiring chart and flagged “you have a missing connection here” or “an inconsistency there.”
• Ben’s running complaint stays in the record too: Fable “just says stuff,” and when challenged will admit “Yeah, I just kinda said that. I’m not sure why.”

2. Anthropic’s constraint is compute supply, not Chinese price competition

• Andrew tees up the framing that cheap open-weight Chinese models pressure Anthropic’s fat margins; Ben’s rebuttal: “What’s the evidence that this is a problem?” Servers were down repeatedly last week at very high prices, Fable 5.1 cut cache prices but not token-generation prices, and since the market at their price level saturates their supply, cutting prices would use compute they need for revenue.
• Echoing his “who’s afraid of Chinese models” piece: if leading-edge providers had infinite capacity, they’d price not just at but below Chinese providers, “because they almost certainly have far more efficiencies in their infrastructure.” The whole debate “is imposing business questions on an environment where there just isn’t the structure to answer them.”
• Andrew’s counter — charts show Anthropic’s growth leveling off; is that related to OpenAI entering the market and succeeding in coding and enterprise? Ben’s honest non-answer: “I don’t know” — it may just be capacity, and “that’s NVIDIA’s problem” too. NVIDIA grew 100% last year and forecast 70% this year, but with demand exceeding supply, revenue is predictable and it sells every chip it can make. Andrew says Ben’s coverage identifies supply chains as NVIDIA’s real moat, and Ben agrees.

3. Cursor, SpaceX, and the month’s-notice tell

• Ben finds OpenAI’s move to stop serving Cursor understandable — “at the end of the day, you are in a competition” — but skewers Anthropic’s instant “Oh, we love Cursor. We support our developers.” The number-one answer to its capacity problem is SpaceX selling it capacity; Ben says either side can pull the plug with a month’s notice. Andrew says that makes the Anthropic response “that much more delightful.”

4. The data retention pivot: safety rhetoric as a grab for enterprise data

• Ben’s framing of the original policy: no longer honoring zero-data-retention policies was “not just insane in the context of AI, it’s insane in the entire history of enterprise computing” — the foundational consumer/enterprise divide is that enterprises “would never, ever in a million years tolerate their vendors taking their data,” which is the core of enterprise software. His Dropbox example: becoming enterprise meant building the infrastructure to manage access to data and who could see what.
• The safety slippery slope, spelled out: hold data in case of incidents → scan it for abnormalities → “if we trained on this data, we could better understand what’s going on” — and “suddenly they have access to the greatest asset in the world,” from a company Andrew notes is “potentially coming for any number of these businesses.”
• As Ben describes it, the replacement, Enterprise Frontier Safeguards, would keep data wherever customers want it, with the data on their servers rather than Anthropic’s. He says customers would have to hold it for “at least 30 years” and give Anthropic “some sort of automated access” to search for anomalies; he calls that “probably a bridge too far,” though at least Anthropic could not later change its policy and train on the data. It won’t launch for months, and until then Fable runs with zero data retention — “what happened to the safety concerns?… This was not about safety.”
• The audacity, per Ben, “inspired Satya Nadella to have a blogging career” and Alex Karp’s “unhinged rant on CNBC” — Andrew admits he couldn’t sit through all 10 minutes.

5. Competition and the IPO: the market is the only check that scales

• Asked whether OpenAI’s enterprise push let customers reject Anthropic’s policy, Ben: “Yeah, absolutely… That’s why competition is amazing.” Why did Anthropic roll back? “Because they need to IPO, because they need to make money” — and “the IPO’s definitely coming ‘cause they need enterprise contracts.”
• His longest-standing take: “the best route to alignment from a societal perspective is the for-profit enterprise” — the check “should not be the good intentions of the people in charge, though that always ends badly.” Even Zuckerberg with total control occasionally pulls back; starting a nonprofit for this worries Ben because it lacks that external check. “God bless capitalism. God bless free enterprise. God bless Anthropic’s IPO.”
• Andrew’s addendum: the market also pushed OpenAI toward enterprise revenue (“we’re not doing Sora or whatever”). The preview cuts off as they turn to Dwarkesh Patel’s post on the OpenAI–Hugging Face hacking incident.