OpenAI Just Told You What AI Is Actually Worth

Sam Altman killed Sora six months after launch. The official story was compute reallocation. The real story, broken by TechCrunch this morning, is that Claude Code was eating OpenAI's lunch.

Let that settle for a second. The most-hyped consumer AI product of 2024 — video generation from text, the thing that made the whole internet gasp — died not because the technology stopped working. It died because a developer coding tool at a competitor was pulling away enterprise customers faster than Sora could generate a single frame.

Sora was burning approximately $15 million per day in inference costs. Lifetime revenue: $2.1 million. Disney was told the product was dead thirty minutes after a working meeting. The person who made the call, Fidji Simo, described Sora as "a side quest." A side quest that consumed years of some of the best AI research talent on the planet.

There is a clean business truth buried inside this wreckage. Consumer AI products that create impressive demos do not pay for server racks. Enterprise tools that make developers 30% faster on a Monday morning do. OpenAI knows this now in the way you only know things after they cost you real money.

The deeper story is about who won. Anthropic did not win by building a better video generator. It won by making Claude Code relentlessly good at the thing software engineers do forty hours a week. TechCrunch reports that Anthropic's paying customers have doubled since January. Not because of a Super Bowl ad, though they ran one. Because Claude 3.7 Sonnet became the model that senior developers actually leave open in their editor all day.

OpenAI built Sora for the top of the funnel — the YouTube video, the conference demo, the "have you seen what AI can do now" text message to your parents. That funnel has diminishing returns in 2026 because everyone has already been amazed. What has compounding returns is being indispensable to someone's daily work.

The ARC-AGI-3 benchmark, launched last week, reinforced this from a different angle. Frontier models scored below one percent on the new interactive reasoning challenge — lower than a simple graph-search algorithm running on decade-old methodology. The benchmark is deliberately designed to require goal inference and environmental adaptation rather than pattern-matching against training data. GPT-5, Claude, and Gemini all scored worse than algorithms your computer science professor demonstrated in 2009. The models are extraordinary at the tasks they trained on and brittle on everything else.

This is not a reason to panic about AI. It is a reason to think carefully about which AI capabilities actually matter for the work being done today. The answer, increasingly, is the boring ones. Code completion. Refactoring. Writing tests. Summarizing documentation no one reads. These are not the demos that get standing ovations. They are the capabilities that justify a $20 per month subscription renewal without anyone thinking twice.

Anthropicfs accidental leak of Claude Mythos, their next frontier model, described it as "far ahead of any other AI model in cyber capabilities" — so far ahead that the company is warning cybersecurity defenders before release. That is a genuinely alarming capability profile. It is also, not coincidentally, a capability that enterprise security buyers will pay a great deal for. Offensive capability research has a natural commercial ceiling in consumer markets. It has a very different ceiling in government and enterprise contracts.

OpenAI's IPO is expected in Q4. The pivot they made by killing Sora is the pivot a company makes when it is preparing to show institutional investors a coherent revenue story. Consumer video generation is a cost center with cultural cachet. Enterprise coding tools are a subscription business with compounding retention. These are not the same thing, and the people who will price the IPO know the difference.

The question worth asking now is which other AI consumer products are Soras in disguise — technically impressive, strategically adjacent to where the money actually is, and quietly burning compute that could be redirected toward something boring and durable. Every AI lab has at least one. The ones that survive the next eighteen months will be the ones that found it before the board meeting did.