Frontier AI Models are Worthless - OpenAI is DEAD

A 34-minute monologue from Eli the Computer Guy, built on CNBC’s report from Salesforce’s Dreamforce conference (18 September 2026), where customers said last year’s models are good enough. He takes it as vindication. His frame is the smartphone: once a technology solves the problem you have defined, you stop upgrading, and in IT the first rule is “you touch it, you own it”. He puts the point where models became good enough at GPT-4, which he thinks handled almost every well-defined problem.

His second argument is about cadence. Adopting a new model means evaluating it, finding where it fits, integrating, testing and deploying, so he expects to have a view on OpenAI’s new flagship, Astra, around August 2027. Releases every few months look good to investors and journalists and wear out the people who have to support them. The third is price. Frontier models are “far overpowered” for 90 to 95 per cent of what organisations need, he says, and he is explicit that he is not calling them worthless, whatever the title says. If customers buy the cheap tier while the valuations assume the expensive one, the trillion-dollar numbers don’t add up. The fourth is margins: moving from subscription software to token-billed agents could pull a SaaS company’s gross margin from about 85 per cent to about 45, and he doesn’t believe Marc Benioff will hand the difference to the model vendors. He ends on architecture. Rather than dumping a database table into the context window, send the model the schema and a few sample rows, have it write the SQL, run the query yourself, and cache it for the next time someone asks.

The report says what he reads from it. Salesforce customers and partners told CNBC that older, cheaper models are plenty for everyday sales and service work. Alec Bronston of the retail data company SPINS said it is “already hard enough to keep up”. Tim Sanders, G2’s chief innovation officer, gave the 85-to-45 margin figure and said most agentic outcomes are driven by last year’s AI. Kevin Lee of NICE said his company doesn’t rely on high-end models such as Anthropic’s Fable for most work. On stage, Jensen Huang urged the labs to “run as fast as you can”. OpenAI’s GPT-6 Astra is priced at $10 per million input tokens, as he says, and Anthropic’s investors are reported to expect a valuation of up to $2 trillion at its listing. The other valuations he quotes, and the 20-cent price he gives for OpenAI’s cheapest tier, could not be pinned down.