Gary Marcus: GPT-6 Astra's costly gains don't fix OpenAI's broken economics
GaryMarcus · x · 2026-09-14
Gary Marcus published a critical summary of OpenAI's GPT-6 "Astra":
- Real progress on formally verifiable problems (partly via neurosymbolic techniques), but very expensive and not translating well outside benchmarks; only modest gains on the ECI
- Astra isn't much better than the latest Claude, and the next major model would cost far more for potentially modest gains
- Fierce competition from China, Anthropic, and eventually Amazon and Google is cutting into already negative margins
- OpenAI's agents have huge security problems the company seems unable to solve
His verdict: OpenAI is a master of the press release, but its economics still don't make sense as it builds GPT-7 at even higher cost.
More from Companies & People
- Microsoft publishes first draft of 'Humanist AI Code of Conduct' for its MAI models — gaganghotra_ · 2026-09-14
- Alpha School reportedly culls lagging students to game its AI teaching metrics — benjaminjriley · 2026-09-14
- Odd Lots interviews OpenAI president Greg Brockman on safety, slowdowns and compute — soumitrashukla9 · 2026-09-14
- SV Angel launches Project Blueprint, hires Jay Carney to shape AI policy consensus — gharik · 2026-09-14
- Colossus Publishes Year-Long Zuckerberg Profile After Interviewing 36 People, From Parents to Frontier Lab Rivals — aarthir · 2026-09-14
- Polymarket traders put 81% odds on tech layoffs exceeding 2025's 447,000 in 2026 — Polymarket · 2026-09-14