OpenAI's newest model, GPT-6 Astra, launched a week ago to widespread praise, including rebuilding Manhattan street by street inside a game engine. Now users say it feels dumber. The backlash matters because it questions whether AI labs quietly cut compute after the hype fades.
What actually happened
Developer synthwavedd wrote on X that "Astra feels significantly dumber for me today," calling it "The Post-Launch Lobotomy." Developer Pranjal Paliwal reviewed code Astra generated and concluded, "We don't have AGI. We have a regression," according to Decrypt. Developer Pankaj Kumar listed symptoms: faster answers, worse quality, and suspicion OpenAI "reduced the juice value," an unofficial term for computing effort spent per answer. Researchers Salio and Md Ismail Sojal both ran identical prompts against launch-day Astra and the current version, and both reported worse output now. Astra costs $10 (USD) per million input tokens and $50 per million output tokens, 2.5 times the launch price of predecessor GPT-5.6 Sol. Not everyone agrees a nerf happened. User Antikythera argued the model "is as dumb as it was on launch," and that hype simply wore off.
How we got here
This is not new. OpenAI's previous flagship, GPT-5.6 Sol, faced the identical backlash in July 2026, when users reported its top reasoning mode had gone shallow overnight. OpenAI executive Tibo Sottiaux denied deliberately weakening Sol, while confirming the company had been experimenting with "reasoning effort," the setting controlling how many steps a model thinks through before answering, per Decrypt. Coding tool Opencode's founder, Dax Raad, said his team already reverted to Sol from Astra, because costs doubled for results not worth it. OpenAI's own president used the term AGI to describe Astra at launch, a claim now being turned against the model by its own users.
Why this matters for you
For builders relying on OpenAI's API, inconsistent output quality has direct cost implications, since Astra charges more per token than Sol did at launch. Teams that priced projects around launch-week performance may need to rebudget or switch models mid-build. For everyday users, the episode is a reminder that AI benchmarks and demos capture a moment, not a guarantee. For the wider AR and wearable hardware push that depends on stable, affordable AI reasoning, unpredictable model quality raises the bar for what companies must test before shipping products tied to a single vendor's model.
The bigger question
If AI labs can silently adjust how much computing effort a model spends per answer after launch, how should users, developers, and regulators verify that an advertised capability still matches what a model delivers days or months later, and who bears responsibility for proving that consistency?
What to watch
OpenAI has not issued a Sol-style statement addressing the Astra complaints yet. Watch for an official response, similar to Tibo Sottiaux's July comments on Sol. Astra remains the first OpenAI model to cross the company's cybersecurity risk threshold, restricted to vetted defenders under its Daybreak program. More side-by-side prompt tests from developers are likely in the coming weeks.
