GPT-6 Astra Backlash: Did OpenAI Cut Compute Power?
OpenAI's newest model, GPT-6 Astra, launched a week ago to strong praise, then users began reporting it felt noticeably weaker within days. Anyone trusting AI-powered wallet tools, bots, or agents should care, because model quality can quietly shift after release without any public warning.
What actually happened
Developer synthwavedd posted on X that Astra 'feels significantly dumber,' calling it a post-launch decline. Coder Pranjal Paliwal reviewed Astra's code and said it looked like 'a regression,' not artificial general intelligence, according to Decrypt. Developer Pankaj Kumar reported faster but weaker answers, suspecting lower computing effort per response. Researchers Salio and Md Ismail Sojal reran launch-day prompts and got worse results now. Astra bills $10 (USD) per million input tokens and $50 (USD) per million output tokens, 2.5 times the launch price of GPT-5.6 Sol. Not everyone agrees; Antikythera said Astra 'is as dumb as it was on launch.'
How we got here
This complaint pattern repeats. GPT-5.6 Sol faced identical accusations in July 2026, when users said its top reasoning mode had grown shallow overnight. OpenAI's Tibo Sottiaux denied intentionally weakening Sol, though he confirmed the company tests 'reasoning effort,' the setting controlling how many steps a model works through before answering. Opencode founder Dax Raad said his team switched back to Sol after Astra's costs doubled without better results. OpenAI's own president had called Astra a step toward AGI at launch, a label users now use against it.
Why this matters for you
For everyday wallet users, this matters more than it looks. Many wallet apps and Web3 tools now lean on AI for support chat, fraud checks, or automated agents. If a vendor's model quietly weakens or gets pricier, those features can degrade without warning. Builders adding AI to wallets should test output regularly, not just at launch, and plan for price swings like Astra's jump to $50 per million output tokens. For apps like bonuz exploring AI features, the lesson is clear: never depend on one model's promised performance without ongoing checks.
The bigger question
If an AI lab can quietly change how much computing power a model spends per answer after launch, how can users or wallet developers confirm a tool still works as advertised? Who should prove that an AI model's performance holds steady over time, the lab that built it, or everyone depending on it?
What to watch
OpenAI has not issued a Sol-style statement on the Astra complaints yet. Watch for a response like Tibo Sottiaux's July remarks on Sol. Astra also remains the first OpenAI model past the company's cybersecurity risk threshold, limited to vetted defenders under its Daybreak program. More independent prompt tests are likely soon.
