KorvaJoin the waitlist
Tech

The Same AI Model Can Be Unusable and Excellent

Coders called GPT-6.2 unusable, while other testers called it fast and cheap, and both can be right.

AI image
The AI Daily Brief: Artificial Intelligence News and Analysis · Why the AI Kids Hate Astra 6.2

OpenAI released two models on Tuesday, GPT-6.2 Astra and a smaller version called Sol, and the first verdicts were brutal. Tester Matt Schumer told his followers, "They were essentially unusable. I do not recommend these models."

The loudest complaints came from coders, people who use AI to write and fix software. Dan Shipper, a founder at the tech company Every, said his 30-person team tested Astra 6.2 for weeks on coding, writing, design, and office research, and no one switched to it.

On coding, Shipper said Astra 6.2 often stopped too early on longer builds, which made end-to-end projects frustrating. On writing, he called it a step back, and on design, he said it defaulted to a noticeable house style.

Developer and entrepreneur Theo was blunter in a post titled "Why I Hate GPT-6.2 Astra." He wrote, "It runs in loops and goes nowhere," and summarized, "This model is clearly super smart, it is insufferable to work with on real projects."

That is when tester Alex Finn pushed back in a reply to fellow tester Ash, after days of online criticism. Finn argued that the verdict depended on what job you gave the model.

Tester Alex Finn put it specifically: "Astra 6.2 finished the benchmark in a third of the time for a third of the price at a quarter of the tokens."

Finn said that showed Astra could do complex work while using far less computing, which in his account meant lower cost. He said it handled incredibly complex tasks while using only about 2% of his usage allowance.

Other noncoding work told the same story, in his view. Ash said Astra had been his main powerhouse for agentic video editing, meaning AI that can operate editing software step by step, and founder Claire Vo said it beat rivals at image generation and animation code.

In Finn's view, reviewers focused only on the best coding score missed the tradeoff between that peak score and speed and price in everyday use. As he put it, "the panicking influencers never seem to take into account efficiency and price. This is a great model."

Questions people ask

What does using fewer tokens actually mean?

Tokens are the small chunks of words and data an AI reads and writes. A model that uses one quarter of the tokens to finish the same task needs less computing, so it usually answers faster and costs less per job.

How can a new model get worse at coding?

A new version can improve at some skills while slipping at others because training involves tradeoffs in data, tuning, and efficiency. In this release, outside measurements found small gains in general scores but a drop for the smaller model on coding tasks that require including key facts.

Should I ignore online reviews when picking an AI model?

Treat reviews as a starting point, then run your own frequent tasks. A model that frustrates a coder building software from start to finish may still be excellent for editing video, making images, or doing research quickly and cheaply.

More in Tech