KorvaJoin the waitlist
Tech

AI Progress Can Mean Cheaper, Not Just Smarter

Scores moved only a few points while task costs fell sharply, but coding slipped and early reviews were harsh.

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

Early testers were mostly negative about OpenAI's GPT-6.2 Astra and Sol. Reviewer Matt Schumer called them essentially unusable, and Dan Shipper from Every said his 30-person team found them disappointing after testing them on everyday jobs including programming, writing and design.

Shipper said no one on the team loved Astra 6.2, adding "No one on our team who tried it switched back to it." Developer Theo also wrote a long complaint, and professor Ethan Mollick called the release uneven and weird.

On coding tests, the split was specific. On the Artificial Analysis Coding Agent Index, Sol dropped three points and slipped behind rivals, while Astra kept the same score and ranking. Artificial Analysis tied Sol's drop to a regression on SWE Atlas, a coding test, where it said Sol was "more likely to omit key facts in its responses."

On broader intelligence scores, the move was small but positive. Artificial Analysis raised Astra five points to 62, in second place, and Sol two points to 57.

The larger shift was in cost per task, what it takes to finish a standard batch of work. Astra fell to $4 per task from $6.24, down 35%, while Sol fell to $1.01 from $1.27, down about 20%.

The numbers suggest two causes working together. The models appear a little more token efficient, using fewer chunks of text to finish the work, alongside a lower price for each token.

High and max are effort settings that control how hard a model works before answering, with max trying harder and usually costing more. In this comparison, Fable 5.2, a rival Anthropic model on high settings, still costs more than double what Astra costs on max settings, while Sol on max settings costs under one-third what Opus 5.5, another Anthropic model, costs on high settings.

Alex Finn, who tested Astra on complex tasks, said, "Astra 6.2 finished the benchmark in a third of the time for a third of the price at a quarter of the tokens."

Ziwen, a tester who posted his feedback on X, acknowledged the quality tradeoff. "If quality is all you care about, then yeah, it flopped bad," he wrote, but added, "the future of AI should focus on efficiency more than quality." He added, "I don't think it's sustainable if the price just keeps going up just to get your job done better."

Hader, a commentator on the release, wrote that he had said before that efficiency was OpenAI's real edge. "The efficiency gain with Astra and Sol 6.2 is massive," he wrote, adding "they seem to be doubling down on it." For heavy everyday use, a slightly less capable model that is much cheaper and faster can be the more useful advance.

Questions people ask

What does cost per task mean?

Cost per task is the total bill for completing a fixed set of benchmark jobs, not the list price for tokens. It captures both price cuts and using fewer tokens, so it is closer to what a business would actually pay.

What is a token?

A token is a small chunk of text, roughly part of a word, that a model reads and writes as it works. Models charge by the token, so needing fewer tokens lowers the bill even before any price cut.

Why did coding problems shape the early reaction?

Coding was the main thing the early testers focused on. A model can look like a flop to that group while still working well for people doing research or other non-coding work.

More in Tech