OpenAI unveils GPT-6.2 Astra

PLUS: Testers call it its worst

Together with

howdy, it’s Barsee again.

happy wednesday, AI family, and welcome back to AI Valley.

here are the biggest things worth knowing today:

  • OpenAI unveils GPT-6.2 Astra. Testers call it its worst

  • Jensen Huang breaks silence, signals willingness to return

  • Plus trending AI tools, posts, and resources

Let’s dive into the Valley of AI…

WISPR FLOW

Courtesy: Wispr Flow

Some problems don't solve themselves at a keyboard. Flow is perfect for walks: talk through the idea, the structure, and the tradeoffs, then get back a clean draft you can paste into your doc or AI tool. It's like having a writing assistant that keeps up with your brain in motion.

*This is sponsored

THROUGH THE VALLEY

“The worst model OpenAI has ever released”

In late September, OpenAI scrapped a model. GPT-6.1 Astra never shipped because, as its head of safety systems put it, it didn’t meet the bar on staying within scope and on how it reported back what work it had done.

On Tuesday its replacement arrived, GPT-6.2 Astra, and on paper it’s a win. It’s 20% cheaper than the last Astra, at $8 per million input tokens and $40 output. Artificial Analysis ranks it just behind Anthropic’s Fable 5.2 at about 15% of the cost per task. Someone from Plotly said their cost per task fell from about $20 to $4.

Then the people who’d actually lived with it started posting.

Theo Browne had been testing it for a while and got so frustrated that he spent a day writing OpenAI a letter. On launch day, sick in bed and unable to film a video, he just posted it. More than 700,000 people read it within hours. He doesn’t call the model dumb. He calls it a great code reviewer and world class at 3D. His problem is everything around the work. Two hours into building a game level, it stopped to ask him which file a reviewer had saved clips to, something he couldn’t possibly know. He started running other agents on a timer just to check whether Astra had stalled itself again. Its summaries were so dense, he wrote, that it “speaks in tongues that only 6.2 Astra understands.”

Others landed in the same place. Matt Shumer called the models “essentially unusable” and said that when Astra worked under Claude models, they had to redo its work. At Every, nobody on Dan Shipper’s 30-person team preferred it. Ethan Mollick called it “very uneven & weird.”

Put those together and a pattern shows up. Astra saves tokens, then spends people. Theo needed extra agents to watch it and another model just to read its summaries. Shumer needed one to redo its work.

OpenAI more or less sees it too. Its system card found Astra left out a known obstacle in 22% of tests, double the last Astra, and OpenAI suspects its push for shorter answers. That’s the September problem again, just milder.

The timing doesn’t help. Anthropic prices its IPO tonight and starts trading tomorrow, and AI commentator Carlos Santana joked that the week couldn’t be going better for Anthropic.

If you try it (and at this price, you should):

  • Revisit “be concise” in your prompts. It already is, and OpenAI’s own guide says as much.

  • Make it list every blocker and skipped step before it says “done.” Codex nudges it to; direct API calls won’t.

  • Use it for reviews, 3D work and short tasks, and check in on long runs.

It’s cheaper on paper. Whether it’s cheaper for you probably depends on how much of your day it eats.

TRENDING TOOLS

  • Devin Bot - a persistent AI engineering lead that runs a team of Devin agents for days or weeks, taking projects from goal to merged PRs

  • Atlas - an AI travel companion that plans your trips and nights out, books them, and keeps watch while you’re out

  • SID-2 - a search model that gets better answers by sending out a swarm of agents and scaling up compute

  • Scout - an AI sales assistant that handles the busywork so reps can spend their time closing

  • Second - a portable AI model of you that you own and can take to any agent, so they know you without starting over

  • Replit Missions - give it a goal and a budget, and a swarm of agents works on it until the goal is done

  • Vercel-Labs Agent-Desktop - lets AI agents use desktop apps by reading the OS accessibility tree instead of guessing from screenshots

  • Daydream - builds outfits, offers virtual try-ons, and can complete fashion purchases

WHAT I'M CONSUMING

THE VALLEY GEMS

What’s trending on social today:

THAT’S ALL FOR TODAY

Thank you for reading today’s edition. That’s all for today’s issue.

💡 Help me get better and suggest new ideas at [email protected] or @heyBarsee

👍️ New reader? Subscribe here

Thanks for being here.

REACH 100K+ READERS

Acquire new customers and drive revenue by partnering with us

Sponsor AI Valley and reach over 100,000+ entrepreneurs, founders, software engineers, investors, etc.

If you’re interested in sponsoring us, email [email protected] with the subject “AI Valley Ads”.