Superpower Daily / Video

OpenAI’s Cheaper GPT-6.2 Sol Nearly Matches Astra on a Code Test

/ 8:01

Watch on YouTube

OpenAI’s GPT-6.2 launch cuts Astra’s standard API rates and positions cheaper Sol close to Astra on a coding benchmark. Anthropic adds Linked Billing for Claude in outside tools, Google rolls out Gemini task tools with checkout caveats, and a physics benchmark result prompts its co-creator to pause PhD admissions.

Published · Read as text

Key takeaways

  • OpenAI cut Astra’s standard input and output API rates by 20
  • Anthropic’s linked billing removes API-key sharing, but your
  • Gemini can make calls and prepare shopping carts, while
  • Strong results on checkable physics calculations raise

Stories covered

Read the transcript

Automatically transcribed from the narration. Timestamps open YouTube at that moment.

0:00

Hey everyone welcome to today's explainer briefing Really glad you could join us Today we're unpacking some absolutely massive shifts happening across the AI landscape right now We're talking everything from computing costs dropping to models actually reaching out and interacting with the real world on your behalf So let's jump right in Here's a quick rundown of our four stories today First up OpenAI's brand new GPT 6.2 models and a major price cut Second Anthropic is fundamentally changing

0:28

how you pay for Claude inside third party tools Third Google is turning Gemini into a hands on personal assistant And finally a stunning AI physics benchmark that literally has professors pausing PhD admissions Okay let's dive into this first story OpenAI has officially released GPT 6.2 Astra and GPT 6.2 Sol They're rolling out right now across the developer API chat GPT work and codecs And right out of the gate there's a major shift in pricing You heard

0:58

that right We are talking about a full 20% cut to Astra's standard input and output API rates compared to the previous GPT 6 Astra model Now Sol is keeping its previous pricing but OpenAI says both models are now optimized to use fewer output tokens on those really demanding tasks Basically your subscriber allowance is going to stretch a lot further For developers scaling apps this literally changes the daily math of what you can actually afford to build

1:24

So how do you actually choose between them in practice Well OpenAI still recommends Astra for the absolute hardest projects Think heavy lifting like taking a screenshot of a slide and adapting it perfectly into a custom presentation template or making deep complex modifications to an existing code base Sol on the other hand is your go to for everyday efficiency Get this on the Frontier Code 2.0 test which measures if code base changes are actually good enough to

1:50

merge Sol came within half a point of Astra Half a point And it costs about one fifth as much making high level AI assistance way more accessible for everyday developers Now both models also feature an ultra fast setting on certain plans generating tokens up to 4.5 times the standard rate But so the crucial point is this faster generation eats up way more of your usage allowance And just because you're generating tokens 4.5 times faster it absolutely

2:18

does not mean your entire project finishes 4.5 times faster It's a strict trade off You know you get that raw speed but you burn through your quota in a flash Let's move to our second story and see how this builds on the API ecosystem Anthropic just launched a new beta feature for outside tools called Linked Billing Historically you know the drill If you wanted to use Claude inside a third party app you had to generate an

2:43

API key paste it in and hope for the best Whoever holds that key can charge your account meaning the outside platform has to store and secure it Well Linked Billing removes that handoff entirely Developers just don't have to handle or protect a customer's API key to bill requests anymore It dramatically simplifies the daily friction Let's walk through a concrete example of how this actually flows Say you're using a launch partner like Linear An admin in your

3:11

Anthropic console simply approves the connection to Linear one single time Then you can set a strict monthly spending limit just for that platform From then on when Linear makes requests to Claude for your workspace Anthropic bills you directly The spending control stays firmly in your hands But you definitely need to be aware of this major limitation Because the actual requests are running inside the outside platform's organization not yours your strict customer security settings are left behind

3:40

So if your org has strict IP restrictions specific encryption keys or custom data retention rules set up in Anthropic yeah those don't travel to the outside tool That's a huge consideration for enterprise security teams Alright moving on to our third story Google's free Gemini tasks Google is taking a massive step here trying to shift Gemini from just answering our questions to actively delegating our chores They are rolling out these new task tools for Gemini and get

4:08

this they're free across six countries right now The US Japan South Korea Canada Australia and India You can actually have Gemini make phone calls to businesses for you log in to inspect your utility bills and assist with shopping To keep things locked down Google says these tasks run in isolated virtual machines with strict spending caps so you always maintain control And this brilliantly illustrates the potential here Google executive Josh Woodward shared a wild concrete example of

4:36

booking a car service At minute zero he gives Gemini his car details and home address By minute two Gemini is simultaneously calling two local dealerships And by minute five an appointment is successfully completely worked around his personal calendar Less than five minutes totally hands free It is a crazy glimpse into the very near future of personal assistance However there's always a catch Will website access blocks and mandatory checkout handoffs prevent this from being a truly hands

5:05

off assistant Right now yes A lot of websites actively block automated bots Plus for things like grocery shopping Gemini can absolutely build your cart but it still has to hand off to platforms like Instacart so you can give the human checkout approval So we aren't quite at the fully autonomous stage yet For our final story we are looking at a striking new benchmark result that isn't just a technical novelty It is already altering real world academic

5:32

decisions at the highest levels The benchmark is called TheoFizzEval It consists of 50 unpublished checkable theoretical physics problems And that unpublished part is a crucial detail because it proves the AI couldn't have just memorized the answers from its training data Well an AI agent using GPT 6 Astra equipped with web search and calculation tools correctly solved 49 out of the 50 calculations 49 out of 50 And it did it for an average cost of about $3

6:01

a problem Now what's really interesting about this is the human impact Physicist Suvarat Raju who actually co created this benchmark has paused new PhD admissions because of this shift He's assessing what this even means for his students prospects because as Andreas Karch pointed out we have a massive apprenticeship problem here For decades doing that mechanical grinding through math problems was exactly how young physicists developed their intuition and picked up the deeper concepts Which leaves the physics

6:31

community with a massive unanswered dilemma If AI handles the calculation grind how will future physicists actually develop the necessary intuition and judgment MIT's Daniel Harlow made a great point The calculations are really just the final step Deciding what question to ask and what to calculate that's the real bottleneck in scientific research If students skip the grind will they still learn how to ask the right questions That is a heavy question to ponder for sure Let's smoothly

6:57

transition into wrapping up today's explainer by distilling the high signal takeaways from all four of these massive stories Here's your daily summary First OpenAI released GPT 6.2 cutting standard Astra prices by 20% while the cheaper Sol model approaches its performance on certain tasks Second Anthropic introduced linked billing letting you pay for Claude in third party apps without exposing your API keys Though remember your security settings won't carry over Third Google launched free Gemini tools in six

7:28

countries that can make calls and build shopping carts but final checkout still largely requires you And finally AI just crushed a theoretical physics benchmark solving 49 out of 50 problems directly disrupting how we think about PhD training and human intuition I'll leave you with this to think about as you go about your day Which routine errands would you gladly delegate entirely to AI And which ones will always require your final human approval Read more at superpowerdaily

7:56

com Stay curious

YOUR READING SPACE

Notifications