@GosuCoder

The new Gemini 2.5 Pro 06-05 update is substantially better, but temperature matters a lot for AI coding assistants. Check out this graph, this shows the average eval score based on temperature. 0.7 is the clear winner here.
https://pbs.twimg.com/media/Gs2MdeM [...] ame=medium

The Google AI Studio free tier isn't going anywhere anytime soon

Hey folks, lots of good discussion going on here, I see all the posts and comments and deeply appreciate the love for AI Studio. The team and I have poured the last year + of our lives into this and it is great to see how important it has become in so many of your workflows. Given all the comments here, I thought I would do a wrap up post to clarify some things and share where we are at.

Moving AI Studio to be API key based does not mean you won't get free access to stuff. We have a free tier in the API used by millions of developers (more people than use the UI experience, by design).

Many folks mentioned 2.5 Pro as not being available for free in the API, this is in large because we offered it for free in the UI as well so we were giving out double free compute in a world where we have a huge amount of demand. I expect there will continue to be a free tier for many models in the future (though subject to many things like how the model is, how expensive it is to run, etc), and 2.5 Pro will hopefully be back in the free tier (we are exploring ways to do this, lifetime limits, different incentives etc)

The goal of AI Studio has long been to be a developer platform. The core experience we have been building for is a developer going to AI Studio, testing the model capabilities, having a wow moment with Gemini, and then going and getting an API key and building something that real people use. It was never built with the intention of being an every day assistant, that has always been the goal of the Gemini app (though acknowledging the feedback from folks on the historical gaps in functionality)

I am a deep believer in winning by building a great product. My hope and exception for the Gemini app is that they are on the cusp of their own "Gemini 2.5 Pro" level moment wrt the product experience really becoming 10x what it is today. In that world, it is going to hopefully be incredibly obvious that for everyday AI use, it is the best product Google has to offer. They have to earn that, I am under no illusion, but I deeply trust Josh Woodward (who was the first person I interviewed with when I was joining Google and the early supporter / builder of AI Studio) + the whole Gemini app team to pull this off.

Some of the historical weirdness in our launch strategy from a model POV has come from the AI Studio teams rapid ability to ship new models. The Gemini app team is deep in building the right infra technically and organizationally in order to do the same thing. They have already made great progress here and in some cases have been shipping faster than AI Studio / the Gemini API.

I saw lots of comments that folks want AI Studio to be part of Google AI Pro and Ultra plans, this is something we will explore, I think it is a cool idea but lots to work out there.

Overall, I hear and see the feedback. We will do this in a thoughtful way to minimize disruption, provide clear messaging, a great product experience, and make sure that Google has the world's best models, consumer products, and AI developer platforms. I will hang out here in the comments if folks have questions!

@GosuCoder
 
The new Gemini 2.5 Pro 06-05 update is substantially better, but temperature matters a lot for AI coding assistants. Check out this graph, this shows the average eval score based on temperature. 0.7 is the clear winner here.
https://pbs.twimg.com/media/Gs2MdeM [...] ame=medium

@chetaslua

Veo 3 Fast is available and it cost 1/5 of veo 3

Veo 3 - 100 credits
Veo 3 Fast - 20 credits

Prompt is not good but you will understand the difference

@AiHandbook
·
Latest Mensa (IQ scores) of top AI Models:

1) OpenAI o3
- IQ Score: 117
- Performance: Consistently high, with scores ranging from 110 to 130.
- Raw Score: 81% correct (13 out of 16).
---------------------------------------------------
2) Gemini 2.5 Pro
- IQ Score: 115
- Performance: Varied, with scores from 97 to 130, indicating some fluctuations.
- Raw Score: 75% correct (12 out of 16).
---------------------------------------------------
3) Claude-4 Opus
- IQ Score: 116
- Performance: Stable, with scores between 110 and 123.
- Raw Score: 75% correct (12 out of 16).

Overall,
- OpenAI o3 shows the highest average IQ score and consistency.
- Gemini 2.5 Pro exhibits more variability in performance.
- Claude-4 Opus maintains a stable performance.

@chetaslua

O4- mini updated

• It was meant to improve response length
• But auto-monitoring flagged a spike in content violations
• Quality > length — not worth the tradeoff
• Staying sharp, rolling back fast

@RubenHssd
BREAKING: Apple just proved AI "reasoning" models like Claude, DeepSeek-R1, and o3-mini don't actually reason at all.
 
They just memorize patterns really well.
 
Here's what Apple discovered:
 
(hint: we're not as close to AGI as the hype suggests)
 

Instead of using the same old math tests that AI companies love to brag about, Apple created fresh puzzle games.
 
They tested Claude Thinking, DeepSeek-R1, and o3-mini on problems these models had never seen before.
 
The result:
 
All "reasoning" models hit a complexity wall where they completely collapse to 0% accuracy.
 
No matter how much computing power you give them, they can't solve harder problems.
 
As problems got harder, these "thinking" models actually started thinking less.
 
They used fewer tokens and gave up faster, despite having unlimited budget.
 
Apple researchers even tried giving the models the exact solution algorithm.
 
Like handing someone step-by-step instructions to bake a cake.
 
The models still failed at the same complexity points.
 
They can't even follow directions consistently.
 

@RubenHssd
BREAKING: Apple just proved AI "reasoning" models like Claude, DeepSeek-R1, and o3-mini don't actually reason at all.
 
They just memorize patterns really well.
 
Here's what Apple discovered:
 
(hint: we're not as close to AGI as the hype suggests)
 

Instead of using the same old math tests that AI companies love to brag about, Apple created fresh puzzle games.
 
They tested Claude Thinking, DeepSeek-R1, and o3-mini on problems these models had never seen before.
 
The result:
 
All "reasoning" models hit a complexity wall where they completely collapse to 0% accuracy.
 
No matter how much computing power you give them, they can't solve harder problems.
 
As problems got harder, these "thinking" models actually started thinking less.
 
They used fewer tokens and gave up faster, despite having unlimited budget.
 
Apple researchers even tried giving the models the exact solution algorithm.
 
Like handing someone step-by-step instructions to bake a cake.
 
The models still failed at the same complexity points.
 
They can't even follow directions consistently.