Feeling great to see something confidential

- Anyone have any idea why it says &#x27;confidential&#x27;?- Anyone actually able to use it? I get &#x27;You&#x27;ve reached your rate limit. Please try again later&#x27;. (That said, I don&#x27;t have a paid plan, but I&#x27;ve always had pretty much unlimited access to 2.5 pro)[Edit: working for me now in ai studio]

Today I guess. They were not releasing the preview models this time and it seems the want to synchronize the release.

Thanks for the hint about Canvas&#x2F;2.5. I have access to 3.0 in AI Studio now, and I agree the results are very similar.

Allegedly it&#x27;s already available in stealth mode if you choose the &quot;canvas&quot; tool and 2.5. I don&#x27;t know how true that is, but it is indeed pumping out some really impressive one shot codeEdit: Now that I have access to Gemini 3 preview, I&#x27;ve compared the results of the same one shot prompts on the gemini app&#x27;s 2.5 canvas vs 3 AI studio and they&#x27;re very similar. I think the rumor of a stealth launch might be true.

It&#x27;s already available. I asked it &quot;how smart are you really?&quot; and it gave me the same ai garbage template that&#x27;s now very common on blog posts: <a href="https:&#x2F;&#x2F;gist.githubusercontent.com&#x2F;omarabid&#x2F;a7e564f09401a64e7daf7f5f84351e46&#x2F;raw&#x2F;9024dd6ca946ac442d0a5fed81da2d56fd72c5db&#x2F;gistfile1.txt" rel="nofollow">https:&#x2F;&#x2F;gist.githubusercontent.com&#x2F;omarabid&#x2F;a7e564f09401a64e...</a>

&gt; <a href="https:&#x2F;&#x2F;gemini.google.com&#x2F;app" rel="nofollow">https:&#x2F;&#x2F;gemini.google.com&#x2F;app</a>How come I can&#x27;t even see prices without logging in... they doing regional pricing?

are you sure its available in cursor? ( I get: We&#x27;re having trouble connecting to the model provider. This might be temporary - please try again in a moment. )

It&#x27;s available in cursor. Should be there pretty soon as well.

On gemini.google.com, I see options labeled &#x27;Fast&#x27; and &#x27;Thinking.&#x27; The &#x27;Thinking&#x27; option uses Gemini 3 Pro

How long does it typically take after this to become available on <a href="https:&#x2F;&#x2F;gemini.google.com&#x2F;app" rel="nofollow">https:&#x2F;&#x2F;gemini.google.com&#x2F;app</a> ?I would like to try the model, wondering if it&#x27;s worth setting up billing or waiting. At the moment trying to use it in AI Studio (on the Free tier) just gives me &quot;Failed to generate content, quota exceeded: you have reached the limit of requests today for this model. Please try again tomorrow.&quot;

you can bring your google api key to try it out, and google used to give $300 free when signing up for billing and creating a key.when i signed up for billing via cloud console and entered my credit card, i got $300 &quot;free credits&quot;.i haven&#x27;t thrown a difficult problem at gemini 3 pro it yet, but i&#x27;m sure i got to see it in some of the A&#x2F;B tests in aistudio for a while. i could not tell which model was clearly better, one was always more succinct and i liked its &quot;style&quot; but they usually offered about the same solution.

I hope some users will switch from cerebras to free up those resources

Looks to be available in Vertex.I reckon it&#x27;s an API key thing... you can more explicitly select a &quot;paid API key&quot; in AI Studio now.

For me it’s up and running. 
I was doing some work with AI Studio when it was released and reran a few prompts already. Interesting also that you can now set thinking level low or high. I hope it does something, in 2.5 increasing maximum thought tokens never made it think more

It&#x27;s available to be selected, but the quota does not seem to have been enabled just yet.&quot;Failed to generate content, quota exceeded: you have reached the limit of requests today for this model. Please try again tomorrow.&quot;&quot;You&#x27;ve reached your rate limit. Please try again later.&quot;Update: as of 3:33 PM UTC, Tuesday, November 18, 2025, it seems to be enabled.

It seem that Google doesn&#x27;t prepare well to release Gemini 3 but leak many contents, include the model card early today and gemini 3 on aistudio.google.com

Can confirm. I was able to access it using GPTel in Emacs using &#x27;gemini-3-pro-preview&#x27; as the model name.

it is live in the api&gt; gemini-3-pro-preview-ais-applets&gt; gemini-3-pro-preview

There&#x27;s a waitlist for using Gemini 3 for Gemini CLI free users: <a href="https:&#x2F;&#x2F;docs.google.com&#x2F;forms&#x2F;d&#x2F;e&#x2F;1FAIpQLScQBMmnXxIYDnZhPtTP3xr5IwHNzKW4nLomuQ1tGOO-UldMdQ&#x2F;viewform" rel="nofollow">https:&#x2F;&#x2F;docs.google.com&#x2F;forms&#x2F;d&#x2F;e&#x2F;1FAIpQLScQBMmnXxIYDnZhPtTP...</a>

With this kind of pricing I wonder if it&#x27;ll be available in Gemini CLI for free or if it&#x27;ll stay at 2.5.

It is so impressive that Anthropic has been able to maintain this pricing still.

Still cheaper than Sonnet 4.5: $3&#x2F;M for input and $15&#x2F;M for output.

Thrilled to see the cost is competitive with Anthropic.

It&#x27;s interesting that grounding with search cost changed from* 1,500 RPD (free), then $35 &#x2F; 1,000 grounded promptsto* 1,500 RPD (free), then (Coming soon) $14 &#x2F; 1,000 search queriesIt looks like the pricing changed from per-prompt (previous models) to per-search (Gemini 3)

API pricing is up to $2&#x2F;M for input and $12&#x2F;M for outputFor comparison:
Gemini 2.5 Pro was $1.25&#x2F;M for input and $10&#x2F;M for output
Gemini 1.5 Pro was $1.25&#x2F;M for input and $5&#x2F;M for output

2025: solve the biking pelican problem2026: cure cancer

It generated a quite cool pelican on a bike: <a href="https:&#x2F;&#x2F;imgur.com&#x2F;a&#x2F;yzXpEEh" rel="nofollow">https:&#x2F;&#x2F;imgur.com&#x2F;a&#x2F;yzXpEEh</a>

Anthropic is also doing this for long context &gt;= 200k Tokens on Sonnet 4.5

Is it the first time long context has separate pricing? I hadn’t encountered that yet

If the model solves your needs in fewer prompts, it costs less.

And of course they hiked the API pricesStandard Context(≤ 200K tokens)Input $2.00 vs $1.25 (Gemini 3 pro input is 60% more expensive vs 2.5)Output $12.00 vs $10.00 (Gemini 3 pro output is 20% more expensive vs 2.5)Long Context(&gt; 200K tokens)Input $4.00 vs $2.50 (same +60%)Output $18.00 vs $15.00 (same +20%)

This is the heroic move everyone is waiting for. Do you know how this will be priced?

How about for Pro (not Ultra) subscribers?

Gemini CLI team member here. We&#x27;ll start rolling out today.

I&#x27;m already seeing it in <a href="https:&#x2F;&#x2F;aistudio.google.com&#x2F;" rel="nofollow">https:&#x2F;&#x2F;aistudio.google.com&#x2F;</a>

Honestly I liked 2.5 Pro preview much more than the final version

Not the preview crap again.
Haven&#x27;t they tested it enough?
When will it be available in Gemini-CLI?

that&#x27;s why everyone using AI for code should code in rust only.

Good personal benchmarks should be kept secret :)

I don&#x27;t think it would be a good idea to publish it on a prime source of training data.

NIBBLES.BAS maybe [1]If you make some assumptions about the species of the snake, it can count as a basic python benchmark ;)[1] <a href="https:&#x2F;&#x2F;en.wikipedia.org&#x2F;wiki&#x2F;Nibbles_(video_game)" rel="nofollow">https:&#x2F;&#x2F;en.wikipedia.org&#x2F;wiki&#x2F;Nibbles_(video_game)</a>

Could be. I&#x27;ll reply to my comment later with pass&#x2F;fail results of a re-run.

Is observed speed meaningful for a model preview? Isn’t it likely to go down once usage goes up?

I moved to using the model from python coding to golang coding and got incredible speedups in writing the correct version of the code

I agree that benchmarks are noise. I guess, if you&#x27;re selling an LLM wrapper, you&#x27;d care, but as a happy chat end-user, I just like to ask a new model about random stuff that I&#x27;m working on. That helps me decide if I like it or not.I just chatted with gemini-3-pro-preview about an idea I had and I&#x27;m glad that I did. I will definitely come back to it.IMHO, the current batch of free, free-ish models are all perfectly adequate for my uses, which are mostly coding, troubleshooting and learning&#x2F;research.This is an amazing time to be alive and the AI bubble doomers that are costing me some gains RN can F-Off!

Everything is about context. When you just ask non-concrete task it&#x27;s still have to parse your input and figure what is tic-tac-toe in this context and what exactly you expect it to do. This is why all &quot;thinking&quot;.Ask it to implement tic-tac-toe in Python for command line. Or even just bring your own tic-tac toe code.Then make it imagine playing against you and it&#x27;s gonna be fast and reliable.

and models are still pretty bad at playing tic-tac-toe, they can do it, but think way too muchit&#x27;s easy to focus on what they can&#x27;t do

A lot of newer models are geared towards efficency and if you add the fact that more efficent models are trained on the output of less efficent (but more accurate) models....GPT4&#x2F;3o might be the best we will ever have

I maintain a set of prompts and scripts for development using Claude Code. They are still all locked to using Sonnet 4 and Opus 4.1, because Sonnet 4.5 is flaming hot garbage. I’ve stopped trusting the benchmarks for anything.

I wonder whether it could be related to some kind of over-fitting, i.e. a prompting style that tends to work better with the older models, but performs worse with the newer ones.

&gt; This is a reminder that benchmarks are meaningless – you should always curate your own out-of-sample benchmarks.Yeah I have my own set of tests and the results are a bit unsettling in the sense that sometimes older models outperform newer ones. Moreover, they change even if officially the model doesn&#x27;t change. This is especially true of Gemini 2.5 pro that was performing much better on the same tests several months ago vs. now.

they are not meaningless, but when you work a lot with LLMs and know them VERY well, then a few varied, complex prompts tell you all you need to know about things like EQ, sycophancy, and creative writing.I like to compare them using chathub using the same promptsGemini still calls me &quot;the architect&quot; in half of the prompts. It&#x27;s very cringe.

&gt;&gt;benchmarks are meaninglessNo they’re not. Maybe you mean to say they don’t tell the whole story or have their limitations, which has always been the case.&gt;&gt;my fairly basic python benchmarkI suspect your definition of “basic” may not be consensus. Gpt-5 thinking is a strong model for basic coding and it’d be interesting to see a simple python task it reliably fails at.

I didn&#x27;t tell you what you should think about the model. All I said is that you should have your own benchmark.I think my benchmark is well designed. It&#x27;s well designed because it&#x27;s a generalization of a problem I&#x27;ve consistently had with LLMs on my code. Insofar that it encapsulates my coding preferences and communication style, that&#x27;s the proper benchmark for me.

How can you be sure that your benchmark is meaningful and well designed?Is the only thing that prevents a benchmark from being meaningful publicity?

It made a working game for me (with a slightly expanded prompt), but the ghosts got trapped in the box after coming back from getting killed. A second prompt fixed it. The art and animation however was really impressive.

Something else to consider. I often have much better success with something like: Create a prompt that creates a specification for a pacman game in a single html page. Consider edge cases and key implementation details that result in bugs. &lt;take prompt&gt;, execute prompt. It will often yield a much better result than one generic prompt. Now that models are trained on how to generate prompts for themselves this is quite productive. You can also ask it to implement everything in stages and implement tests, and even evaluate its tests! I know that isn&#x27;t quite the same as &quot;Implement pacman on an HTML page&quot; but still, with very minimal human effort you can get the intended result.

I like to ask &quot;Make a pacman game in a single html page&quot;. No model has ever gotten a decent game in one shot. My attempt with Gemini3 was no better than 2.5.

after taking a walk for a bit i decided you’re right. I came to the wrong conclusion. Gemini 3 is incredibly powerful in some other stuff I’ve run.This probably means my test is a little too niche. The fact that it didn’t pass one of my tests doesn’t speak to the broader intelligence of the model per se.While i still believe in the importance of a personalized suite of benchmarks, my python one needs to be down weighted or supplanted.my bad to the google team for the cursory brush off.

No, do not share it. The bigger black hole these models are in, the better.

Using a single custom benchmark as a metric seems pretty unreliable to me.Even at the risk of teaching future AI the answer to your benchmark, I think you should share it here so we can evaluate it. It&#x27;s entirely possible you are coming to a wrong conclusion.

I find this hard to understand. I have AI completely choke on my code constantly. What are you doing where it performs so well? Web?I constantly see failures in trivial vectors projections, broken bash scripts that don&#x27;t properly quote variables (fail if space in filenames), and near completely inability to do relatively basic image processing tasks (if they don&#x27;t rely on template matches).I accidentally spent $50 on Gemeni 2.5 Pro last week, with Roo, trying to make a simple Mock interface for some lab equipment. The result: it asks permission to delete everything it did and start over...

I&#x27;m dying to know what you&#x27;re giving to it that&#x27;s choking on. It&#x27;s actually really impressive if that&#x27;s the case.

Google reports a lower score for Gemini 3 Pro on SWEBench than Claude Sonnet 4.5, which is comparing a top tier model with a smaller one. Very curious to see whether there will be an Opus 4.5 that does even better.

I&#x27;m sure this is a very impressive model, but gemini-3-pro-preview is failing spectacularly at my fairly basic python benchmark. In fact, gemini-2.5-pro gets a lot closer (but is still wrong).For reference: gpt-5.1-thinking passes, gpt-5.1-instant fails, gpt-5-thinking fails, gpt-5-instant fails, sonnet-4.5 passes, opus-4.1 passes (lesser claude models fail).This is a reminder that benchmarks are meaningless – you should always curate your own out-of-sample benchmarks. A lot of people are going to say &quot;wow, look how much they jumped in x, y, and z benchmark&quot; and start to make some extrapolation about society, and what this means for others. Meanwhile.. I&#x27;m still wondering how they&#x27;re still getting this problem wrong.edit: I&#x27;ve a lot of good feedback here. I think there are ways I can improve my benchmark.

Simon says if he gets a suspiciously good result he&#x27;ll just try a bunch of other absurd animal&#x2F;vehicle combinations to see if they trained a special case: <a href="https:&#x2F;&#x2F;simonwillison.net&#x2F;2025&#x2F;Nov&#x2F;13&#x2F;training-for-pelicans-riding-bicycles&#x2F;" rel="nofollow">https:&#x2F;&#x2F;simonwillison.net&#x2F;2025&#x2F;Nov&#x2F;13&#x2F;training-for-pelicans-...</a>

I have tried combinations of hard to draw vehicle and animals (crocodile, frog, pterodactly, riding a hand glider, tricycle, skydiving), and it did a rather good job in every cases (compared to previous tests). Whatever they have done to improve on that point, they did it in a way that generalise.

That is pretty impressive.So impressive it makes you wonder if someone has noticed it being used a benchmark prompt.

It hadn&#x27;t occurred to me until now that the pelican could overcome the short legs issue by not sitting on the seat and instead put its legs inside the frame of the bike. That&#x27;s probably closer to how a real pelican would ride a bike, even if it wasn&#x27;t deliberate.

What we have all been waiting for:&quot;Create me a SVG of a pelican riding on a bicycle&quot;<a href="https:&#x2F;&#x2F;www.svgviewer.dev&#x2F;s&#x2F;FfhmhTK1" rel="nofollow">https:&#x2F;&#x2F;www.svgviewer.dev&#x2F;s&#x2F;FfhmhTK1</a>

Interesting that they added an option to select your own API key right in AI studio‘s input field.
I sincerely hope the times of generous free AIstudio usage are not over

This is a list of questions and answers that was created by different people.The questions AND the answers are public.If the LLM manages through reasoning OR memory to repeat back the answer then they win.The scores represent the % of correct answers they recalled.

I estimate another 7 months before models start getting 115% on Humanity&#x27;s Last Exam.

Every time I see a table like this numbers go up. Can someone explain what this actually means? Is there just an improvement that some tests are solved in a better way or is this a breakthrough and this model can do something that all others can not?

This is easily shown that the numbers are for GPT 5.1 thinking high.Just go to the leaderboard website and see for yourself: <a href="https:&#x2F;&#x2F;arcprize.org&#x2F;leaderboard" rel="nofollow">https:&#x2F;&#x2F;arcprize.org&#x2F;leaderboard</a>

If you believe another thread the benchmarks are comparing Gemini-3 (probably thinking) to GPT-5.1 without thinking.The person also claims that with thinking on the gap narrows considerably.We&#x27;ll probably have 3rd party benchmarks in a couple of days.

Supposedly this is the model card. Very impressive results.<a href="https:&#x2F;&#x2F;pbs.twimg.com&#x2F;media&#x2F;G6CFG6jXAAA1p0I?format=jpg&amp;name=medium" rel="nofollow">https:&#x2F;&#x2F;pbs.twimg.com&#x2F;media&#x2F;G6CFG6jXAAA1p0I?format=jpg&amp;name=...</a>Also, the full document:<a href="https:&#x2F;&#x2F;archive.org&#x2F;details&#x2F;gemini-3-pro-model-card&#x2F;page&#x2F;n3&#x2F;mode&#x2F;2up" rel="nofollow">https:&#x2F;&#x2F;archive.org&#x2F;details&#x2F;gemini-3-pro-model-card&#x2F;page&#x2F;n3&#x2F;...</a>

&quot;Gemini 3 Pro Preview&quot; is in Vertex

Can’t wait to test it out. Been running a tons of benchmarks (1000+ generations) for my AI to CAD model project and noticed:- GPT-5 medium is the best- GPT-5.1 falls right between Gemini 2.5 Pro and GPT-5 but it’s quite a bit fasterReally wonder how well Gemini 3 will perform

IMO it&#x27;s not about art, but a completely different path than all these images are going down. The pelican needs tools to ride the bike, or a modified bike. Maybe a recumbent?

Some time I think I should spend $50 on Upwork to get a real human artist to do it first to know what is that we&#x27;re going for. What a good pelican riding a bicycle SVG is actually looking like?

i think anything that makes it clear they&#x27;ve done that would be a lot worse PR than failing the pelican test would ever be.

At this point I&#x27;m surprised they haven&#x27;t been training on thousands of professionally-created SVGs of pelicans on bicycles.

The blue lines indicating wind really sell it.

It’s a good pelican. Not great but good.

Did you notice that this embedded a Gemini API connection within the app itself? Or am I not understanding what that is?

Great improvement by only adding one feedback prompt:
Change the rotation axis of the wheels by 90 degrees in the horizontal plane. Same for the legs and arms<a href="https:&#x2F;&#x2F;aistudio.google.com&#x2F;app&#x2F;prompts?state=%7B%22ids%22:%5B%221LqaF4gknC3yji4uI7uWTjjAo7liZEZ-p%22%5D,%22action%22:%22open%22,%22userId%22:%22115978321886005652329%22,%22resourceKeys%22:%7B%7D%7D&amp;usp=sharing" rel="nofollow">https:&#x2F;&#x2F;aistudio.google.com&#x2F;app&#x2F;prompts?state=%7B%22ids%22:%...</a>

2D SVG is old news. Next frontier is animated 3D. One shot shows there&#x27;s still progress to be made: <a href="https:&#x2F;&#x2F;aistudio.google.com&#x2F;apps&#x2F;drive&#x2F;1XA4HdqQK5ixqi1jD9uMgmGUoxBeu8fFX?showPreview=true&amp;showAssistant=true" rel="nofollow">https:&#x2F;&#x2F;aistudio.google.com&#x2F;apps&#x2F;drive&#x2F;1XA4HdqQK5ixqi1jD9uMg...</a>

Pelican riding a bicycle: <a href="https:&#x2F;&#x2F;pasteboard.co&#x2F;CjJ7Xxftljzp.png" rel="nofollow">https:&#x2F;&#x2F;pasteboard.co&#x2F;CjJ7Xxftljzp.png</a>

When will they allow us to use modern LLM samplers like min_p, or even better samplers like top N sigma, or P-less decoding? They are provably SOTA and in some cases enable infinite temperature.Temperature continues to be gated to maximum of 0.2, and there&#x27;s still the hidden top_k of 64 that you can&#x27;t turn off.I love the google AI studio, but I hate it too for not enabling a whole host of advanced features. So many mixed feelings, so many unanswered questions, so many frustrating UI decisions on a tool that is ostensibly aimed at prosumers...

I just tried &quot;analyze this audio file recording of a meeting and notes along with a transcript labeling all the speakers&quot; (using the language from the parent&#x27;s comment) and indeed Gemini 3 was significantly better than 2.5 Pro.3 created a great &quot;Executive Summary&quot;, identified the speakers&#x27; names, and then gave me a second by second transcript:<pre><code> [00:00] Greg: Hello.
 [00:01] X: You great?
 [00:02] Greg: Hi.
 [00:03] X: I&#x27;m X.
 [00:04] Y: I&#x27;m Y.
 ...
</code></pre>
Super impressive!

I made a simple webpage to grab text from YouTube videos:
<a href="https:&#x2F;&#x2F;summynews.com" rel="nofollow">https:&#x2F;&#x2F;summynews.com</a>
Great for this kind of testing?
(want to expand to other sources in the long run)

Yes, this is the best solution for that goal. Use the MacWhisper app + Parakeet 3.

Parakeet TDT v3 would be really good at that

It&#x27;s not even THAT hard. I am working on a side project that gets a podcast episode and then labels the speakers. It works.

Agreed. I don’t see the need for Gemini to be able to do this task, although it should be able to offload it to another model.

I&#x27;d do the transcript and the summary parts separately. Dedicated audio models from vendors like ElevenLabs or Soniox use speaker detection models to produce an accurate speaker based transcript while I&#x27;m not necessarily sure that Google&#x27;s models do so, maybe they just hallucinate the speakers instead.

The worst when it fails to eat simple pdf documents and lies and gas lights in an attempt to cover it up. Why not just admit you can’t read the file?

I wonder if you could get around this with a slightly more sophisticated harness. I suspect you&#x27;re running into context length issues.Something like1.) Split audio into multiple smaller tracks.
2.) Perform first pass audio extraction
3.) Find unique speakers and other potentially helpful information (maybe just a short summary of where the conversation left off)
4.) Seed the next stage with that information (yay multimodality) and generate the audio transcript for itObviously it would be ideal if a model could handle the ultra long context conversations by default, but I&#x27;d be curious how much error is caused by a lack of general capability vs simple context pollution.

Now try an actual speech model like ElevenLabs or Soniox, not something not made for it.

My audio experiment was much less successful — I uploaded a 90-minute podcast episode and asked it to produce a labeled transcript. Gemini 3:- Hallucinated at least three quotes (that I checked) resembling nothing said by any of the hosts- Produced timestamps that were almost entirely wrong. Language quoted from the end of the episode, for instance, was timestamped 35 minutes into the episode, rather than 85 minutes.- Almost all of what is transcribed is heavily paraphrased and abridged, in most cases without any indication.Understandable that Gemini can&#x27;t cope with such a long audio recording yet, but I would&#x27;ve hoped for a more graceful&#x2F;less hallucinatory failure mode. And unfortunately, aligns with my impression of past Gemini models that they are impressively smart but fail in the most catastrophic ways.

My favorite benchmark is to analyze a very long audio file recording of a management meeting and produce very good notes along with a transcript labeling all the speakers. 2.5 was decently good at generating the summary, but it was terrible at labeling speakers. 3.0 has so far absolutely nailed speaker labeling.

Reading the introductory passage - all I can say now is, Ai is here to stay.

Also recently: 
Code Wiki: <a href="https:&#x2F;&#x2F;codewiki.google&#x2F;" rel="nofollow">https:&#x2F;&#x2F;codewiki.google&#x2F;</a>

DeepMind page: <a href="https:&#x2F;&#x2F;deepmind.google&#x2F;models&#x2F;gemini&#x2F;" rel="nofollow">https:&#x2F;&#x2F;deepmind.google&#x2F;models&#x2F;gemini&#x2F;</a>Gemini 3 Pro DeepMind Page: <a href="https:&#x2F;&#x2F;deepmind.google&#x2F;models&#x2F;gemini&#x2F;pro&#x2F;" rel="nofollow">https:&#x2F;&#x2F;deepmind.google&#x2F;models&#x2F;gemini&#x2F;pro&#x2F;</a>Developer blog: <a href="https:&#x2F;&#x2F;blog.google&#x2F;technology&#x2F;developers&#x2F;gemini-3-developers&#x2F;" rel="nofollow">https:&#x2F;&#x2F;blog.google&#x2F;technology&#x2F;developers&#x2F;gemini-3-developer...</a>Gemini 3 Docs: <a href="https:&#x2F;&#x2F;ai.google.dev&#x2F;gemini-api&#x2F;docs&#x2F;gemini-3" rel="nofollow">https:&#x2F;&#x2F;ai.google.dev&#x2F;gemini-api&#x2F;docs&#x2F;gemini-3</a>Google Antigravity: <a href="https:&#x2F;&#x2F;antigravity.google&#x2F;" rel="nofollow">https:&#x2F;&#x2F;antigravity.google&#x2F;</a>

Curious to see it in action. Gemini 2.5 has already been very impressive as a study buddy for courses like set theory, information theory, and automata.
Although I’m always a bit skeptical of these benchmarks. Seems quite unlikely that all of the questions remain out of their training data.

They are not even leading in Terminal-Bench... GPT 5.1-codex is better than Gemini 3 Pro

Yeah, they mention a benchmark I&#x27;m seeing the first time (Terminal-Bench 2.0) and are supposedly leading in, while for some reason SWE Bench is down from Sonnet 4.5.Curious to see some third-party testing of this model. Currently it seems to primarily improve of &quot;general non-coding and visual reasoning&quot; primarily, based on the benchmarks.

claude 4.5 gets 82% on their own highly customized scaffolding. (parallel compute with a scoring function). That beats Doubao

SWEBench-Verified is probably benchmaxxed at this stage. Claude isn&#x27;t even the top performer, that honor goes to Doubao [1].Also, the confidence interval for a such a small dataset is about 3 percent points, so these differences could just be up to chance.[1] <a href="https:&#x2F;&#x2F;www.swebench.com&#x2F;" rel="nofollow">https:&#x2F;&#x2F;www.swebench.com&#x2F;</a>

Thus far, this is one of the best objective evaluations of real world software engineering...

Why is this particular benchmark important?

They overly emphasize tasks with small context without noise and red herrings in the context.

Honestly, I am inclined to think a lot of the people who are wowed by benchmarks and simple tech demos probably aren&#x27;t doing very much at their day job and if they&#x27;re either working on simple codebases or ones that don&#x27;t have very many users(more users == more bugs found). When you throw these models at complex software projects like SOAs, big object-oriented codebases, etc. their output can be totally unusable.

Does anyone trust benchmarks at this point? Genuine question. Isn&#x27;t the scientific consensus that they are broken and poor evaluation tools?

&gt;&quot;It doesn&#x27;t matter, the real benchmark is taking the community temperature on the model after a few weeks of usage.&quot;Indeed. It&#x27;s almost impossible to truly know a model before spending a few million tokens on a real world task. It will take a step-change level advancement at this point for me to trust anything but Claude right now.

Imho Gemini 2.5 was by far the better model on non-trivial tasks.

It doesn&#x27;t matter, the real benchmark is taking the community temperature on the model after a few weeks of usage.

I mean... it achieved 76.2% vs the leader (Claude Sonnet) at 77.2%.That&#x27;s a &quot;loss&quot; I can deal with.

Understanding precisely why Gemini 3 isn&#x27;t front of the pack on SWE Bench is really what I was hoping to understand here. Especially for a blog post targeted at software developers...

Is it just me or is that link broken because of the cloudflare outage?Edit: nvm it looks to be up for me again

Is it censored... or just biased towards edge-lord MechaHitler nonsense whenever Musk feels like tinkering with the system prompt?

Grok got to hold the top spot of LMArena-text for all of ~24 hours, good for them [1]. With stylecontrol enabled, that is. Without stylecontrol, gemini held the fort.[1] <a href="https:&#x2F;&#x2F;lmarena.ai&#x2F;leaderboard&#x2F;text" rel="nofollow">https:&#x2F;&#x2F;lmarena.ai&#x2F;leaderboard&#x2F;text</a>

I thought it was a workaround to Google&#x27;s complete disinterest in productizing the AI research it was doing and publishing, rather than a way to balance their dominance in a market which didn&#x27;t meaningfully exist.

Pffft. OpenAI was conceived to be Open, too.

Balance is too weak of a word. OpenAI was conceived specifically to prevent Google from getting AGI first. That was its original goal. At the time of its founding Google was the undisputed leader of AI anywhere in the world. Musk was then very worried about AGI being developed behind closed doors particularly Google, which was why he was the driving force behind the founding of OpenAI.

Ironically, OpenAI was conceived as a way to balance Google&#x27;s dominance in AI.

Google always has been there, its just that many didn&#x27;t realize that DeepMind even existed and I said that they needed to be put to commercial use years ago. [0] and Google AI != DeepMind.You are now seeing their valuation finally adjusting to that fact all thanks to DeepMind finally being put to use.[0] <a href="https:&#x2F;&#x2F;news.ycombinator.com&#x2F;item?id=34713073">https:&#x2F;&#x2F;news.ycombinator.com&#x2F;item?id=34713073</a>

Yes, this is correct, and it happens everywhere. App Store, Play Store, YouTube, Meta, X, Amazon and even Uber - they all play in two-sided markets exploiting both its users and providers at the same time.

They&#x27;re not a moral entity. corporations aren&#x27;t people.I think a lot of the harms you mentioned are real, but they&#x27;re a natural consequence of capitalistic profit chasing. Governments are supposed to regulate monopolies and anti-consumer behavior like that. Instead of regulating surveillance capitalism, governments are using it to bypass laws restricting their power.If I were a google investor, I would absolutely want them to defeat ad-blocking, ban yt-dlp, dominate the ad-market and all the rest of what you said. In capitalism, everyone looks out for their own interests, and governments ensure the public isn&#x27;t harmed in the process. But any time a government tries to regulate things, the same crowd that decries this oppose government overreach.Voters are people and they are moral entities, direct any moral outrage at us.

This is an understandable, but simplistic way of looking at the world. Are you also gonna blame Apple for mining for rare earths, because they made a successful product that requires exotic materials which needs to be mined from earth? How about hundreds of thousands of factory workers that are being subjected to inhumane conditions to assemble iPhones each year?For every &quot;OMG, internet is filled with ads&quot;, people are conveniently forgetting the real-world impact of ALL COMPANIES (and not just Apple) btw. Either you should be upset with the system, and not selectively at Google.

Suppressed wages to colluding with Apple to not poach.

&gt; They&#x27;ve poisoned the internetAnd what of the people that ravenously support ads and ad-supported content, instead of paying?What of the consumptive public? Are they not responsible for their choices?I do not consume algorithmic content, I do not have any social media (unless you count HN for either).You can&#x27;t have it both ways. Lead by example, stop using the poison and find friends that aren&#x27;t addicted. Build an offline community.

People love getting their content for free and that&#x27;s what Google does.Even 25 years ago people wouldn&#x27;t even believe Youtube exists. Anyone can upload whatever they want, however often they want, Youtube will be responsible for promoting it, they&#x27;ll provide to however many billions users want to view it, and they&#x27;ll pay you 55% of the revenue it makes?

What kind of world do you live in? Actually Google ads tend to be some of the highest ROI for the advertiser and most likely to be beneficial for the user. Vs the pure junk ads that aren&#x27;t personalized, and just banner ads that have zero relationship to me. Google Ads is the enabler of free internet. I for one am thankful to them. Else you end up paying for NYT, Washinton Post, Information etc -- virtually for any high quality web site (including Search).

They&#x27;ve poisoned the internet with their monopoly on advertising, the air pollution of the online world, which is an transgression that far outweighs any good they might have done.  Much of the negative social effects of being online come from the need to drive more screen time, more engagement, more clicks, and more ad impressions firehosed into the faces of users for sweet, sweet, advertiser money.  When Google finally defeats ad-blocking, yt-dlp, etc., remember this.

frankly when compared against TikTok, Insta, etc, YouTube is a force for good. Just script the shorts away...

Why stop at YouTube? Blame Apple for creating an additive gadget that has single handedly wasted billions of hours of collective human intelligence. Life was so much better before iPhones.But I hear you say - you can use iPhones for productive things and not just mindless brainrot. And that&#x27;s the same with YouTube as well. Many waste time on YouTube, but many learn and do productive things.Dont paint everything with a single, large, coarse brush stroke.

Forgot to mention absolutely milking every ounce of their users attention with Youtube, plus forcing Shorts!

&gt;most of those examples are acquisitionsTaking those products from where there were to the juggernauts they are today was not guaranteed to succeed, nor was it easy. And yes plenty of innovation happened with these products post aquisition.

For what it&#x27;s worth, most of those examples are acquisitions. That&#x27;s not a hit against Google in particular. That&#x27;s the way all big tech co&#x27;s grow. But it&#x27;s not necessarily representative of &quot;innovation.&quot;

&gt; Drive vs WordYou mean Drive vs OneDrive or, maybe Docs vs Word?

Google is using the typical monopoly playbook as most other large orgs, and the world would be a &quot;better place&quot; if they are kept in check.But at least this company is not run by a narcissistic sociopath.

All those examples date back to the 2000s. Android has seen some significant improvements, but everything else has stagnated if not enshittified- remember when google told us not to ever worry about deleting anything?- and then started backing up my photos without me asking and are now constantly nagging me to pay them a monthly fee?They have done a lot, but most of it was in the &quot;don&#x27;t be evil&quot; days and they are a fading memory.

If you consider surveillance capitalism and dark pattern nudges a good thing, then sure. Gemini has the potential to obliterate their current business model completely so I wouldn&#x27;t consider that &quot;waking up&quot;.

Something about bringing balance to the force not destroying it.

Outlook is not better in ways that email or gmail users necessarily care about, and in my experience gets in the way more than it helps with productivity or anything it tries to be good at. I&#x27;ve used it in office settings because it&#x27;s the default, but never in my life have I considered using it by choice. If it&#x27;s better, it might not matter.

Outlook is much better than Gmail and so is the office suite.It&#x27;s good there&#x27;s competition in the space though.

Seriously? Google is an incredibly evil company whose net contribution to society is probably only barely positive thanks to their original product (search). Since completely de-googling I&#x27;ve felt a lot better about myself.

I&#x27;ve been so happy to see Google wake up.Many can point to a long history of killed products and soured opinions but you can&#x27;t deny theyve been the great balancing force (often for good) in the industry.- Gmail vs Outlook- Drive vs Word- Android vs iOS- Worklife balance and high pay vs the low salary grind of before.Theyve done heaps for the industry. Im glad to see signs of life. Particularly in their P&#x2F;E which was unjustly low for awhile.

It also tops LMSYS leaderboard across all categories. However knowledge cutoff is Jan 2025. I do wonder how long they have been pre-training this thing :D.

Op is right.<a href="https:&#x2F;&#x2F;imgcdn.stablediffusionweb.com&#x2F;2024&#x2F;4&#x2F;19&#x2F;8e54925a-0048-4f5a-88d4-aef3eedfd93e.jpg" rel="nofollow">https:&#x2F;&#x2F;imgcdn.stablediffusionweb.com&#x2F;2024&#x2F;4&#x2F;19&#x2F;8e54925a-004...</a>For the above pic I asked &quot;What is wrong with the image?&quot;Result:
- It totally missed the most obvious one - six fingers.Instead it said this:Based on a close inspection of the image, here are the things that are &quot;wrong&quot; or anomalous, indicating that it is likely an AI-generated image:- The Thumb Anatomy: This is the most obvious error. The digit in the thumb&#x27;s position (far left) looks exactly like a long index finger. A natural human thumb has two segments (phalanges) and is shorter and thicker, opposing the other fingers. This digit appears to have three segments and is roughly the same length and shape as the other fingers, giving the hand the appearance of having five fingers and no thumb.- Unnatural Skin Texture: The skin is unnaturally smooth, waxy, and free of normal human imperfections like distinct pores, fine wrinkles, or hair. It has a &quot;plastic&quot; or &quot;airbrushed&quot; look typical of generative AI.- Palm Lines: The creases on the palm are vague and generic. They don&#x27;t follow the distinct pattern of major flexion creases (like the heart, head, and life lines) found on a real human hand.- Wrist Connection: The connection between the hand and the wrist looks slightly distorted, with tendons that appear too prominent and smooth for the pose.[Edit: 3.0 is same as 2.5 - both answered almost identically]

ah interesting. I wonder if this is a &quot;safety guardrails blindspot&quot; due to the placement.

I strongly suspect it&#x27;s a tokenization problem. Text and symbols fit nicely in tokens, but having something like a single &quot;dog leg&quot; token is a tough problem to solve.

Why should it have to be expensive computationally? How do brains do it with such a low amount of energy? I think catching the brain abilities even of a bug might be very hard, but that does not mean that there isn&#x27;t a way to do it with little computational power. It requires having the correct structures&#x2F;models&#x2F;algorithms or whatever is the precise jargon.

Perception seems to be one of the main constraints on LLMs that not much progress has been made on. Perhaps not surprising, given perception is something evolution has worked on since the inception of life itself. Likely much, much more expensive computationally than it receives credit for.

The point of benchmarking that is checking for hallucinations and overfitting. Does the model actually check the picture to count the legs or does it just see it&#x27;s a dog and answer four because it knows dogs usually has four legs?It&#x27;s a perfectly valid benchmark and very telling.

&quot;[a photoshopped picture of a dog with 5 legs]...please count the legs&quot;Meanwhile you could benchmark for something actually useful. If you&#x27;re about to say &quot;But that means it won&#x27;t work for my use case of identifying a person on a live feed&quot; or whatever, then why don&#x27;t you test that? I really don&#x27;t understand the kick people get of successfully tricking LLMs on non productive task with no real world application. Just like the &quot;how many r in strawberry?&quot;, &quot;uh uh uh it says two urh urh&quot;.. ok but so what? What good is a benchmark that is so far from a real use case?

In reality it used the term &quot;male anatomy&quot; heh

&gt; Gemini 3 however, while still failing, at least recognized the 5th leg, but thought the dog was...well endowed.I see that AI is reaching the level of a middle school boy...

It still failed my image identification test ([a photoshopped picture of a dog with 5 legs]...please count the legs) that so far every other model has failed agonizingly, even failing when I tell them they are failing, and they tend to fight back at me.Gemini 3 however, while still failing, at least recognized the 5th leg, but thought the dog was...well endowed. The 5th leg however is clearly a leg, despite being where you would expect the dogs member to be. I&#x27;ll give it half credit for at least recognizing that there was something there.Still though, there is a lot of work that needs to be done on getting these models to properly &quot;see&quot; images.

I&#x27;m not sure how concerned people should be at the trend lines. If you&#x27;re building a product that already works well, you shouldn&#x27;t feel the need to upgrade to a larger parameter model. If your product doesn&#x27;t work and the new architectures unlock performance that would let you have a feasible business, even a 2x on input tokens shouldn&#x27;t be the dealbreaker.If we&#x27;re paying more for a more petaflop heavy model, it makes sense that costs would go up. What really would concern me is if companies start ratcheting prices up for models with the same level of performance. My hope is raw hardware costs and OSS releases keep a lid on the margin pressure.

Every recent release has bumped the pricing significantly. If I was building a product and my margins weren’t incredible I’d be concerned. The input price almost doubled with this one.

Pretty happy the under 200k token pricing is staying in the same ballpark as Gemini 2.5 Pro:Input: $1.25 -&gt; $2.00 (1M tokens)Output: $10.00 -&gt; $12.00Squeezes a bit more margin out of app layer companies, certainly, but there&#x27;s a good chance that for tasks that really require a sota model it can be more than justified.

They&#x27;re a bit less bad than they used to be. I&#x27;m not exactly happy about what this means to incentives (and rewards) for doing research and writing good content, but sometimes I ask a dumb question out of curiosity and Google overview will give it to me (e.g. &quot;what&#x27;s in flower food?&quot;). I don&#x27;t need GPT 5.1 Thinking for that.

&quot;Since then, it’s been incredible to see how much people love it. AI Overviews now have 2 billion users every month.&quot;Cringe. To get to 2 billion a month they must be counting anyone who sees an AI overview as a user. They should just go ahead and claim the &quot;most quickly adopted product in history&quot; as well.

You&#x27;re right, I&#x27;m probably being a little uncharitable!Normal users (i.e. not grumpy techies ;) ) probably just go with the flow rather than finding it irritating.

Maybe you ignore it, but Google has stated in the past that click-through rates with AI overviews are way down. To me, that implies the &#x27;user&#x27; read the summary and got what they needed, such that they didn&#x27;t feel the need to dig into a further site (ignoring whether that&#x27;s a good thing or not).I&#x27;d be comfortable calling a &#x27;user&#x27; anyone who clicked to expand the little summary. Not sure what else you&#x27;d call them.

&quot;AI Overviews now have 2 billion users every month.&quot;&quot;Users&quot;? Or people that get presented with it and ignore it?

i&#x27;m very doubtful gmail mails are used to train the model by default, because emails contain private data and as soon as this private data shows up in the model output, gmail is done.&quot;gmail being read by gemini&quot; does NOT mean &quot;gemini is trained on your private gmail correspondence&quot;. it can mean gemini loads your emails into a session context so it can answer questions about your mail, which is quite different.

The real question is, &quot;For how long?&quot;

Isn&#x27;t Gmail covered under the Workspace privacy policy which forbids using that for training data. So I&#x27;m guessing that&#x27;s excluded by the &quot;in accordance&quot; clause.

This seems like a dubious conclusion. I think you missed this part:&gt; in accordance with Google’s relevant terms of service, privacy policy

I&#x27;m pretty sure they mention in their various TOSes that they don&#x27;t train on user data in places like Gmail.That said, LLMs are the most data-greedy technology of all time, and it wouldn&#x27;t surprise me that companies building them feel so much pressure to top each other they &quot;sidestep&quot; their own TOSes. There are plenty of signals they are already changing their terms to train when previously they said they wouldn&#x27;t--see Anthropic&#x27;s update in August regarding Claude Code.If anyone ever starts caring about privacy again, this might be a way to bring down the crazy AI capex &#x2F; tech valuations. It is probably possible, if you are a sufficiently funded and motivated actor, to tease out evidence of training data that shouldn&#x27;t be there based on a vendor&#x27;s TOS. There is already evidence some IP owners (like NYT) have done this for copyright claims, but you could get a lot more pitchforks out if it turns out Jane Doe&#x27;s HIPAA-protected information in an email was trained on.

By the year 2025 I think most of the HN regulars and IT people in general are so jaded regarding privacy that it is not even surprising anyone. I suspect all gmails were analyzed and read from the beginning of google age, so nothing really changed, they might as well just admit it.Google is betting that moving email and cloud is such a giant hassle that almost no one will do it, and ditching YT and Maps is just impossible.

I expect almost no-one to read the Gemini 3 model card. But here is a damning excerpt from the early leaked model card from [0]:&gt; The training dataset also includes: publicly available datasets that are readily downloadable; data obtained by crawlers; licensed data obtained via commercial licensing agreements; user data (i.e., data collected from users of Google products and services to train AI models, along with user interactions with the model) in accordance with Google’s relevant terms of service, privacy policy, service-specific policies, and pursuant to user controls, where appropriate; other datasets that Google acquires or generates in the course of its business operations, or directly from its workforce; and AI-generated synthetic data.So your Gmails are being read by Gemini and is being put on the training set for future models. Oh dear and Google is being sued over using Gemini for analyzing user&#x27;s data which potentially includes Gmails by default.Where is the outrage?[0] <a href="https:&#x2F;&#x2F;web.archive.org&#x2F;web&#x2F;20251118111103&#x2F;https:&#x2F;&#x2F;storage.googleapis.com&#x2F;deepmind-media&#x2F;Model-Cards&#x2F;Gemini-3-Pro-Model-Card.pdf" rel="nofollow">https:&#x2F;&#x2F;web.archive.org&#x2F;web&#x2F;20251118111103&#x2F;https:&#x2F;&#x2F;storage.g...</a>[1] <a href="https:&#x2F;&#x2F;www.yahoo.com&#x2F;news&#x2F;articles&#x2F;google-sued-over-gemini-ai-151508072.html" rel="nofollow">https:&#x2F;&#x2F;www.yahoo.com&#x2F;news&#x2F;articles&#x2F;google-sued-over-gemini-...</a>

It’s not a format, but in my mind it implies designs that are supposed to be functional as opposed to models that are meant for virtual games.It generated a blender script that makes the model.

When I see CAD, I always think of Casting Assistant Device.

I&#x27;m not familiar enough with CAD what type of format is it?

Based on my benchmarks (run 100s of model generations).2.5 stands between GPT-5 and GPT-5.1, where GPT-5 is the best of the 3.In preliminary evals Gemini 3 seems to be way better than all, but I will know when I run extended benchmarks tonight.

Sorry to ask 7 days late, but what sort of prompt do you use to get it to do it? I tried the same exercise but it just placed the image in 2D in the 3D world. Much like Paper Mario but not what I was going for! Thank you.

Just hand sketched what 5 year old would do on the paper - the house, trees, sun. And asked to generate 3d model with tree.js.Results are amazing! 2.5 and 3 seems way way head.

Just generated a bunch of 3D CAD models using Gemini 3.0 to see how it compares in spatial understanding and it&#x27;s heaps better than anything currently out there - not only intelligence but also speed.Will run extended benchmarks later, let me know if you want to see actual data.

Yeah, at this point I want to see the failure modes. Show me at least as many cases where it breaks. Otherwise, I&#x27;ll assume it&#x27;s an advertisement and I&#x27;ll skip to the next headline. I&#x27;m not going to waste my time on it anymore.

People would have had a heart attack if they saw this 5 years ago for the first time. Now artificial brains are “meh” :)

Our ability to adapt to new things is both a blessing and a curse.

It is. But understandably the people who need to push back on what is still magic may get a bit tired.

I think it&#x27;s fun to see what is not even considered magic anymore today.

Why?Not trying to challenge you, and I&#x27;d sincerely love to read your response. People said similar things about previous gen-AI tool announcements that proved over time to be overstated. Is there some reason to put more weight in &quot;what people on HN said&quot; in this case, compared to previous situations?

Only reasonable thing is to not listening to anyone who seem to be hyping anything, LLMs or otherwise. Wait until the thing gets released, run your private benchmarks against it, get a concrete number, compare against existing runs you&#x27;ve done before.I don&#x27;t see any other way of doing this. The people who keep reading and following comments either here on HN, from LocalLlama or otherwise will continue to be misinformed by all the FUD and guerilla marketing that is happening across all of these places.

No LLM has ever been as good as people said it was. That doesn&#x27;t mean this one won&#x27;t be, but it does make it an unlikely bet based on past trends.

I agree but if Gemini 3 is as good as people on HN said about the preview, then this is the wrong announcement to sleep on.

it&#x27;s not AI fatigue, its that you just need to shift mode to not pay attention too much to the latest and greatest as they all leap frog each other each month.  Just stick to one and ride it thru ups and downs.

And by this time next year, this comment is going to look very silly

My test for the state of AI is &quot;Does Microsoft Teams still suck?&quot;, if it does still suck, then clearly the AIs were not capable of just fixing the bugs and we must not be there yet.

I think I am in this AI fatigue phase. I am past all hype with models, tools and agents and back to problem and solution approach, sometimes code gen with AI , sometimes think and ask for a piece of code. But not offloading to AI and buying all the bs, waiting it to do magic with my codebase.

And you&#x27;re implying they&#x27;re being 100% truthful?Marketing is always somewhere in the middle

Gemini app != Google search.You&#x27;re implying they&#x27;re lying?

I don&#x27;t know for sure but they have to be counting users like me whose phone has had Gemini force installed on an update and I&#x27;ve only opened the app by accident while trying to figure out how to invoke the old actually useful Assistant app

&gt; it&#x27;s still a choice that users are making to go there rather than being an existing product they happen to user otherwise.Yes and no, my power button got remapped to opening Gemini in an update...I removed that but I can imagine that your average user doesn&#x27;t.

&gt; Gemini app surpasses 650 million users per monthUnless these numbers are just lies, I&#x27;m not sure how this is &quot;pushing their AI into everything they can&quot;. Especially on iOS where every user is someone who went to App Store and downloaded it. Admittedly on Android, Gemini is preinstalled these days but it&#x27;s still a choice that users are making to go there rather than being an existing product they happen to user otherwise.Now OTOH &quot;AI overviews now have two billion users&quot; can definitely be criticised in the way you suggest.

&gt; The Gemini app surpasses 650 million users per month, more than 70% of our Cloud customers use our AI, 13 million developers have built with our generative models, and that is just a snippet of the impact we’re seeingNot to be a negative nelly, but these numbers are definitely inflated due to Google literally pushing their AI into everything they can, much like M$. Can&#x27;t even search google without getting an AI response. Surely you can&#x27;t claim those numbers are legit.

<a href="https:&#x2F;&#x2F;blog.google&#x2F;technology&#x2F;developers&#x2F;gemini-3-developers&#x2F;" rel="nofollow">https:&#x2F;&#x2F;blog.google&#x2F;technology&#x2F;developers&#x2F;gemini-3-developer...</a><a href="https:&#x2F;&#x2F;aistudio.google.com&#x2F;prompts&#x2F;new_chat?model=gemini-3-pro-preview" rel="nofollow">https:&#x2F;&#x2F;aistudio.google.com&#x2F;prompts&#x2F;new_chat?model=gemini-3-...</a>