I'm not at all an industry pundit. But I suspect there's a reason we're not seeing leading models from Google recently.
Judging from my own frustrating attempts to use Gemini for vibe-coding, it seems like Google is badly over-sold (i.e., under-provisioned).
From all those promos giving away their pro-level subscription with phones; spinning up a mid-level subscription to undercut other providers and (probably most significantly) putting AI queries into ever search response because their flagship search product had become useless; they're promising a lot more processing to customers than they can reliably deliver.
The recent iterations seem to be intended not to push the capabilities forward, but to deliver capabilities at the current level while consuming less resources. That will allow them to maintain their trajectory until (I'm expecting) they get the huge infusion of extra compute resources from Space X later this year.
If I'm right, then I expect we should see Google start pushing forward again (rather than more of this lateral stuff) by the end of the year.
There's no way these are using a significant amount of compute. I'm not 100% convinced they're actually LLM-generated rather than an old-school Markov model. Both the relevance and accuracy numbers of the responses flirt with 0%. It's possible they just have a few million stashed responses and choose one at random, from what I can tell as a user.
This model is not for builders and engineers. DeepSWE score of 49% is behind gpt 5.4 and muse spark. It's clearly intended to be an efficient model for google gemini usage.
What is interesting is how this is announced before any Gemini Pro progress. From the outside it seems as though Google cannot keep up with other frontier models.
Everyone wants to announce as late as possible (i.e., last) to chart the highest. Google is in a position financially to take a hit for these last few months.
I assumed google would lean into the efficiency stuff more and try to eat the easy 80% of workloads, winning market share on volume instead of frontier if they were not able to produce frontier level models.
They're very well equipped to be the volume discount store of inference.
Just as ChatGPT is now called "Chat" by the kids (go ahead and finish grinding your teeth, I'll wait...), if Antigravity takes off then it'll get an irksome nickname like Antigrav or just Grav.
I would be fine lauding Gemini models if the only benefit of them was superior understanding of intent (read between the lines). I don't need it to code because other models are tuned for that explicitly, but I would like a model that is tuned to produce less mechanical output.
I think there is another angle here - maybe, just maybe google doesn't want to release a pro/stronger model sooner because of two reasons- 1) they're afraid they'll have to get their hands dirty in a certain war?? 2) what's the benefit of being on top of this chart?
so I think they're just focussed on releasing whatever helps their bottomline (improving search?). its not like anthropic and openai are making a lot of money being on the top of charts! this might just be my crazy pills talking though.
It’s frankly embarrassing at this point. I’ve got free access through buying a Pixel phone and it’s not even worth using as it’s a waste of my time. Here’s my experience so far using it for basic sysadmin Linux type stuff.
Gemini 3.1 Pro just feels a generation behind, from when models would miss easy things and make bad assumptions. Its not actively detrimental in bad way but the opportunity cost vs using something like Opus to be productive is large.
Gemini 3.5 Flash is the most annoying model I have ever used. It loves to respond in ALL CAPS like “LOOK AT THAT” for no apparent reason. I realize it’s a flash model but I will give it a basic list of tasks and the output will simply vomit “Now I will X” “Now I will Y” “Now I will Z” over and over again filling my screen with garbage. It’s also no smarter than 3.1 Pro and consumes just as many tokens as 3.1 Pro, it’s really pointless without a newer Pro model in place.
3.1 Pro: while it's coding performance is mediocre, a lot of coding work requires minimum actual thinking. I use it often for light refactoring, boilerplate generation, testcase skeleton generation, code review, language questions ("is there a better way to write this code block?"). I don't have a corporation behind me so costs matter. Considering that I'm a Pro subscriber, it's quite cost effective.
3.5 Flash: excellent model for general (re)search. I use it for everyday tasks with Thinking instead of Flash-Lite. It's a much better version of Google Search for general queries like gaming tips, cooking, day-to-day first aid, tax and investment questions, etc.
Google is clearly aiming for cost-benefit here and considering that it gets bundled with YouTube Plus and Google One at $20/month, it's a killer deal IMHO.
PS: I don't work for Google and don't even like Google very much. But this is a good product.
It was a decent deal ~6-8 months ago. I had been using 3.1 pro almost since release, but it really is feeling old. After using other models more in the past two months though...I really can't go back to 3.1 pro, as I just have to explain my reasoning so damn much to get it on the right path, where as opus or fable just "get it" from the context of the project much better.
Sonnet is roughly the same level as 3.1 pro for me.
Of the models I've tried lately, I get more value out of Gemini for personal stuff like researching products or learning how to do things than the others. It seems more factually correct in domains I know about and hallucinates minor stuff less than the others.
I know that Flash is obviously no Opus or Fable. That being said, when I want a very quick response, my go-to is Gemini Flash. It's just really damn fast and generally tends to be accurate enough.
Yeah, Gemini Flash is my daily go to. It's certainly quite capable - yesterday it talked me though fixing my Ubuntu desktop (rebuilding NVIDIA driver) after an update caused it to boot into low resolution mode.
I don't like how it compared itself to 5.6 Luna instead of Sol, and still losing on some metrics. You will not see me using this over even 5.6 Sol-light
Claude's free usage is so limited that it's useless. Gemini is extremely generous - never hits a limit for me despite being used all day.
A couple of days ago I was comparing Claude vs Gemini asking both what the size of YACC generated parsing tables for ANSI C was. Claude was very gung-ho and rushed off to download a C grammar, install bison, and find out for itself (not really necessary), and then hit daily free usage limit half way through without even getting to an answer (despite this being the only thing I'd asked it that day).
I'm not at all an industry pundit. But I suspect there's a reason we're not seeing leading models from Google recently.
Judging from my own frustrating attempts to use Gemini for vibe-coding, it seems like Google is badly over-sold (i.e., under-provisioned).
From all those promos giving away their pro-level subscription with phones; spinning up a mid-level subscription to undercut other providers and (probably most significantly) putting AI queries into ever search response because their flagship search product had become useless; they're promising a lot more processing to customers than they can reliably deliver.
The recent iterations seem to be intended not to push the capabilities forward, but to deliver capabilities at the current level while consuming less resources. That will allow them to maintain their trajectory until (I'm expecting) they get the huge infusion of extra compute resources from Space X later this year.
If I'm right, then I expect we should see Google start pushing forward again (rather than more of this lateral stuff) by the end of the year.
> putting AI queries into ever search response
There's no way these are using a significant amount of compute. I'm not 100% convinced they're actually LLM-generated rather than an old-school Markov model. Both the relevance and accuracy numbers of the responses flirt with 0%. It's possible they just have a few million stashed responses and choose one at random, from what I can tell as a user.
Official announcement
https://blog.google/innovation-and-ai/models-and-research/ge...
This model is not for builders and engineers. DeepSWE score of 49% is behind gpt 5.4 and muse spark. It's clearly intended to be an efficient model for google gemini usage.
What is interesting is how this is announced before any Gemini Pro progress. From the outside it seems as though Google cannot keep up with other frontier models.
Who tested it on DeepSWE?
Edit: Oh it's in the other link
https://blog.google/innovation-and-ai/models-and-research/ge...
Everyone wants to announce as late as possible (i.e., last) to chart the highest. Google is in a position financially to take a hit for these last few months.
I assumed google would lean into the efficiency stuff more and try to eat the easy 80% of workloads, winning market share on volume instead of frontier if they were not able to produce frontier level models.
They're very well equipped to be the volume discount store of inference.
I find their models are pretty good for answering every day questions, and with their large user base, I’m sure they’re getting lots of usage.
They also provide the most usage for free and in the cheap paid plans.
They're cannibalizing their own search ad revenue. Very strange decision.
3 replies →
Model card: https://storage.googleapis.com/deepmind-media/Model-Cards/Ge...
It's available in Antigravity as well -
Antigravity is such a dumb name. An-ti-gra-vi-ty (5) is long, Ge-mi-ni (3) was better.
Aren’t Marketing/PR people at FAANG supposed to be among the best in the industry?
They actually paid [1] Lexicon Branding for this dumb ass name (probably millions of dollars).
[1] https://www.lexiconbranding.com/case-studies/google-antigrav...
Bard is still the best
Zune would like a word
Just as ChatGPT is now called "Chat" by the kids (go ahead and finish grinding your teeth, I'll wait...), if Antigravity takes off then it'll get an irksome nickname like Antigrav or just Grav.
3 replies →
just call it google grav
1 reply →
remember "kubernetes" ?
2 replies →
Not yet in Antigravity CLI 1.1.5 for Antigravity Business
I would be fine lauding Gemini models if the only benefit of them was superior understanding of intent (read between the lines). I don't need it to code because other models are tuned for that explicitly, but I would like a model that is tuned to produce less mechanical output.
It is both less intelligent and more expensive than GLM-5.2, while being closed weight.
Their niche is that you get the quality of a Chinese model for the price of an American model.
(Quoting myself from 2 days ago.)
I think there is another angle here - maybe, just maybe google doesn't want to release a pro/stronger model sooner because of two reasons- 1) they're afraid they'll have to get their hands dirty in a certain war?? 2) what's the benefit of being on top of this chart?
so I think they're just focussed on releasing whatever helps their bottomline (improving search?). its not like anthropic and openai are making a lot of money being on the top of charts! this might just be my crazy pills talking though.
Flash 3.5 does OK for various tasks. It's not super smart but is a workhorse and if 3.6 one is better than it, that's a positive for me.
I just saw it in agy and started using it without asking any questions. I did not notice any major difference.
Most conversation is taking place in this thread: https://news.ycombinator.com/item?id=48993414
https://blog.google/innovation-and-ai/models-and-research/ge...
at least openai had the guts to call code red and improve.
releasing a model worse than luna is pretty bad. Its clear that internally they did not decide coding was a thing until relatively recently.
It’s frankly embarrassing at this point. I’ve got free access through buying a Pixel phone and it’s not even worth using as it’s a waste of my time. Here’s my experience so far using it for basic sysadmin Linux type stuff.
Gemini 3.1 Pro just feels a generation behind, from when models would miss easy things and make bad assumptions. Its not actively detrimental in bad way but the opportunity cost vs using something like Opus to be productive is large.
Gemini 3.5 Flash is the most annoying model I have ever used. It loves to respond in ALL CAPS like “LOOK AT THAT” for no apparent reason. I realize it’s a flash model but I will give it a basic list of tasks and the output will simply vomit “Now I will X” “Now I will Y” “Now I will Z” over and over again filling my screen with garbage. It’s also no smarter than 3.1 Pro and consumes just as many tokens as 3.1 Pro, it’s really pointless without a newer Pro model in place.
I respectfully disagree, at least for most tasks.
3.1 Pro: while it's coding performance is mediocre, a lot of coding work requires minimum actual thinking. I use it often for light refactoring, boilerplate generation, testcase skeleton generation, code review, language questions ("is there a better way to write this code block?"). I don't have a corporation behind me so costs matter. Considering that I'm a Pro subscriber, it's quite cost effective.
3.5 Flash: excellent model for general (re)search. I use it for everyday tasks with Thinking instead of Flash-Lite. It's a much better version of Google Search for general queries like gaming tips, cooking, day-to-day first aid, tax and investment questions, etc.
Google is clearly aiming for cost-benefit here and considering that it gets bundled with YouTube Plus and Google One at $20/month, it's a killer deal IMHO.
PS: I don't work for Google and don't even like Google very much. But this is a good product.
It was a decent deal ~6-8 months ago. I had been using 3.1 pro almost since release, but it really is feeling old. After using other models more in the past two months though...I really can't go back to 3.1 pro, as I just have to explain my reasoning so damn much to get it on the right path, where as opus or fable just "get it" from the context of the project much better.
Sonnet is roughly the same level as 3.1 pro for me.
This would be more useful if you could compare to a local model like Qwen3.6 27B or 35B.
where is gemini-3.5-pro ??
my question exactly - it was announced months ago on IO
yeah, I was surprised when Sundar said they will release it in a month, since training/evals is not so predictable
You're not supposed to ask this question.
It may be a dud. The blog post assures us they are now pretraining Gemini 4.
In june. Year unspecified
have we defined what ai limbo is yet?
rumor has it that pro is not so pro in internal evaluations and that there are internal stakeholders holding it up
I've heard almost nothing about Gemini in my circles, are they still in the race to build AGI?
Of the models I've tried lately, I get more value out of Gemini for personal stuff like researching products or learning how to do things than the others. It seems more factually correct in domains I know about and hallucinates minor stuff less than the others.
I flop between 3.1 and 3.5 when I need it to be flawless. Fast and relatively affordable. Up to 6x cheaper than competitors.
DeepMind is absolutely still in the race to build AGI, but they see AGI as more than just a big LLM.
Google has not had a single SOTA model since about a year now.
You are exaggerating. 3.1 Pro came out in February of this year.
3 replies →
I know that Flash is obviously no Opus or Fable. That being said, when I want a very quick response, my go-to is Gemini Flash. It's just really damn fast and generally tends to be accurate enough.
Yeah, Gemini Flash is my daily go to. It's certainly quite capable - yesterday it talked me though fixing my Ubuntu desktop (rebuilding NVIDIA driver) after an update caused it to boot into low resolution mode.
Why isn't there an official annoucement? And is pro dead haha?
Is this supposed to be a replacement for 2.5 flash?
I don't like how it compared itself to 5.6 Luna instead of Sol, and still losing on some metrics. You will not see me using this over even 5.6 Sol-light
Claude's free usage is so limited that it's useless. Gemini is extremely generous - never hits a limit for me despite being used all day.
A couple of days ago I was comparing Claude vs Gemini asking both what the size of YACC generated parsing tables for ANSI C was. Claude was very gung-ho and rushed off to download a C grammar, install bison, and find out for itself (not really necessary), and then hit daily free usage limit half way through without even getting to an answer (despite this being the only thing I'd asked it that day).
[flagged]