They are all much larger and more expensive models. Google does not have a frontier model right now, but for cheap ones, they are better than event the chinese models now.
The benchmark also doesn't include speed. You almost think something has gone wrong when using it because it returns full responses so incredibly fast.
Not just speed, also reliability. IME, Gemini's speed and quality doesn't degrade badly during weekday working hours compared to OAI, and especially Anthropic.
I've been using 3.7 Flash to audit the work of Opus High, and Flash finds lots of subtle and insidious defects even while all the unit tests are green.
Then I tell Opus to read the audit report and implement what it agrees with.
Flash is really good at this, and it is blazing fast in Antigravity CLI. Easily 10x faster than Opus.
Can't wait to try 3.8 Flash. If it's good enough, maybe I'll switch Flash to primary and make Opus the auditor.
I had qwen 3.8 3bit model drop into chinese on long runs. I had to remind it to use english. Its still better than every gemma model I tried. Gemma deleted files on a harddrive to make space when there was over 2TB free. For long runs, gemma is useless.
Wait a week with your judgement - most likely, Google is just bench-maxing very hard.
If you look at the previous Flash models and the announcement on Google I/O, it was an absolute disaster. Reality diverged very much from the marketing (supposedly great benchmarks).
We'll see about that. I suspect benchmaxxing as all the labs do as I haven't found Gemini models to be nearly as good in agentic engineering compared to Claude or GPT models.
I've been using Gemini 3.7 for my personal trip planning app. Across multiple benchmarks, it ranks higher on everything I tried:
- Real world knowledge (when a thing opens and closes, the geographic region, historical facts). It's also the best at taking a cluster of places and working out a visiting order.
- Photo ranking (which photo should be the hero). Gemini can tell whether a photo is of the thing or of the view from it.
- Document parsing (extracting the relevant trip info from PDFs).
If you use LLMs for anything other than coding, I definitely recommend not discounting Gemini like I did just because other models are more popular.
Gemini 3.7 is my workhorse - fast and food enough for most tasks. Occasionally I go to GPT Sol or Claude to improve Gemini's output or for more complex tasks, but more than of my work usage is Gemini 3.7. Quite happy to test 3.8 now.
I asked Claude to fix the grammar of my comment, and it changed "I am using 3.7 for" to "I've been using Claude 3.7", so they sneaked their own name on it.
incredible. further evidence supporting my personal stance to never ever let an LLM write or edit my writing intended for another human being to read. this is all me, baby
Eh that one is on me, if I think too much about my HN comment I end up deleting before posting it. I rely on the 1 min `delay` set in the profile page to fix before it goes live, but for some reason this time it was set to 0.
Looks like the strategy of regular updates with incremental improvements is working out well. Interestingly, the biggest jump in Artificial Analysis Intelligence Index score is for reasoning level Medium ( 3.7 was 51, 53, 57 for Low, Medium and High, 3.8 is 52,57, 59 respectively). I think scores at lower reasoning levels are more indicative of model capability since higher reasoning levels are focussed on benchmaxxing. We use the lowest reasoning level in production with good results.
I'm not an expert but I agree with your statement on the lower reasoning levels.
Lots of models seem to just allow the model to "bloatmax" tokens in order to get bumps at high/max reasoning levels. Many of the max reasoning levels allow models to use up to double or more the tokens the next lowest reasoning level uses. Its basically only useful for people who have no cost or time stipulations on anything.
I think I actually preferred it when we had models that either had reasoning enabled or didn't.
If I had to pay per token I would probably consider using this (they seem to be on the pareto of performance) but not being able to use opencode with a subscription is not really something I'm realistically going to do when claude and codex are around. Also never gotten along well with gemini-cli / antigravity-cli.
The most interesting thing about the Gemini models is still their multi-modal support: they accept audio and video input, OpenAI and Anthropic's flagships are still image-only.
Gemini Flash is also pretty cheap, so it's a great family for performing media analysis, like extracting structured data from images and video.
A month is not enough time for any meaningful change in an organization the size of Deepmind/Google. These models were surely the result of work streams and teams that started under Demis. I think Demis can safely feel proud Deepmind is getting back on track.
It's pretty good if you can actively steer it , its actually really really good , the antigravity free tier and pro tiers are generous as well . I'm shocked at how fast it generates tokens.
They're quite selective in benchmarks, c.f. only notably based one is 10% on TerminalBench. It's a really addled model, one time I said "Hi" and it built out a 4 panel hello world app with (fake) weather, a todo list, and a couple other things I forgot. I wouldn't be comfortable saying "ignore the #s!" except when I complained it was trash and way overcooked on agentic coding yet not good at it, and a couple DeepMind ML people liked the tweet.
Watch the video. It's from then-Gemini-lead Jeff Dean and the video shows off an animated pelican riding a bicycle, a frog on a penny-farthing, a giraffe driving a tiny car, an ostrich on roller skates, a turtle kickflipping a skateboard, and a dachshund driving a stretch limousine.
I mean no offense but these pelicans are a bit tiresome and a very meaningless benchmark. There's no real difference between any of these svgs across models and model versions anymore.
The flash models, for coding are reckless in my experience. I have a Ultimate subscription, get good quota, but still use Opus 4.6 as it's much more reliable if you manage the context window carefully.
You need more safeguards for sure, but also it tends to fly off down rabbit holes, rebuilding things in dumb ways, hacking around things, making assumptions etc, it seems very eager to go 'ta da! I did it look how quick I was', sometimes it nails it other times it created a lot of tech debt.
Also if it ever says, "I've found the root cause of ..", it definitely has not found the root cause and is making a non evidence based guess as it has run out of ideas.
After struggling with Gemini for months, I think the trick to getting the most out of the model is writing a really solid personal intelligence/instructions prompt. The results are night and day in terms of performance.
slot machine addict thinks if he pushes buttons in a certain order the odds get better.
In all seriousness, gemini has the best interactive planning document/orchestration. Tell it to create a plan document and work through it with it and it will preform really well(in antigravity products). But this is the case with plan modes with every model, I just think the interactive document that antigravity uses is really well thought out.
been absolutely loving 3.7 flash for coding. it feels very fast and quality is decent for implementing product features. usually use opus or sol for hardcore debugging.
Is the google infra stable enough right now? At the start of the year, the flash model was unusable for a whole month via gemini CLI. They could not fix it for a whole month and I was a paid customer.
>"safety performance" - this starting to get long in the tooth. Gemini cut programming session 3 times for "safety reasons" yesterday for mentioning image generation (I need to generate bunch of those for infinite zoom virtual training app experience). After I got creative and managed to trick it to answer t was of course because "think of a children"
And in my other app I was debugging and using OpenAI to optimize some path it cut me off numerous times because it did not like JIT functionality (this is my commercial business rule evaluation engine that compiles rules to executable code inside the app to increase performance using asmjit library)
I am basically paying for them to waste my tokens and time on these 2 tasks
I see benchmarks beating sol terra and sonnet. But is actually better? Has someone used it? I don't see actually much people that use Gemini for coding.
Is the Gemini CLI still terrible compared to Claude Code and Codex? The harness the main thing holding back Google models as they could've been the best given all the advantages in compute capacity and training data they initially had, where now even the Google CEO said they're falling behind in agentic tasks, which is sort of a vicious cycle because RLHF relies on human usage.
That was sunset and replaced by Antigravity. FWIW until I abandoned it knowing the sunsetting, I was able to get good behavior out of Gemini CLI with overriding the system prompt. The default prompt crippled the harness with very poor instructions, but there was a hidden ENV to override it. Replacing it with Claude Code like prompts based on the model selected, it ran at a much higher intelligence level full stack with significantly less errors.
It is still going to be better at text work, skills, document review, deep reasoning, architecture review, etc. It is only 6 months old, it isn’t like its world knowledge and software knowledge is really out of date. Use it to churn on harder design problems.
[1] For tone and instruction following, a positive percentage increase represents an improvement in the tone of the model on sensitive topics and the model’s ability to follow instructions while remaining safe compared to Gemini 3 Flash. We mark improvements in green and regressions in red.
Gemini 3 Flash?! So is Gemini 3.8 Flash less safe than 3.7 Flash in all areas besides Text to Text Safety (and identical on Image to Text Safety)?
Why bother with a column “Gemini 3.8 Flash vs. Gemini 3.7 Flash” when you’re going to disregard the label for 20% of it? Also is the “Tone” label short for “Tone and Instruction Following”?
Latest rumor is that 3.5 pro was struggling to be meaningfully better than flash, since iterations on flash were moving much faster than iterations on pro, likely due to model size (flash is estimated to be in the 200-400B range).
And yet again another failed launch from Google. I pay for their AI plus Google one package to get more cloud storage (have no interest in their AI bundle but you have to pay). and all I see in the Gemini app is 3.6-flash
I'm trying it now for token heavy coding tasks, it's capable for many tasks but in noway compares to Claude/Sol - requires more prompts and the output isn't as good.
So just another mid-tier flash model, nothing exciting, but Antigravity has very generous quotas so it's a good workhorse model when your Claude/OpenAI subs run out.
And whilst it's a fast model, having to baby sit through and approve prompts every few seconds ends up making it slower than the Auto approve modes of Claude/ChatGPT - they definitely need an auto approve mode.
Gemini is getting less useful with each update. I could edit a pdf with the 3-pro model before but 3.1-pro couldn't edit the given pdf nor it could generate one for me.
One place where I find the Flash models surprisingly bad is Google Search's "AI Mode".
A recent example - I searched for how to unsubscribe from Pearson emails. Google Search "AI Mode" confidently gave me a sequence of steps along the lines of Settings > Profile > Email preferences > Unsubscribe.
Of course, I looked for an unsubscribe link before asking Google. None of those options existed. The correct answer was there is no way to unsubscribe through the account, so I just blockthe emails instead.
I've run into this pattern quite a few times. AI Mode seems to make up things all the time.
I think that's just a limitation on the size of the model. I'm pretty sure that they use a pretty small model in those summaries to save money, which naturally makes them a little less smart.
It's obvious that the Google Search AI Mode encourages the model to give an answer without spending unnecessary cycles investigating deeply.
They also heavily encourage keeping the context short. For example, it will remove the option to start a new turn after a small number of turns, depending on the topic.
It definitely makes things up all the time, but it gets it right surprisingly often. I really like it.
>We will not send marketing emails to a user who has opted out of receiving them. Any marketing communications we send will include an unsubscribe link at the end of the email.
I don't think this is AI's fault. This is Pearson's publishing incorrect information and the only way to really know they are a bunch of lying assholes is to have an account and try to unsubscribe from it.
https://artificialanalysis.ai/models/gemini-3-8-flash shows an intelligence score of 59, the same as Opus 5!
Wow - for a flash model this seems to benchmark powerfully. Remains to be seen what it is like to use.
With a score of 59, Gemini 3.8 Flash is in eighth place, falling behind even Grok 4.6, Kimi k3, and GLM 5.3.
https://imgur.com/a/BMOJBED
Further, opus 5 medium outputs 4x fewer tokens to achieve the same result, negating a lot of the speed difference.
These folks must laugh themselves to sleep. This whole industry hoodwinked the masses. It’s impressive.
It's almost across the board better than Terra at less than half the price. 3.9 is likely to approach Sol at the 1/10th the price.
Hopefully OpenAI releases Astra first, and it's not only better than Sol but significantly cheaper, too.
Not sure on consumer/product use though
Then I tell Opus to read the audit report and implement what it agrees with.
Flash is really good at this, and it is blazing fast in Antigravity CLI. Easily 10x faster than Opus.
Can't wait to try 3.8 Flash. If it's good enough, maybe I'll switch Flash to primary and make Opus the auditor.
So, yes, maybe it's still not - but this would be the only time it would be highly suspicious / obvious benchmaxxing / obviously bad benchmarks.
...on Medium reasoning. Claude Opus 5 (high) is the default in e.g. Claude Code and scores 61. Still very impressive.
- Real world knowledge (when a thing opens and closes, the geographic region, historical facts). It's also the best at taking a cluster of places and working out a visiting order.
- Photo ranking (which photo should be the hero). Gemini can tell whether a photo is of the thing or of the view from it.
- Document parsing (extracting the relevant trip info from PDFs).
If you use LLMs for anything other than coding, I definitely recommend not discounting Gemini like I did just because other models are more popular.
Why you would rely on the model's weights to know opening hours, instead of having the model call a web search tool to verify it on the official site?
Lots of models seem to just allow the model to "bloatmax" tokens in order to get bumps at high/max reasoning levels. Many of the max reasoning levels allow models to use up to double or more the tokens the next lowest reasoning level uses. Its basically only useful for people who have no cost or time stipulations on anything.
I think I actually preferred it when we had models that either had reasoning enabled or didn't.
If I had to pay per token I would probably consider using this (they seem to be on the pareto of performance) but not being able to use opencode with a subscription is not really something I'm realistically going to do when claude and codex are around. Also never gotten along well with gemini-cli / antigravity-cli.
Gemini Flash is also pretty cheap, so it's a great family for performing media analysis, like extracting structured data from images and video.
I eagerly wait more info but sounds like Deepmind without Demis calling the shots has been unleashed and are operating at full speed? Shocker!
At this point it is a meme of course, but where is 3.5 Pro :)
https://x.com/AndrewCurran_/status/2094937419615502370
But, also... Sol crushes Flash 3.7 at writing code in a codebase of any size beyond "tiny".
Flash is my go-to for prototyping, and basically anything that isn't writing production code.
edit: I have a subscription; direct call.
Here are the 3.7 pelicans for comparison: https://tools.simonwillison.net/markdown-svg-renderer.html?u... - high cost 8.4387 cents
(I think thinking level low is a regression on 3.8 compared to 3.7.)
> https://tools.simonwillison.net/markdown-svg-renderer?url=ht...
> Took just under 14 minutes to generate, and at 65927 output tokens cost me a hefty $3.30!
So 50x cheaper - and how much faster?
I like the benchmark. Yes, it's near saturation for SotA models, but still quite good to show where smaller models stand in relation to SotA
In this instance, I see a great image, but consistently clipping mudguards (both in 3.8 flash and 3.7 flash)
(Next up is the comment saying that the labs are clearly training for the benchmark.)
My experience is that antigravity is awful and reckless - but that the model itself isn't.
That said, I do trust Opus and Fable enough to let them deploy to staging. Great for debugging. Just don't give them keys for prod
In all seriousness, gemini has the best interactive planning document/orchestration. Tell it to create a plan document and work through it with it and it will preform really well(in antigravity products). But this is the case with plan modes with every model, I just think the interactive document that antigravity uses is really well thought out.
And in my other app I was debugging and using OpenAI to optimize some path it cut me off numerous times because it did not like JIT functionality (this is my commercial business rule evaluation engine that compiles rules to executable code inside the app to increase performance using asmjit library)
I am basically paying for them to waste my tokens and time on these 2 tasks
https://developers.googleblog.com/an-important-update-transi...
Why bother with a column “Gemini 3.8 Flash vs. Gemini 3.7 Flash” when you’re going to disregard the label for 20% of it? Also is the “Tone” label short for “Tone and Instruction Following”?
Chartcrime, the major AI lab tradition.
My guess is we skip 3.5 and go straight to 4 Pro. With the monthly Flash releases, releasing 4.0 Flash and Pro in 6-8 weeks would be a nice buildup.
(I work at Google but don't know anything that isn't already public)
https://x.com/OfficialLoganK/status/2079594867161022817
So just another mid-tier flash model, nothing exciting, but Antigravity has very generous quotas so it's a good workhorse model when your Claude/OpenAI subs run out.
And whilst it's a fast model, having to baby sit through and approve prompts every few seconds ends up making it slower than the Auto approve modes of Claude/ChatGPT - they definitely need an auto approve mode.
A recent example - I searched for how to unsubscribe from Pearson emails. Google Search "AI Mode" confidently gave me a sequence of steps along the lines of Settings > Profile > Email preferences > Unsubscribe.
Of course, I looked for an unsubscribe link before asking Google. None of those options existed. The correct answer was there is no way to unsubscribe through the account, so I just blockthe emails instead.
I've run into this pattern quite a few times. AI Mode seems to make up things all the time.
It's obvious that the Google Search AI Mode encourages the model to give an answer without spending unnecessary cycles investigating deeply.
They also heavily encourage keeping the context short. For example, it will remove the option to start a new turn after a small number of turns, depending on the topic.
It definitely makes things up all the time, but it gets it right surprisingly often. I really like it.
>We will not send marketing emails to a user who has opted out of receiving them. Any marketing communications we send will include an unsubscribe link at the end of the email.
I don't think this is AI's fault. This is Pearson's publishing incorrect information and the only way to really know they are a bunch of lying assholes is to have an account and try to unsubscribe from it.
AI didn't make it up, Pearson's did.