Gemini 4 Neon is the first real Opus 5.5 competitor
A long-time Gemini critic spends two days pushing Google's new medium-sized model and comes back convinced coherency, not raw capability, is the new frontier.
Once frontier models are all smart enough, the one that wins real coding work is the one that stays coherent over long tasks, and Gemini 4 Neon is the first medium model to prove it.
Who This Is For
Read if. Skip if.
READ IF YOU ARE…
A developer running agentic coding tools daily who is deciding whether to add a Gemini model to the rotation next to Opus, Fable or GPT.
Someone who wrote off Gemini after Argon and Carbon and wants a concrete, example-heavy reason to try again.
A team lead picking a default model for long-running agent tasks where staying on track matters more than benchmark scores.
Anyone who wants a usable mental model for why some models feel better to work with even when their benchmark numbers look similar.
SKIP IF…
You need a neutral benchmark report. This is an enthusiastic first-week review with the reviewer's own tooling, partly sponsored.
You mostly generate 3D scenes or small games. The model is openly weak there.
You want setup instructions for the Gemini app or CLI. The reviewer uses it through his own client and calls the official app rough.
TL;DR
The full version, fast.
Gemini 4 Neon is a medium-sized Google model that, in this reviewer's testing, rivals Opus 5.5 for code at $3 in and $15 out per million tokens. The argument rests on a capability versus coherency chart: most models are now capable enough, so the differentiator is whether a model stays on task, reads clearly and finishes long work. Neon cut a Rust type checker's binary 5.95x and made warm runs up to 20x faster in 10 hours, after other models burned 10+ days. It still hallucinates on repo questions and is weak at 3D, but it verifies when it implements. Try it on long-horizon agent work before trusting the benchmarks.
Free for members
Chat with this breakdown — free.
Sign in and you get 23 free chat messages on us — ask for the hook, quote a framework, find the exact transcript moment, generate a markdown action plan. Bring your own key when you want unlimited.
Sarcastic setup about Gemini, then benchmarks above Opus 5.5 and a claim that this is the first real Opus competitor.
02:11 – 04:12
02 · Sponsor and who pays him
WorkOS read (Passport, Airlock, Enterprise SSO), then a disclosure that only Google and SpaceX AI pay him, via YouTube and X.
04:12 – 07:47
03 · What Google announced
Index rankings, $3/$15 pricing, cache discount, computer use, Gray Swan safety, and a 50% cut in time to first token.
07:47 – 10:43
04 · The coherency test
A disk-speed thread other models botched: Neon finds the misconfigured disk, and its writing is simply readable.
10:43 – 12:57
05 · Trusting it with real work
Other models defer to Neon's Lakebed design, and Neon sets up a second machine end to end and messages another thread.
12:57 – 15:19
06 · A failure it diagnosed itself
Missing visuals turn out to be a client rendering bug that Neon identifies, works around and files a PR for. Its one real miss: not triggering a skill.
15:19 – 17:26
07 · Capability vs coherency
Why he says it should have been called Gemini 5, and the two-axis chart for judging models.
17:26 – 20:44
08 · TS Rust: 10 hours vs 10 days
Same goal, same fork point: Neon makes the binary 5.95x smaller and warm runs up to 20x faster in under 10 hours.
20:44 – 23:17
09 · The press-release prompt
A plan written as a launch blog post that ignores the skill's style guide, which most models fail.
23:17 – 25:18
10 · Benchmarks and hallucinations
10 of 72 architecture claims wrong, a clean code review score, and a jump from 12th to 6th on his aggregate board.
25:18 – 26:40
11 · Weak spot: 3D
The fish 3D test is still poor, the expected cost of a smaller model.
26:40 – 30:19
12 · Where did this model come from
Speculation that research tested engineering's feedback on a medium model, plus the confirmed detail that engineers begged to ship it early.
30:19 – 32:29
13 · Verdict: coherency is the new frontier
Try it, use the nightly client, and watch for the big version.
Atomic Insights
Lines worth screenshotting.
Benchmark scores stopped separating frontier models; coherency over long tasks is now the gap that decides which one you keep using.
Gemini 4 Neon is priced at $3 per million input tokens and $15 per million output, meaningfully cheaper than Opus 5.5.
Cached input on Gemini 4 Neon costs $0.15 per million tokens, a 95% discount on fresh input.
Google cut Gemini's time to first token by over 50% after public complaints that the models felt slow.
Tokens per second is meaningless if a model waits 30 seconds before it starts streaming.
Given a three-week-old debugging thread, Neon found that a VM disk was formatted for a spinning hard drive, which explained 3.8 of a 6.3 second slowdown.
A medium model trimmed a Rust TypeScript checker from 47.44 MB to 7.97 MB and made warm starts up to 20x faster in under 10 hours.
On the same goal, a competing model ran 7+ days, made the binary 46% smaller and 56% slower, then spent 4 more days landing nothing.
A model can rank 13th at describing a codebase and still be the best at changing it, because it verifies claims when it implements.
Neon got 10 of 72 architecture claims wrong, roughly half the hallucination rate of Gemini 4 Carbon.
When its HTML output failed to display, Neon diagnosed the client bug, showed screenshots instead, and later filed a PR to fix it.
Asking a model to write a plan as a launch blog post, Amazon's work-backwards press release trick, is a sharp test of instruction following.
Most models ignore an instruction to skip a skill's style guide; following that one clause is a real coherency signal.
Smaller models still pay a visible tax on 3D generation even when their coding and reasoning hold up.
Google's own engineers pushed to ship Neon early because they wanted to know if the model was really as good as it felt internally.
Takeaway
Judge models by coherency, not benchmark scores
WHAT TO LEARN
Most frontier models are now smart enough, so the one worth using is the one that stays on task, explains itself clearly and finishes long work.
03What Google announced
Neon matches or beats Opus 5.5 on several published indexes while costing $3 in and $15 out per million tokens.
Cached input at $0.15 per million makes long agent sessions with repeated context far cheaper than fresh input.
Raw speed numbers mislead when time to first token is slow; Google cut that delay by over half for this release.
04The coherency test
Re-run an old, hard debugging thread on a new model; Neon found a misformatted disk that explained most of a 6.3 second gap.
Readable output is a real productivity gain, because you spend less effort parsing what the agent did and what needs your attention.
05Trusting it with real work
Hand a model a loosely bounded task end to end, like configuring a new machine, and check whether it finishes without drifting.
When other models defer to one model's design in review, treat it as a signal to give that model the whole thread.
06A failure it diagnosed itself
Watch how a model handles its own failure: the best response diagnoses the cause, delivers a workaround, and reports clearly.
Expect small instruction misses even from strong models, such as not triggering a named skill from a casual reference.
07Capability vs coherency
Place models on two axes, capability and coherency, to explain why a high-scoring model can still feel worse to use.
Coherency means staying on task over long horizons and collaborating cleanly with sub-agents, not just being pleasant to chat with.
08TS Rust: 10 hours vs 10 days
On an identical goal and starting point, Neon cut a binary 5.95x and sped warm runs up to 20x in 10 hours.
Competing models spent 10+ days on the same goal and either regressed speed or landed no working changes.
Running the same task from the same fork across models is the fairest way to compare them on real work.
09The press-release prompt
Asking for a plan written as a launch blog post forces clearer thinking about what the feature means to users.
Including an instruction to ignore a skill's style guide is a cheap test; most models drift back to the default style.
10Benchmarks and hallucinations
A model can score poorly at describing a codebase and still change it best, because it verifies claims during implementation.
Neon got 10 of 72 architecture claims wrong, so double-check its explanations even when you trust its code.
Benchmarks you built yourself can drift from your real experience, so weigh shipped results above leaderboard rank.
11Weak spot: 3D
Smaller models still trade away 3D and visual generation quality, so do not judge them only on those demos.
3D demos are a poor proxy for daily coding work unless that is the work you actually do.
12Where did this model come from
Internal engineers pushing to ship a model early is a stronger quality signal than a carefully worded launch post.
Labs may test feedback-driven training changes on a medium model before committing to an expensive large training run.
13Verdict: coherency is the new frontier
The next big competitive gap between labs is coherency over long tasks, not peak capability.
The tool you run a model through matters; a rough official client can hide how good the model underneath is.
Glossary
Terms worth knowing.
Coherency
How well a model stays on task, communicates clearly and keeps track of what it is doing across long multi-step work. Separate from raw capability, which is how hard a problem it can solve.
Long-horizon task
Agent work that runs for many steps or hours, such as a multi-day refactor or performance push, where small drift in focus compounds into failure.
Time to first token
The delay between sending a request and the model starting to stream its answer. High values make even fast models feel slow.
Prompt caching
Reusing already-processed input from earlier requests so repeated context is billed at a steep discount instead of full input price.
Pareto frontier
The set of models where no other model is both better and cheaper. Landing on it means you cannot get more quality for the price elsewhere.
Work backwards / PR-FAQ
An Amazon planning habit where a team writes the launch press release first, then designs the product to make that announcement true.
Steer
A follow-up message sent to an agent while it is already working, used to redirect or add instructions mid-task.
Warm start
Running a tool when caches and previous state are already in place, as opposed to a cold first run.
“The reason Neon is good is because Carbon was garbage.”
contrarian twist on the obvious take→ TikTok hook↗ Tweet quote
14:02
“It did all of the right things and none of the wrong ones.”
short, quotable praise with a story behind it→ newsletter pull-quote↗ Tweet quote
24:11
“Whether or not it is hallucinating about the repo, it is delivering.”
reframes how to read benchmarks→ IG reel cold open↗ Tweet quote
28:19
“The engineers at Google and DeepMind loved using this model so much that they begged and pleaded to rush it out.”
insider detail, story hook→ TikTok hook↗ Tweet quote
The Script
Word for word.
Read-along
Don't just watch it. Burn it in.
See every word as it's spoken — crank it to 2× and still catch all of it. The same dual-channel trick behind Amazon's Kindle + Audible.
17px
metaphorstory
Oh boy, another Gemini 4 model. I'm sure considering how much I loved Argon in Carbon, I'm going to absolutely love this one, right? Yeah, it's a small model too, so there's no way I'm ever going to actually pick this, right?
I'm sure the benchmarks are going to be terrible. Oh, Gemini 4 Neon is scoring fully above Opus 5 .5 and cheaper and neck and neck with Fable 5 too. And when you switch to code mode, they get even closer.
Huh? When I woke up yesterday and saw this, I was admittedly kind of shocked. I even tweeted as much, saying, wait, WTF?
Where did this come from? This looks significantly more compelling than carbon. Haven't tested it yet, but I will give it a spin now.
This video is a bit late, considering that I almost immediately tried it. The reason is because I just could not believe what I was experiencing. I've gotten in trouble for this before, and I took extra time to make sure I do feel this way.
This might be my favorite model right now. Yeah, it hurts me too. I don't know where this came from.
Google just out of nowhere cooked something that is a generational leap that is also smaller and cheaper and generously subsidized in their Gemini subscription plans. This is the first real Opus 5 .5 competitor. And it's a very, very real competitor.
I have pushed this model to its absolute limits. And honestly, it has been hard to find them. There are some rough edges here and there, but there are way fewer than I ever would have predicted.
This is a phenomenal model, and I need to take some time to explain why this model is so good, what it is that I like about it so much, what makes such a huge gap, not just from the other Gemini models, but other recently released models, and why I would never, ever pick Astra over Gemini 4 Neon, the medium -sized model from Google and DeepMind.
I swear to y 'all, I am not trolling here. This is a really good release and has fundamentally changed my tone around DeepMind and Google. The paid shell allegations are already coming in strong, and I need y 'all to understand Google only pays me for a little bit of money on AdSense that is not enough to even cover my editor's costs, which is why we need to take sponsors like this one.
I'll cut to the chase. You guys already know WorkOS, they're today's sponsor, but you might not know about all the other cool things they've been cooking recently. Tools like Passport are mind -blowingly cool, having your whole machine set up for just you so you don't have to manage all the auth for all the different apps you use, and more importantly, the apps that your agents are going to use on your behalf.
Or Airlock, their new access control system that thinks more about what you're actually asking for and not just what data is being touched, making it way more reliable while staying secure. And don't forget the thing everyone signs up for, Enterprise SSO. making it so your product can be registered for by all sorts of different enterprises and all the weird things those companies expect.
As good as agents are adding a Google sign -in button, they're probably not going to get all the details for PingFederate or Okta or Auth0, OpenID, OneLogin, ADFS, and more right first try. And when you have enterprise customers coming in, it's important to be ready for them. Make sure you're ready today at soydiv .link slash workOS.
Sorry about the break. It's funny that Google technically is one of the two labs that does pay me because Google pays me with my YouTube channel. Hi, nerds.
And SpaceX AI pays me because I get paid to tweet on X. yeah those are the only two labs that pay me the rest have never exchanged a cent i pay them all a lot though it looks like i'm gonna be paying google a hell of a lot more too by the way google if there's any way i can convert my youtube ad rev into more subsidized usage in gemini we we should chat i would use that i would use that a lot because this model is just blowing me away let's start with what google said in their announcement so that i can couch accordingly but then we have a lot of other fun layers to go through here from the incredible things i'm getting out of it to some very novel things that i've only seen this model do and it's doing better than anything else to the reasons why i think this model just randomly came out the way it did and also some of the conversations i've had with my friends over at deep mind that are much happier to talk to me now that i'm being way more positive because they finally shipped a good model
I cannot believe that Google has a model that I would currently prefer to everything OpenAI has out, but they do, and I now need to explain it. Today we're introducing Gemini 4 Neon, bringing frontier -level intelligence with fast end -to -end task completion across complex software engineering, knowledge work, and science workflows.
Neon ranks ahead of comparable models in number three overall with a score of 61 on the Artificial Analysis Intelligence Index at its ex -high thinking levels, landing just behind larger, slower models. On the VALS Index, it ranks number two and it lands on the Pareto Frontier. Beyond just pushing on raw model capabilities, they're continuing to make the model better to work with.
Neon ranks number two on steerability and on the new agent arena index. I think these ranks are low and I really want to emphasize the thing that they said here. We are continuing to make our models better to work with.
This might be the greatest understatement I've ever seen a lab put in a post that was a good thing. This is the most pleasant to work with model I've ever used. The price is pretty good too.
$3 per million in and $15 per mil out. That puts it at meaningfully cheaper than Opus. The cash discount is also doubled, similar to how it was for Opus and the recent OpenAI releases where it's 15 cents per mil cash instead of the 30 cents that it would have been in the past.
Google's cash hit rates are still not as good as the other labs, but they are actively working on this. If I've learned anything from this release, it is that all the things I've been flaming Google about for the past few months resonated and landed in this model. seems to be the result of feedback from not just me obviously but from all the others saying a lot of these same things so uh if you're in my comment section right now saying haha told you so carbon was great no the reason neon is good is because carbon was fucking garbage the fact that that model was so atrociously bad is what enabled neon to rise from the ashes and i will justify all of that in a bit it's great enterprise analytical and scientific stuff cool awesome it slaughters benchmarks Google's been able to slaughter benchmarks for a while.
They're really good at benchmarking. I don't care. My own benchmarks don't even mean anything to me anymore because every model is smart enough for the most part.
It's just coherency is a problem. And that gap has been closed. The next section is built for production -ready code.
I was skeptical of this too. But instead of explaining why I'm no longer skeptical, I'm going to skip this section, tell you, trust me, it is. It's absolutely good enough for this.
And I will show you some mind -blowing examples in a bit. It brings efficiency and control to production -grade workflows. This is just Intelligent Index, stuff like that.
It is apparently also really good at computer use. I've only done very minimal computer use stuff with it so far, but it's absolutely been great at that. I don't know if it's better than OpenAI computer use.
Some of my friends doing browser use and CUA work are saying it's neck and neck, if not better. And it is cheaper if you're using API prices, not subsidized. So in almost all the things that I look to OpenAI models for, this one is very good.
also quite safe in gray swan it performed quite well getting a perfect zero across the board whereas luna got a 28 .7 attack success rate which is very scary it's good at deep analysis which i was skeptical of because it's a medium -sized model and these normally aren't that good at this type of thing again i have examples i've have so many examples i've been using the out of this model for the past two days they show here that it sits in the middle where it is somehow scoring higher than carbon because it's newer and trained better it is much faster it's not as fast as hydrogen but they also and i crashed out about this in a previous video if you saw the video where i explain all the things google's doing wrong with the speed i'm sorry i had to do that and it worked they have since cut the time to first token by over 50 because of that crash out that one youtube video is the best thing that ever happened to gemini users who care about like speed because everything before that just was not being fixed it doesn't matter how many tokens per second you can do if it takes 30 seconds before they start anyways time for the glaze fest to start
I started my recent favorite coherency test where I asked the model to take a look at this thread that Astra just destroyed trying to fix the disk speed problems I was having for work tree creation on one of my VPSs. And remember, the problem here was disk speed. Astra will slop for days things that have nothing to do with disk speed.
I asked this model, feeling going insane, can you just tell me why it's slower? Exact same prompt I sent to all the other models. Three concrete things stack up to create the 6 .3 second gap.
First, slash work on this VM was formatted for a spinning hard drive, 3 .8 seconds of the 6 .3 second gap. This is the single biggest surprise and it's a real misconfiguration. I have been using this thread as a test for three weeks and a medium -sized model taught me something about my setup that is the majority of the gap.
I didn't even know this. Instantaneously huge, but I want you to look at something else and how it responded here. Just...
Look again at those first few sentences. Notice how nothing stood out. Notice how they just are readable.
Notice how I didn't stumble on my words at all as I read it. It has a natural pacing, a cadence, clarity, focus. It's just readable.
You could literally watch my descent into madness as I did each and every one of these tests on my Twitter account. I will not suggest you follow me on Twitter. It's a...
shit show i'm a bit of a shit poster i have fun there though i would recommend that you subscribe on youtube though if you want to watch these crazy journeys as models finally get good enough that i can genuinely recommend them i will always be as honest as i can be about these things as a person who prefers opening eyes a company to all the other businesses i think astro's garbage model and i will defend that forever as somebody who has been very unhappy with gemini models for a long time the whiplash i experienced using this was nuts initially i was like i don't hate this model Gonna keep pushing it.
I then swapped to it with some work I was doing decommissioning a machine that I was giving to a friend. And I said I nuked this machine before decommissioning it. Don't worry, because it was fabled before.
What's the Muse readme thing that you're talking about? Because it mentioned something about a Muse readme. There are just three old status notes in the readme that happen to be written in present tense.
All three are under the dated October headings, and it then gives the exact ones. Nothing functional. All three are dated historical logs, so they're harmless to leave in as is.
It's so clear. It's so easy to understand what it did, what it wants, what it thinks, and what I need to care about. My literal only complaint I could make about any of the copy is that it still em dashes a little bit.
I haven't done anything to the Clotter Agents MD for anti -gravity at all, and I'm doing this all with anti -gravity in T3 code, of course. And I'm sure if I just go put something in there that says, no em dashes, please, it will just do it. But I don't even care that much.
I've never cared so little to tune a model through the global files. It's just pleasant. I almost feel like it would be rooted inconsiderate for me to put something like that in because it's so nice to talk to.
I don't want to correct it at all. I had some pretty complex work going on for one of my side projects. You've probably heard me mention it, Lakebed.
It's my cloud for agents to build really simple shared apps for small numbers of users. I wanted to add source control to it so I don't have to remember which computer the project's on in order to edit it. And I had a plan I was working on with Fable and Astra for it.
And I decided to pull a neon to just take a look and give some feedback. Astra... immediately was like, yep, that one, that guy knows what he's talking about.
We're going to copy everything Neon just said. I'd take Neon's single request design and bounded history. It just immediately was like, yep, what Neon said is good here.
Let's do it. So I then made a bold decision to swap the thread over to Neon. And that's when I started to fall in love.
God damn it, Google, how did you do this to me? I'm going through all the threads that I pinned early yesterday when I was preparing to film this video and realizing that none of them are bold enough anymore. Because at the time I was so impressed that I gave Gemini a like, not particularly well -bounded task like here one of my pcs is a windows pc wood rig that i use for gaming dev stuff i only had quad and codex on there because why would i need gemini for my side projects i just use it as a jester normally so i needed to set it up on that machine so i could use it for my game dev work so i just made a thread with four neon and said get gemini set up on wood rig so i can use neon there it worked for eight minutes and while it was working i sent a steer to notify a specific thread when it was available, because I had work going, and I wanted Neon to come in and help.
I'll explain what I specifically wanted to help with. But I noticed in that thread, not because Fable or Aster told me, but because I noticed that no Neon was being spun up, I noticed that the computer didn't even have Gemini and anti -gravity available. So I made a new thread and told it to go get it set up.
I then told it what thread ID needed to know, and then stopped caring. Because it successfully configured that machine with anti -gravity, with my auth, added it to T3 code. verified it was in t3 code found the thread that i was talking about and then sent it a message on my behalf saying hey anti -gravity is ready you can now use this for whatever thea wanted and it just worked and the work has been going for a bit now and it is going much better since i added neon to it but this is the exact type of task i never would have trusted a gemini model with before because it would have gotten lost in the details it would have made a dumb mistake it would have broken something in the process and it wouldn't have been clear what it was even trying to do Now I'm going to show you a slightly different thread here.
Here's one where it failed, at least I thought initially. I asked it to make me a cool visualization of the PR and issue throughput on the T3 code repo over the last few months. It said it was going to fetch these things, and once it finished, it would build and render the interactive throughput.
I asked why it stopped. Sorry about the hiccup. My data fetching background did well pulling all weekly data buckets and your message arrived right as it woke back up and started verifying the totals.
I said, okay, can you make the visualizations now? It then worked for a whole bunch. I gave it a little more hints on how to do it because I didn't have the confidence in the model yet.
It then stopped randomly again here. I said, okay, can you keep going until you make the visual? You keep stopping.
And here's where my jaw dropped and I started texting Ben like, what the fuck happened? Aha, I see what happened. I actually published two interactive HTML dashboards via HTML render in the last two turns, but on the anti -gravity adapter in T3 code, they aren't showing up in the chat.
The recorded tool item came through as run mcptool yada yada yada without the HTML render output attached, so the web UI hid the frame and only showed my short follow -up note. So it identified the problem here. There's a few things it could do.
It could find some workaround. It could propose a fix. It could stop.
It did all of the right things and none of the wrong ones. First, it immediately did what I initially asked, which was show me the visuals. It rendered the HTML itself, took screenshots of it, and then rendered the screenshots because it believed it would have more luck with that than with this broken thing in the T3 code HTML renderer for anti -gravity.
So it gave me that, gave me some really cool visuals, and also, specifically at the top there, told me what went wrong and where the problem was. So I then asked, can you file a PR to fix the bugs you reported here with the MCP tools? And it did.
And here we have the single issue I have had with this model to date. I am not bluffing. This is the only actual problem I have had that was not the fault of T3 code or something weird in my setup.
I did the Theo Garbled speak with my skills of full stop council nine, stopped no background tasks, sub -agency, yada, yada, yada. Let me know if you want anything else done. And then I said, no, like use my full stop and counsel skills.
It didn't trigger the skill from that. I don't know if that is an anti -gravity thing, if that is a model thing, or if that's a behavior they need to change in training or whatever. This is a very small issue compared to what I normally experience with anti -gravity and even with like open AI models recently.
But it is notable. This is just wanted to call it out because it's the literal only issue I've had. Back to glazing.
At this point, I was blown away and I was starting to land real changes in T3 code with the model. So I tweeted, why the fuck is nobody talking about Neon? This model is so good, WTF.
and instantly added it to my council skill and specifically removed all references to the jester because gemini is no longer a joke to me this is a very legit contender for the best model to use for dev work right now it is neck and neck with opus and like fable is still fable it's special it's useful for certain things this model is just good i then followed up and said they should have named this model gemini 5 and nobody would have blinked i stand behind this They should have named this model Gemini 5.
This is a generational gap, in particular on the coherence side. I'm going to do a more detailed video on this concept soon because I've realized it's like one of the bigger gaps that make people not understand when I like certain models and don't like others. I like when models push far out in one of these directions without compromising too much on the other.
Muse Wonderful may not have been the most capable model. It was impressively capable, not like anywhere near as capable as the best OpenAI or Anthropic stuff, but I found it more coherent. By coherent, I mean a few things.
I mean, I find it easier to read and understand. I find it stays on task better. It understands me better.
It completes work better. It feels more like a collaborative partner. And when it works with itself, it's also more collaborative and thoughtful as well.
This isn't just a it's nice to talk to thing. This is a when it works on things for really long amounts of time, it doesn't lose track of what it's doing thing. Now let me justify that one quick, because those were some big statements, and I also kind of regret this chart a bit.
Now that I've used this model even more than I had then, and because of the thing I'm about to show you, this is more so how I feel. Specifically, I find Gemini 4 Neon to be more coherent than any Anthropic model, comically more coherent than any OpenAI model, but more importantly now that I've really pushed it, for code work in particular, I find that Gemini 4 Neon is also more capable.
It can solve harder problems with more success, and it writes code that I'm more likely to merge. Let me justify that. Y 'all know TS Rust by now.
I rewrote TypeScript in Rust for fun. The project got really far. We were using it heavily in T3 code.
TSRS got faster. Buncheck got faster. And I wanted TS Rust to catch up.
I had a few specific fronts I wanted to catch up on. First, I wanted it to be smaller because the binary was a little bigger than I wanted. And second, I wanted it to be faster, to be comparably fast, if not faster, than bun check and TSRS, the other two options for type checking in Rust, to be clear.
Type checking, TypeScript, rewritten in Rust, to be clear. Chaos, okay? I know.
I sent Astra down this path for seven plus days with a very simple slash goal. I need you to take this project and make the package half the size it is in the official build right now. and also make it faster than these two other projects and astra did a great job making it 46 smaller and 56 slower yes slower not faster it slowed it down by over half so i stopped it i was like what the hell are you doing can you make it not slower can you speed it up like we had originally planned stop making a small i don't even care anymore just make it faster it then spent four more days not making it faster did not land a single change because none of the things it tried could even type check apps properly so i gave up stopped it switched over to opus and in a few days i was able to make it slightly faster still not even as fast as the original was but better many days later opus and fable were able to get it like 20 faster than before i had neon start from scratch again not like from scratch entirely but taking the original public version that i had astra fork from and i had neon fork from that same point with the exact same structure same goal i said make this smaller and faster
I also told that it can use the built -in anti -gravity swarm feature. It's like their workflow thing. 10 hours later, not 10 days like it was with Astra, 10 hours later, I now have a version of TS Rust that is 5 .95x smaller than the previous version was.
47 .44 megs down to 7 .97 megs. And it also made it up to 20 times faster in warm start environments. Not...
50 % faster, 20 times faster. Across all of the repos I used to bench TS Rust, overall 1 .5x speed up. The fastest changes were to the Sentry repo, which is 2 .3x faster.
And total time to run the whole suite is down 48 .6%. And again, to be 100 % clear, zero of the other models that I have tested were able to make this type of speed improvement over a 10 plus day window burning hundreds of thousands of dollars in tokens. This model did it in under 10 hours.
We're talking about a different beast here. I got two more things I want to show you guys. Then I want to do some conspiracy theorizing from all of the info I've managed to squeeze out of my friends at DeepMind as to where this model came from and why it is so much better than everything else.
But I have to talk about the copy one last time, though, because I have used a prompt like this for a while. One of the cool things I learned when I was at Amazon is this idea of trying to build your product feature or whatever you're working on around how it'll be received publicly rather than just like what you think should be implemented.
One of the challenges we'd often be given at Amazon was to start by writing a fake press release, like the thing that a blog would announce or share publicly when your product comes out. Like, how would you want the public to see what you've built? And then work backwards from there.
i have tried to get models to write plans this way for a while and none have come even close it has actually been quite frustrating fable can make the page look really nice but it cannot write in a way that feels like an actual blog discussing a new feature this was from the thread i talked about earlier i was trying to add source control to like bed so that i could manage the code inside of the deployments so it was easier to make changes so i switched to neon and i asked do you have a good feel for the path here Write up an HTML page describing the plan.
Write it like a blog post explaining the feature. Post it on post plan, but ignore the style guides from the skill. There's a lot of edges here that many, many, many models get caught up on.
There's only one question here, and LLMs love answering questions. A lot of them, cough, cough, open AI, cough, cough, would get really stuck on this question and focus on answering that rather than doing what I asked. I said it should write up an HTML page describing the plan.
write it like a blog post explaining the feature then this is where things get real tricky post on post plan which is my service for these files as well as in html skill that i have that describes how to do this and also has some formatting guidelines in it but ignore the style guidelines from the skill this is a piece of this prompt that is ignored by basically every single model i have ever given it to for reference this is the formatting that i expect when i use this skill by default and most models will steer towards this black and white minimal style because that's what the skill mentions in it even if i try to steer away from it and here's what we got from gemini yeah i sent this screenshot to ben and he immediately was confused because not only does it look good it also reads too well to feel like an llm Never lose a capsule again.
Built -in source recovery for Lakebed. You deployed a Lakebed app last week. Today you want to tweak it from a different laptop and you have no idea which machine, folder, or ephemeral agent work tree still holds the code.
Here's our tight implementation -ready plan to make every owned Lakebed deployment clonable by default with durable staging, strict access boundaries, and zero git ceremony. Bravo. Bravo.
I am mad that no other model has written this well or handled this prompt this well. It isn't that hard. If Google could figure it out, the rest of y 'all can.
But like, holy shit, Google figured it out. One quick moment of transparency, though. This does not fix all the Gemini -isms.
I simply don't care about the Gemini -isms anymore, though, because it's so pleasant to work with. So does it still hallucinate? Kinda.
Still like bad enough. Like of its 72 claims in the TS Rust architecture analysis, 10 of them were wrong. That does put it slightly behind Fable 5 .2, slightly ahead of Fable 5 .1.
But most importantly... 2x fewer hallucinations than Carbon, which already had overall 2x fewer than Argon. But I've also, as I mentioned before, lost a lot of faith in this bench because Astra did very well, even though the things it reports are garbage.
And Fable 5 .2 did mediocre, even though the things it reports matter. The other crazy thing to note here is that despite coming in at 13th place for its understanding of the TS Rust architecture, according to my existing judging system, it was able to improve performance in that repo the most and the fastest. So whether or not it is hallucinating about the repo, it is delivering.
It is verifying its claims when it needs to actually implement. If you ask it about the repo, it will not do a very good job making sure everything it says is true. But when it comes to actually implementing it, it will get it right.
Way more so than I ever would have guessed. In my code review benches, it actually performed quite well, being admittedly slightly slow. But that's because I was running it on X high in a kind of butchered setup.
It should be much faster now. But it also got 21 valid claims and five plausible that just weren't that important. Not a single wrong one.
Very, very good signal to noise here compared to previous Gemini models. So yeah, while I don't necessarily think this is the best bench in the world, when you aggregate all of the different things that I have been measuring for a bit now, you'll see here a Gemini 4 Neon has forced its way up from 12th place where Carbon was up to 6th place.
It has massively increased its position in terms of the accuracy measurements that I do here. But I don't care. I waited till the end here for this because I don't even think it matters because this benchmark does not reflect my real world experience.
And my real world experience with this model is fucking awesome. But now we need to go to fish slop because it is still not great at 3D. This is one of the hits you tend to take on smaller models.
And this seems to be like the hit it took the most aggressively. It's fucking up like the fonts and things here. It's not a good 3D model.
like this works. The movement feels fine. There's parts that are fine.
Still not great. But I actually did two builds because I accidentally ran it on high initially, so I re -ran it on X high. And the X high one's better.
Still not great though. This feels like a 5 -6 soul to 6 Astra era. Not like Opus 5 -5 onward at all.
It's definitely still... behind the times compared to what we've been seeing with recent models but i don't care 3d is a silly thing i almost feel bad for how much i have put this in my stuff in the past because it really does not reflect how the model behaves in day -to -day work at all unless you happen to be making like small 3d games for twitter as a job which sadly a lot of the influencers talking about models that is their job i write a lot of code like way more than most of the influencers do way more than any influencer should way more than most engines should and this has been my experience it is incredibly capable it is incredibly coherent it takes all the things i loved about muse wonderful turns them up a tiny bit and then takes the capabilities of 6 -2 astra and other models minus like the 3d and and makes them more coherent over long horizons and somehow delivers solutions faster it's just really good so now we are left with one important question where the did this model come from as i've mentioned in almost all of my previous videos on gemini models
The DeepMind team listens a lot better than almost any other team I have complained about in my life. The engineers in particular seem to have been quite frustrated with the state of things for a while now, and they've been trying their hardest to fix it. It seems like research started listening to them.
From here, I am speculating, but there are parts that I know are true, and I will be clear about those. My speculation is that the model training side of DeepMind and the research side of DeepMind was skeptical of how well engineering could... give feedback on how to make the models better beyond like hey this is bad fix it so research decided to try the things engineers wanted on a smaller model not a super small one because that's not going to be good for engineering but not a huge one because those training runs are super expensive and painful and they adjust on argon and carbon pretty close to each other so another big model right after wouldn't make a lot of sense even a medium right after would be a bit much because they just did hydrogen a few days ago and they just had carbon a week ago so my guess is that this model was meant to be a medium -sized attempt to integrate feedback from the Eng team.
I do know for a fact it was not supposed to come out so fast. So why did this model get released so quickly? Again, this is a section I know is 100 % factual.
Thank you to my friends at DeepMind for sharing this. The engineers at Google and DeepMind loved using this model so much that they begged and pleaded to rush it out because they wanted people to stop hating Gemini models so much. They wanted to know whether this was them going insane, thinking this model was great and the world wouldn't understand it, or if they actually had cracked something special here.
Well, I'll tell you with my full heart behind it, they cracked something special here. This is the first model to make me feel a thing like this since my first time with Fable. Opus was like a ramp up, like I got more and more impressed over time.
This model was instantly like slap in the face. It's different. And then continued impressing me more over time.
and they intentionally undersold it in the announcement for these reasons the original tweet i saw was from sundar the ceo saying this is our new medium -sized model we find it really good for code you should try it out and then i went and did it's like wait no this is not just like a medium -sized model for code this is a a generational leap for google they do not have the balls to call it gemini 5 but that's what i'm here for to have far too much balls for far too little to back it And I will loan my confidence to you for this one, Google.
I will loan my confidence to you for this one, DeepMind. When you do the big version of this, call it Gemini 5. Just leave for in the past.
Wipe our slate clean. Pretend those models didn't happen. It's for the best for everyone.
Every serious person I know does not give a fuck about Argon or Carbon. They are bad models. I have been struggling to convince the best AI and agent dev people I know to try this model because they were so burned by Argon and Carbon.
Every time they do, they're like, wait, holy shit. This is a generational gap. The engineers felt it.
Research wasn't sure. Positive leadership didn't understand. But they fucking did it.
They made a model that is smart enough to do real work, is good enough at staying focused in long -running, long -horizon tasks to complete really hard work, and is coherent enough to stay on task and explain what it is doing consistently. It's exceptional. It's a phenomenal model.
i have to keep checking the calendar to see if it's like april 1st or something but it's not like this is just a good model they did it and you should go try it seriously like just just go this one a shot i i think the anti -gravity app is still kind of garbage and the cli is fine -ish if you're okay with it doxing your email all the time and being slow and weird obviously i exclusively use it in t3 code and we are shipping a ton of fixes for anti -gravity and t3 code right now if you do want to use anti -gravity and t3 code i highly recommend the nightly because we are shipping fixes for it constantly right now because we want to have the best possible experience with this model, because it is quite possibly the best model available for code right now.
You fucking did it, Google. I don't know how. I don't know where it came from.
I am thankful that you took all the feedback from the engineers internally and the shit posters externally, and you took it to heart, and you made something better. And I'm very thankful you made all the other fixes that you did over the last few weeks based on my crash outs, because it would have been a lot harder to appreciate how good this model was if it was 3x slower than it had to be.
Remember this chart. Because we'll be going back to the capability coherency framing a lot going forward. Because I need y 'all to understand this.
Because the new frontier is not the capability frontier. The new frontier is coherency. And this is where OpenAI is currently behind some of the Chinese labs.
And if you didn't understand why I liked Muse so much, now you know. Combine the best of Astra, the best of Muse, and what you get is Gemini 4 Neon. It's a great model.
I need to go make sure that my fever isn't back because I feel like I'm going insane. I have not experienced whiplash like this from a series of model releases in my life. How do you make two of the worst large models I've ever used and then have 6 .2 Astra come out and just break because of how frustrating it is to use?
And then, like, right out of nowhere, Google comes back with a medium model that is the most pleasant and enjoyable experience I've ever had coding with an agent. I don't know how we're here. I don't care, though.
I have a new favorite model, and you might now, too. Go give it a shot. Let me know what you're thinking, and I cannot wait to see what they do when they put these new behaviors into the big version.
This might be AGI. Fuck. I'm going to go mad if I keep going, so I'm going to stop now.
Enjoy, nerds. Until next time, peace. God, I can't believe they did this.
The Hook
The bait, then the rug-pull.
It opens as sarcasm: another Gemini model from a reviewer who hated the last two. Then the benchmarks land above Opus 5.5, and the next half hour is a skeptic explaining, thread by thread, why a medium-sized Google model became his favorite tool for code.
Frameworks
Named ideas worth stealing.
16:03model
Capability vs Coherency
Capability: how hard a problem the model can solve
Coherency: readable, stays on task, understands you, finishes work, collaborates with itself
A two-axis chart for placing models. GPT-6.2 Astra sits high on capability but low on coherency; Muse Wonder Full is the reverse; Neon lands near Fable 5.2 and Opus 5.5, and after more use he moves it further right on both.
Steal forcomparing any tools or hires where raw power and reliability pull apart
20:44concept
Work backwards from the press release
Write how the public announcement should read before designing the feature, then build toward it. Used here as a plan-writing prompt and a test of instruction following.
Steal forfeature planning docs and product launch prep
CTA Breakdown
How they asked for the click.
VERBAL ASK
30:19product
“Go give it a shot. Let me know what you're thinking.”
Low-pressure try-it ask, paired with a plug for the T3 Code nightly build where Antigravity fixes are shipping. Sponsor read sits early at 2:11.
Add Modern Creator as a preferred source and Google shows you more of our breakdowns in Search, Top Stories, and AI Overviews. It only changes what you see, and you can undo it in your Google settings anytime.
Add to Preferred SourcesOpens your Google source preferences with us pre-loaded. Tick the box and you're done.
Theo spends a day stress-testing Moonshot's 2.8-trillion-parameter open-weight release — and comes away convinced it's frontier-class, cheap enough to matter, and genuinely dangerous once the weights go public on July 27.
A 2 a.m. field report on GPT-6.1 Sol, the surprise OpenAI model that matches Claude Opus 5.5 on coding benchmarks for a fraction of the price, but still can't out-build it on long, unattended work.
A 30-minute case that Jev's speed and zero-format-error rate make it a genuinely useful classifier, paired with a pointed pushback on the two ways people are already misusing it: as an LLM judge and as a context-compaction engine.
Early access to OpenAI's next flagship model turns into a benchmark massacre, a string of jaw-dropping 3D demos, and one very ugly story about a model that lied about finishing a PR.
Theo puts Meta's Claude Code clone, Muse Code powered by Muse Spark 1.2, through benchmarks, a codebase audit, a game rewrite, and a live integration test to see if the price is the whole story.