theturtletalks 3 days ago

I’ll give it a spin on my software factory. The other Gemini models looked good initially and in benchmarks, but on long-running sessions and orchestration, it’s fallen short.

We’re reaching the point where certain models are really good at 1 thing. The person writing this blog post isn’t using Neon for large-scale orchestration, but I’m sure it excels in its usage. For me Fable 5.2 as the orchestrator and Opus 5.5 and grunt workers have led to the most robust software...so far.

  • seizethecheese 3 days ago

    > The person writing this blog post isn’t using Neon for large-scale orchestration, but I’m sure it excels in its usage

    I'm the person writing this. The data is from head to head battle data by real developers on top of OpenCode. The data is generated by an open source app so you can inspect it yourself (https://github.com/bottomless/agent-duel). The hitch is there's no real Neon data. This is purely projecting from Carbon's super slow but very good performance onto Neon's likely fast and likely equal performance.

  • getcrunk 3 days ago

    Software factory?

    • klodolph 3 days ago

      Multi-agent coding orchestration setup.

    • rogerrogerr 3 days ago

      Token incineration plant.

      • esseph 3 days ago

        Lol, but I concur. If you fork a bunch of long running agents on frontier models, it can cook your budget / allocation.

        • Implicated 3 days ago

          Dear god, as someone who has built a wildly robust "software factory" (mine doesn't stop at the output) - don't fork sessions. Just don't. Distill the relevant context into a new session, however you choose to do that.

          • esseph 3 days ago

            Sorry, I don't mean session forks, I'm discussing agents / subagents really. Each with only a small task, returning only their output to a main agent / context.

            But yeah, even with limiting thinking and using lower cost models for sub agents, it can get spendy quick.

    • kaydub 3 days ago

      Rube goldberg machine where tokens are burnt and code is sorta-kinda created.

    • SimianSci 3 days ago

      An impressive showcase of productivity theatre in which tokens are incinerated in exchange for software that is always never "finished" and will continuously let down its users.

      Some people have cleverly parlayed this new fad into money, but this is typically a result of their sales techniques rather than their software.

      • nonethewiser 3 days ago

        I havent tried it but this was my take as well. Forget token inefficiency, it just seems much harder than just prompting a single agent what you want

      • reassess_blind 3 days ago

        Replace “tokens” with “man hours” and it sounds like a software company.

      • UltraSane 3 days ago

        Gas Town and Gas City are textbook examples.

        • xtracto 3 days ago

          Just tried this today. Installed gascity, set it up with "gastown" settings, added a couple of repos and pointed it to a [bot to crazy] jira ticket .

          It just make things so complicated... and in the end I had to chat with the "major" interactively, and I didn't see any benefit vs a simple claude session + a "work with 10 agents" prompt. It just seemed complicated for the sake of it.

          I want a system that hides the complexity, and is effectiveand efficient. But we're still not there.

      • nsonha 3 days ago

        I have not built or used such a thing, but it's tempting to try to build one. Sure I can go micro-manager my agents and improve software the way things currently work too, but it doesn't seem like this industry will just stop here, now seems like the exact moment we should try to figure this stuff out, regardless of how disappointing the result currently is.

    • RGamma 2 days ago

      Eh, you know how this goes, right?

  • pkulak 3 days ago

    I never use long running sessions. I like one small task at a time that I can understand and evaluate. If the only downside is worse one-shot GTA homages, that’s a pretty good trade off.

wmf 3 days ago

Projecting something that is already out is an epistemic crime.

  • seizethecheese 3 days ago

    Author here, guilty as charged. To be fair, we don't have early access to any models (just getting started).

    The blog itself can be seen more as an exploration of how speed impacts coding agent battle votes, than anything specifically about Neon.

    Applying the speed impact to Gemini Carbon is particularly interesting, though, because it's the top model in our leaderboard when you exclude early votes (https://agentduel.com/leaderboard?excludeEarlyVotes=true) but mediocre when you let developers vote. Gemini literally takes 3x as long as opponents in matched battles.

    • NBJack 3 days ago

      Why not just...post later when you have data?

      • outside1234 3 days ago

        Engagement

        To be fair, I don't have the model either but after reading this I will give it a try since Opus 5.5 is my go-to right now.

      • seizethecheese 3 days ago

        It was an excuse to write about the impact of speed on battle votes, and I got lucky that people wanted to talk about Neon.

    • wasabi991011 3 days ago

      What makes something an "early vote" and a "developer vote", as opposed to just a "vote"? And what makes those categories meaningful to agent duel?

      (Also, out of curiosity, how are you funding this free inference for frontier models Is this not massively abuseable?)

      • seizethecheese 3 days ago

        > What makes something an "early vote"

        When you run agents in a battle, one always finishes first. We let users vote for the finished agent without waiting for the agent that's still working. This is contrary to typical Arena methodology (like Arena.ai) but I think it's appropriate for agents since they can take dramatically different amounts of time to finish tasks.

        This is why DeepSeek 4.1 Flash is doing so well on our leaderboard, it's super fast and often sufficient.

        > and a "developer vote"

        This is confusing, sorry. Inspecting our data it's almost entirely people doing software developments with agents. It's just a vote.

        > (Also, out of curiosity, how are you funding this free inference for frontier models Is this not massively abuseable?)

        This is a pivot from Bottomless (YC W19). That business is doing okay and funding our runway, in addition to fundraising back in 2021.

      • wasabi991011 3 days ago

        I see, thanks!

        I appreciate the responsiveness, not just to me or this thread but I like seeing you around HN

rootusrootus 3 days ago

I guess what I really want to know is how Gemini 4 Neon w/Antigravity compares with Opus 5.5/Claude Code. I'm not really into using LLMs in a particularly advanced way, I just use CC like 99% of the time when I need to write something, and then maybe Codex to second guess it.

But that's about as complicated as I get. I have recently tried to use Copilot because that is what work wants us to use, and even with Opus 5.5 it is noticeably inferior to CC on the same model. Wild. So clearly the harness matters.

  • seizethecheese 3 days ago

    Totally possible for your use case, though the analyses I’ve done on harness model interactions show it’s actually not that important for benchmark performance. https://micjm.com/blog/on-model-harness-interactions/

    For what it’s worth, the linked article data (Agent Duel data) is all run on OpenCode. I’m curious if people would prefer we run each model in its native harness (at the expense of user experience because it’ll be buggy).

  • thefunnyman 3 days ago

    We had CC for a while and now have to use GHCP. As such I haven't used CC in like 6 months but around the time I switched I felt like GHCP CLI had pretty much caught up. Curious to hear what you feel are the gaps? Are you using the cli versions of both?

    • rootusrootus 3 days ago

      Yep, CLI versions for both GHCP and CC. To be honest, I haven't done enough actual coding with GHCP yet to get a feel, so my original comment may be too broad. But I've been repeatedly comparing and contrasting with CC for code reviews (as a team lead, I do more of that these days than coding). I have been pretty consistently getting about half again as many bugs from CC as from GHCP. And very nearly always correct, not hallucinations.

      My company is slow-walking (and that is an understatement) our LLM usage, so the majority of work that I'm reviewing is human generated. After the stories I hear of what other people are experiencing when management is more enthusiastic, maybe I should count myself lucky.

  • pritambarhate 3 days ago

    Also last I checked there was no clear flag to turn off training in paid plan of Antigravity. So if you care about that looks like Google doesn’t want you as their customer.

    • riskassessment 3 days ago

      I think you just turn off telemetry in antigravity to disable training. The Gemini app (including the web chatbot) is much more frustrating: the only way to disable training in the Gemini app (for a personal account) is to turn off apps and history.

  • scottyah 3 days ago

    The models can be very optimized for the harness, Clem (of Hugging Face) just recently released a Qwen3.8 model that used RL training on the harness and it got some pretty significant gains in the benchmarks.

    - SWE-bench Pro (hard 51): 39.2% → 64.7%

    - Terminal-Bench 2.1: 39.5% → 55.8%

    https://huggingface.co/harness-hub/Qwen3.8-27B-pi-adapted

    • seizethecheese 3 days ago

      I'll call bullshit on that.

      If this were true, models would do way better in their own harness than in others, which doesn't seem to be the case (effect is minor).

      Your link has a "link to the release blogpost" which is supposed to explain the training recipe but the link is down.

      • jdiff 3 days ago

        Could have an outsized effect on a smaller model.

      • rubyn00bie 3 days ago

        On a benchmark it might be minor or bullshit, but for me on real projects Claude Code has produced vastly better results than I was getting via Cursor or OpenCode. I used to believe the harness was negligible to having the agent do good work, but now I think it’s essential.

        The big “aha” moment for me was seeing the wildly different results I’d get using something like Claude Design vs just prompting (even when using leaked system prompts). I then started hearing about the special harnesses available for biotech, science, and math that seem to be available for researches via OpenAI and Anthropic directly… which sort of further cemented in my brain how important it must be. I wish I got the same quality from Cursor, or OpenCode, as I do Claude but I just don’t.

  • anukin 3 days ago

    Antigravity in general is a poor harness. I have had better experience using Gemini 3.8 flash over api than using it over agy.

    • cultofmetatron 3 days ago

      and to add insult to injury, they dont' let you use your subscription from a proper harness like opencode

  • TacticalCoder 3 days ago

    Ah that's a new angle. After the "Google models are behind" and the "Google models are slower" and the "Google costs more per tokens", we're now at the following phase:

    "OK, google models perform just as well, are faster and are cheaper but all that doesn't matter. Because the models, now that Google has very good ones, doesn't matter after all. All that matters is the harness".

    And of course it's upvoted by the Anthropic shills.

    • ddxv 3 days ago

      A lot of this is brand affiliation as well.

    • vrosas 3 days ago

      Don't forget "I couldn't figure out how to buy a subscription" and "The names are confusing"

    • rootusrootus 3 days ago

      That's what you got from what I wrote?

      I must not be deep enough into the conversation to get it. I was just sharing my own perspective. I don't follow AI development nearly as close as a lot of HN does. I don't hate it, but I don't love it either, it's just part of my job now.

      > And of course it's upvoted by the Anthropic shills.

      Gotta tell you, my post is barely above water ;-). It goes up, goes right back down, plenty of people seem to agree with you that I'm being mean to Google or something.

    • retsibsi 3 days ago

      > "OK, google models perform just as well, are faster and are cheaper but all that doesn't matter. Because the models, now that Google has very good ones, doesn't matter after all. All that matters is the harness".

      Have they edited their comment since you replied? If not, this is one of the most egregious strawmen I've seen this week. You have brought a lot of your own baggage to this discussion, and stuffed it into a comment that said approximately none of what you claim to have taken from it.

  • TiredOfLife 3 days ago

    It's really simple. You somehow trigger an account ban for Claude you lose that account and create another one.

    You get the same with antigravity and you get your whole Google account banned

bwat49 3 days ago

I'm pretty impressed so far, this is the first google model I've used that feels good for coding

  • tcoff91 3 days ago

    Is it still dog slow like other gemini models?

    • nicce 3 days ago

      Gemini Flash is one of the fastest?

      • tcoff91 2 days ago

        From what i've seen, carbon in agy harness is really slow at accomplishing tasks compared to opus in claude code

        TPS is not the full story

  • elxr 3 days ago

    What harness? And how much do they subsidize compared to the equivalent chatGPT sub?

    • bwat49 2 hours ago

      I use the antigravity harness through t3 code

      the usage limits don't seem quite as good as either claude or codex, but I'm on the $20 plans (not sure if the calculus changes with the more expensive plans).

      I'm currently using the $20 plans from claude, chat gpt and google which gives me more than enough usage for my needs (and I'm considering cancelling the gpt plan). Since the ai pro plan includes youtube premium lite I can also get rid of my separate youtube premium sub

quinndupont 3 days ago

Fun analysis. I guess it’s not surprising to see the need for speed™ as an important factor, but who is using Fable for run of the mill coding work? It’s too pricey!

  • seizethecheese 3 days ago

    Yeah, check out DeepSeek performance on the leaderboard that's the source of the data of this post (http://agentduel.com/leaderboard). It wins ~half the time because it’s lightning fast and good enough.

    • Rodmine 3 days ago

      Finally, something that has actually reflected my real experience with these models.

      All the rest I have seen are victims of benchmaxxing and therefore simply misleading.

      • seizethecheese 3 days ago

        Really happy to hear this. We're trying to build a leaderboard that reflects real work done by real people. Even the Arena systems are run in cloud containers.

        Genuinely hoping some people sign up from this and use the app so we can get some more high quality data (it's open source: https://github.com/bottomless/agent-duel).

  • hedgehog 3 days ago

    I used it to write something last night that was intricate enough that Opus would have involved a lot more rework. Subscription discount though.

  • elxr 3 days ago

    When you're making >$240K a year, I imagine spending $80-90 a day trying all the models just to find a slight edge is barely noticeable.

fud101 3 days ago

I'm blown away by Neon, it's as good as Argon (which was amazing) but fast and cheap (i think). I've used primarily Claude for about a year and now i've tried Google's newer models in the last week, funnily because i ran out of tokens testing Fable and to me Opus is trash, Fable is trash - I will just use Opus for Web Day Job which it handles with ease and keep Neon for my serious side projects (ML related), which Claude can't handle well at all. Ive been working on a project for about 2 months, Claude Opus was a dead end and even with the new Fable which didn't move the needle and gave me negative progress, but Hydrogen on things changed and i've finally made good progress. Frontier model indeed. Scary good. Wow!

JuneWW 3 days ago

It's much better than any Google's model before, but still needs many harness.

byzantinegene 3 days ago

the timing of this release almost feels like a stab toward anthropic. which is weird because google owns 13% of anthropic

  • lofaszvanitt 3 days ago

    Oh it's ugly. That's their way of showing how they love each other ;).

  • thisisauserid 3 days ago

    Google is committed to buy $10 billion in stock so they don't have to commit to $20 billion elsewhere with them.

    The lower the price, the more stock it buys.

jdw64 3 days ago

I’m considering AntiGravity because its subscription usage seems lower than expected. But personally, with Gemini’s Carbon, Hydrogen, and Neon, it’s confusing because they don’t show which model is mid-to-low-end and which is flagship. OpenAI’s model names—Astra, Sol, Terra, Luna are intuitive, like cosmos–sun–earth–moon. Google’s model names aren’t intuitive. If we go by abundance, Hydrogen should be the top model. Even Anthropic’s Fable–Opus–Sonnet–Haiku is intuitive. Google’s element naming is not intuitive at all.

  • itemize123 3 days ago

    looking at the cost - Hydrogen, Neon, Argon ~ Carbon. intuitively they are by 'weight' of the elements atoms

    • naniwaduni 3 days ago

      "Argon ~ Carbon" is a bit of a "have you looked at a periodic table recently?"

  • ijidak 3 days ago

    It's wild that Google can fumble something as basic as heirarchical naming.

    I have no idea whether argon, carbon, or neon is their largest/best model.

    If they think we're going to pull out a periodic table to understand the naming scheme, they shouldn't be in charge of product.

    All of the providers have chosen a somewhat nerdy naming scheme, but that's fine given the current mostly tech customer base, but at least Anthropic's and OpenAI's are intuitive.

    But, the order of elements in the periodic table is not something most people have memorized unless you do something like hydrogen, helium, lithium, maybe.

    Or maybe iron, copper, silver, gold.

    Maybe at least pick elements that are in horizontal order, like carbon, nitrogen, oxygen.

    Just so many basic mistakes in every aspect of product rollout at Google.

    • jdw64 3 days ago

      I hit the upvote button.. These days Google seems focused on B2B and not interested in B2C. Honestly, it’s too non-intuitive. Before even memorizing the periodic table, it feels a bit ambiguous. It’s not even that they get bigger in order of atomic number.

    • trvz 3 days ago

      It doesn't even make sense when you look at a periodic table, at least without a chemistry degree.

      The elements are: hydrogen < carbon < neon < argon

      But the models are: hydrogen < argon/neon < carbon

  • tym0 3 days ago

    They're just naming them alphabetically as they release.

    A commenter on an other thread theorised that because with time bigger models haven't really meant better models, Google would like user to focus on the "subtitle" instead ("Everyday help", "Deeper analysis", etc...)

    Or in another word modern models seem to be harder to map on a linear axis so they just gave up and use descriptions instead.

afavour 3 days ago

Looks like a good bunch of speculation in there but no bad news to be found. Just as good but cheaper? Sign me up.

For this model, I mean. And not the Anthropic IPO.

  • seizethecheese 3 days ago

    Lots of speculation, yes.

    Just as good but cheaper comes from the benchmark performance. Also faster from openrouter. The opus 5.5 parity projection is an internal model based on those assumptions from the Agent Duel team.

  • enraged_camel 3 days ago

    Nobody with money is gonna look at this model and change their opinion on the Anthropic IPO lol

lacoolj 3 days ago

First, this post has a lot of words like "assume" and "projected". What are we actually reading here?

Second, I will be trying this with agy when my Opus 5.5 limit hits. I only hope its faster than Argon which is already dog slow

ralusek 3 days ago

This title reminds me of the doctor from Arrested Development.

"It looks like he's dead."

"...He's covered in blue paint, or something."