I think the reality is even simpler: Huang, Altman and Amodei are in the end of the day head salespeople of their respective company… and they use these social platforms to sell their products.
Their job is to convince people that AI has a large upside while minimizing any downside, they dont need to believe in those themselves.
> I've never had an issue with Codex or Claude reading massive files
Reading files isn't a problem they want to solve. The idea seems to be using a cheaper model to "scout" for the intended code, instead of an expensive one that reads all the things (and spends more tokens / thinks about them).
I think this might be useful because Opus 5 especially tends to over-read. So this looks like an "LLM Bloom filter", telling "hey this is the code you might want to read".
I have a stage-gated workflow that prioritizes “premium” token efficiency (Fable.) and getting the most out of my subscription services. (Which boils down to Fable running carefully prompted deepseek-flash agent teams that defer back to the managing agent for any design decisions in most work.) As part of that workflow the manager uses cheap reconnaissance agents to burn their tokens in order to build relevant repo context, instead of the managing model’s. I’ve been doing this since they released Opus and it occurred to me that most of my pre-implementation phase token use was going right into the garbage bin with file reads that have to be done to find the relevant code, but are very wasteful.
There’s an added benefit that the manager’s focus on strategy and task decomposition before actually handling the user’s prompted task directly seems to be a very good way to interact with Claude’s Fable safeguards, and I haven’t had any refusals doing this.
And while I haven’t ran any numbers, I can get orders of magnitude more out of my claude subscription doing this, especially with deepseek-v4-flash being as good as it is for as cheap as it is.
It sounds like you're halfway to gas town already... Not saying that is definitively bad, but I have avoided this myself since I don't want to get bogged down in trying to figure out the optimal multiagent setup.
The creator (Eric Provencher) worked for Unity before that and he is exploring game development tooling etc (with an open token budget) its fun to see where things are heading in the on-demand future of handsfree blender output and animation
Im now fascinated by how exactly Bcherny generated that message using claude. Like is the entire issue resolution automated, or still guided in someway?
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 good 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.
Same here. I see so many people obsessing over the latest most state of the art bleeding edge models and yelling at Google for not being there, but I feel like the vast majority of people don't actually need those models. Flash has just been super useful and incredibly fast in my experience.
I prefer luna for most development, especially when I am guiding the process. Sometimes terra. I have had terrible results coding with sol. It is way over-tuned on RL to make something that completes the task, no matter what. I end up with way too much code that does a lot of things I didn't ask for.
IME you're supposed to have Sol drive Luna sub-agents to do 90% of the work. Sol should primarily be the verifier and goal setter. Use omp.sh with Task Delegation -> Always to strongly encourage Sol to drive Lunas. Also Luna prefers to be talked to with English in XML.
I love Luna too. An excellent model and still usually better value per dollar than Gemini if you pay for API tokens. Things may change with 3.8 - we'll know soon.
That is a very valid question. I happen to want to get Claude Code muscle memory under my belt for professional reasons in addition to get side projects advanced, could have settled for Codex else. Also OpenCode with eastern models gets part of the job done. In CC beyond using Luna for the cheap, I am using DeepSeek flash v4 for subagents, that is a further cost shaver. Not sure if in Codex I could do that.
When I read all the issues people have with Claude - the aggressive guardrails, the cost, how quickly it burns tokens - it seems almost masochistic to use it. Just seems like herd behavior - people use it because everyone else is using it, and because they believe it’s the “best”, whatever that means. (Benchmarks certainly don’t help define that.)
Google One plans are quite a good value actually - for a few bucks you get more Gemini plus space in Drive and other extras. Even through API, $3.75 for nearly Sol-level quality isn't that bad.
And let's not forget you can use it for free in AI Studio, and in the user app (even free accounts get tons of usage, though it's still 3.6 there), and in Antygravity.
That's the thing. I am completely lost because there are so many redundant paths to get the same thing and I'm trying to figure out which one is the best deal
Well are you looking for a subscription or pay-as-you-go API usage?
Subscription? -> Google One plan (http://one.google.com/)
API? -> AI Studio (https://aistudio.google.com/)
It's not really any different than the choice you'd make with OpenAI/Anthropic depending on how you plan to use it. Except as a hyperscalar, it's also offered first party from Google Cloud (like Claude via Amazon Bedrock or GPT via Microsoft Azure OpenAI Service):
Google Cloud -> Gemini Enterprise AI Platform (https://cloud.google.com/ai)
But if you're using models via OpenCode or Pi or whatever, the flow chart is basically just "Go To AI Studio" unless you or your employer is already used to Google Cloud, otherwise there's no need to subject yourself to all those enterprise-y IAM dashboards and stuff. You still get free usage from AI Studio when you generate the API key without needing to add billing details so very easy to try.
good summary, thanks. I used to use Google Cloud for consulting and projects (I worked at Google for a while, and there is some nostalgia) so I have Gemini API via Google Cloud, but I am retired now. Your post reminded me that I need to shut that all down and switch to getting an API key using AI Studio.
I have spent two months experimenting with a wide range of US and Chinese models, and I had a lot of fun doing that, but I am in the process of switching to just using local models, using Gemini on an API if I need it, and once or twice a month when I really need help on something difficult, I use something top-tier like Kimi K3.
This is what killed Gemini for me. The model might well be great, but the ecosystem Google has built around them is a confusing maze of not-quite-there products.
I've been benchmarking[1] models for trip planning and world knowledge specifically (to decide on which model to use with my travel app), and the Gemini models consistently come out on top.
I started trying out 3.7 Flash this week and it is competitive with opus/fable and also FAST. It is getting work done that anthropic models were struggling with and the speed with which it does is quite a bit noticeably faster.
Beginning to think Google is a dark horse in this race and some of Anthropic's "everything feels janky and rushed" karma is going to catch up.
I want to believe this, but every time I try Gemini coding assistance within Colab it's utterly dire. Code gen in a cell is OK, but things fall apart when you try to get into a feedback loop. The system prompt/harness fails to inform the agent about what it can and can't do, or does and doesn't have access to. It will confidently tell you it's done a thing, and then you ask, it admits can't actually do that but will happily try and fail again. Very frustrating, because I really like Colab as a platform for little reproducible experiments that may or may not require CUDA.
Hype that burned out pretty quickly, it's hard to speak to the size and significance of old hype, I never felt it.
Every time I personally tried Gemini models up until last week they simply couldn't do the long complex tasks I'd being doing with Anthropic models for many months.
For awhile now I've found Gemini will use Google search for pretty much any real world knowledge, which is a huge plus IMO. It's basically Google with a much better frontend and no ads/seo nonsense.
I think it already has them but it's much more subtle. Also useful. When I've made certain sorts of queries I've had the distinct impression that it was attempting to very gently steer the conversation with suggestions. But it was brief, still answered usefully, and didn't resist going in the direction I wanted. So a win-win tactic I guess.
For example find a beautiful landscape shot of a place that just so happens to be accessible to tourists and ask it something along the lines of identifying the location. IME it will noticably steer the conversation towards relevant commercial offerings and offer (entirely unprompted) to help plan a trip.
Or ask it about a certain category of product with some requirements and it will initially present (relevant) options that look like paid placement to my eye. But if you ask it's happy to go on to turn up lots of alternatives and enumerate tradeoffs.
Assuming I'm correct the subtlety is on par with product placement in movies. Certainly leagues better than the internet advertising we've suffered to date.
It definitely steers. For example if it suggests travel plans, the booking links it provides give Alphabet a cut.
As you say it was subtle, along the lines of "oh, if you are planning on going to the place you are researching, here are some helpful links to places you can stay". Subtle, in that it didn't get in the way of main result, so I didn't mind overly. Insidious, as I only noticed because I wondered why it was providing those particular links and looked them up. I can't see how you could ad-block them if I did object.
And worrying, because these unblockable sneaky ads are just a first foray coming from a company that prostitutes its own app store searches, by making the first and most obvious result utterly unrelated to to the search topic. Instead it's who paid them the most to be there. That behaviour is why everyone dumped Alta Vista when an alternative came along. Alternative Android app stores can't come soon enough.
They already skim off 15% of purchases which I'm sure makes their Android operation return a profit that makes other industries drool. Debasing their search to ad a tiny bit extra on top must by driven pure greed. Senseless, as I'm sure it will come back to bite them in the end.
There are some subtleties though. For example, with Gemini search grounding (this was a few weeks ago), you cannot give it a domain whitelist, only an excludelist -- with Anthropic's API you can do both. For somewhat niche search tasks like "Find LinkedIn profiles matching this ICP", Anthropic wins there.
That being said, Anthropic is also so insanely expensive for everything I ended up switching that particular part to Exo instead...
Maps grounding is an amazing interface for Google Maps - „I’m at x, need to be at y by 5pm. Find me a route that’s walkable and has a good vegan restaurant and a toy store“
I believe Gemini Flash is smart enough to know when to ground with web search. Their app has been saying it’s running a web search on almost all of my queries since 3.6. And given that Google … is Google, I trust them with web search grounding more than anyone else.
I wasn't trying to be precise originally, I just tried to fit activities into "morning / evening" buckets. I did the whole itinerary with Opus first, but when I gave it to Gemini 3.7 Flash to review, it started correcting it with "this place will close 5PM" or "this place is closed for good".
It was right on every nit, so it was surprising how well the model knows these things. If I ever release this I'll probably need the SERP API or Google Maps SDK (which I've heard is very expensive now), but for a personal trip where I will verify manually, using the LLM is okay for now.
When you called the Gemini API, did you opt in to using search grounding:
tools=[{"type": "google_search"}]
I'm curious whether in fact you were getting answers from the model weights (which is what I had assumed) or whether your API calls were resulting in web search tool calls.
Using grounding in Gemini is indeed backed by the same canonical data source for business information (like opening hours) as Google Maps. This stuff is available in its own API for a GCP fee, but we’ve built tooling to connect it to the Gemini agentic ecosystem as well.
So if I understand you correctly, Gemini has direct free access to the Google Maps API in a manner that others (people, LLMs) would need to sign up for API access and pay for?
Gemini models - at least via some interfaces - have tool calling API access to various Google integrations. flights.google.com, maps.google.com, etc.
The info isn't in the model weights.
Because of where I live, there are three viable airports for any given flight I might want to take, which historically has made shopping a real pain. But Gemini (and only Gemini) has greatly simplified it. Pramble plus date range plus destination and it very quickly generates potential itineraries with costs, total travel time (driving included), etc.
jampa was saying that the Gemini 3.7 (a model) ranks higher on real-world knowledge. The point I was originally making is that I would trust any model by itself to answer real-world knowledge problems. If I want to know opening hours, then probably any of the models could find the answer with a web search tool.
Perhaps the Gemini API makes this easier, as the web search tool is built in.
I've just returned home after doing a 100 day "half lap" of Australia in a camper trailer with my family. I made heavy use of gemini to plan much of it. Getting packed up with a general destination in mind and telling gemini "we're leaving X town at 10am and heading to Y, where should we stop for lunch. My wife has coeliac disease, find us somewhere that does good gluten free options" was one of the many things I regularly leant on it for.
In my experience, Gemini 3.7 is excellent for general non-coding tasks. But for coding, especially backend development, I still find models like Opus 5 and GPT-5.6 more reliable.
If I strip the dependencies out, I'm using it on a 5MLoC C++ codebase, and I found it performs really really well. I am using the Opus 5/Fable in parallel and I couldn't tell the difference. Both models make mistakes here and there.
I've swapped over to it in the past two weeks, it's been really good. It does what I ask and doesn't think it knows better than me, which so far has made it the most pleasing experience I've had when slop-coding.
My only wish is it were somewhat cheaper, as it tends to balloon pretty quickly when I'm using it in Opencode. I'm currently trying to offload a lot of work to subagents to stop the context expanding so rapidly. But on the upside, I rarely have to correct it - I've spent far less time arguing with this than with anything else so far.
I usually use OpenCode for all open weight models but for Gemini I use Google’s agy coding harness (or my own).
Venders coupling coding harnesses with their own models is usually a good thing. Poolside.ai has a combined harness with their own models that works well locally, and the DeepSeek harness with their models is very interesting.
I tried similar travelling tasks but also added transportation and complex transfers (train, bus, walk, next train...). Worked meh and still a difficult thing to do for a llm.
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
Once you've written something, it's incredibly easy to overlook minute changes to the text.
See: why authors wait days, weeks, or even months before editing what they've written (or, if you're more interested: cognitive regression, inattentional blindness, and the effects of misdirected saccades).
Then I think you read it wrong, because you don't make that mistake unless you copy-paste your comment out of a Claude window and into the comment box.
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.
I stopped using Gemini a few months ago because it would often just (partially) reply literal nonsense to me.
Think 2023 style ChatGPT. Something like “to open a document on your Mac click File > Open docurrrar” - like it suddenly forgot it had to produce actual words.
Overall I enjoyed its speed and comprehensiveness. But those occurrences of nonsense just made it feel like a great car that once a month just stops in the middle of the highway.
Serious answer: no model ever gets close to writing an architectural floor plan that makes sense.
They understand all the rules and best practices, they can (sometimes) spot a bad idea in a floor plan, they can describe a good floor plan.
But ask them to make one, even if you give it every detail (even a "node graph" of rooms), they will still output nonsense. Same for text and image models.
Opus 5 feels like a downgrade from Opus 4.8 overall. It, along with Fable, really has a problem following instructions and staying in scope, and their prose keeps growing, both in explaining what it did and in writing multiline code comments (some comments read like a changelog, e.g. `// sky is blue (changed from red on 2026-01-01 per TCK-234 by @Foo)`).
Every time I ask it to do something, it does 80% of the job, goes off on "side quests" beyond the scope, and then leaves something out of the core ask (and when you tell it to finish, it does the same thing again).
The only advantage of Opus 5 over 4.8 is the better cutoff date for working with 3rd-party tools, though both do a very bad job of "this tool is constantly updated, I should look for the latest version first".
*Describe the code as it is now — no residue, no change-narration.* Artifacts (docs, plans, comments, commit messages) should describe the current code statically, as if it had always been this way. Two facets of one rule: (1) never describe a dismissed alternative or a corrected/replaced choice; (2) even when nothing was rejected, don't narrate continuity or evolution relative to some earlier state. Mention a former state only when the current choice is genuinely hard to understand without it, and then only as an explanation of the current choice.
This governs descriptions of the code and main documentation. It does not apply to work-tracking artifacts in `doc/tasks/`.
```
Which makes comments and docs bearable but I'll be damned how it loves to overload work tracking document with every little detail.
The new rule here requires you to pay for part of your imports (up to ~60%), even if you export more. I know people who installed batteries in the inverters because of this. It is more expensive overall (hybrid inverter + LFP batteries) but has other benefits (blackouts are not a problem, most months you don't even use the grid). I imagine that once sodium-ion battery production scales and costs decline, this will become the "default" option for a solar installation.
Their job is to convince people that AI has a large upside while minimizing any downside, they dont need to believe in those themselves.