YHK - yet another hacker news clone

I have an 'ask' alias in my shell that just uses Haiku. I use it daily for pretty much everything

For more long work, I now use Fable to create a PLAN.md. I tell it to make a plan that will be executed by other models, and most of the time it ends up choosing Opus or Sonnet.

I didn't start doing this recently. Before that, I would just use the top model for everything. Splitting the work across different models depending on the task has helped a lot. They run faster, and I usually get much better results

mimo-v2.5-free, mimo-v2.5, deepseek-flash in that order, honestly don’t even bother using qwen3.6-35b-a3b now unless i need uncensored tasks finished, mostly reverse engineering

most engineering tasks don’t require frontier llms

when they get stuck, then i consider moving up to more capable models

purchasing a claude plan seems widely unnecessary to me

the tasks they do better than the average engineer cut both ways: unless you have an existing portfolio of well written and designed work done pre-llms, it looks like you’re producing slop that pretends to be well designed

poor typography choices despite using the mode,

poor layout choices despite using popular CSS frameworks

etc

bad engineers will always be bad engineers

tools don’t make up for it

edit: a follow up to this— everyone is using eyebrows in their layouts and have no fucking clue why it was done to begin with

everyone has a status pill floating above their front page hero display text and its not fucking status related

so gross

opusplan, for the same reasons.

Astra low for pretty much everything gets me through the work week with a bit to spare on the 5x max plan.

I've found good success with the new Gemini models on Antigravity. Granted I use my models either:

- like a fancy auto complete (here are some stub methods, they should do X, fill them in)

- using fairly detailed plans and test harnesses, so blowing up the world is hard

The 3.X Flash family have been fairly capable models, and the selling point for me is just raw speed. Gemini is noticeably faster than the competition, about 3-4x, and I just get work done faster with it.

That said I'm keeping an eye on Open Weights. DS4 Flash was good until price hikes, and finding a provider that serves at high speed and without quantisation at the prior price is tricky.

Claude, not necessarily because it's better. I find deepseek flash to be of similar quality. But because it's so heavily subsidized.

I think for me stuff peaked around Opus 4.7, I was leaning heavily on the model with paired supervision from reviewing the output manually every step of the way. Ever since that things got a little more complicated and in an unsustainable pace for me, I am trying to remove myself from the equation and build verifiable and reliable tests with quick feedback loops that let frontier models run autonomously but in all honesty not seeing it scale well, I need to take a step back and reassess if the trade off was worthwhile. Frontier models are being incredible at making me feel like they passed my tests only to eventually reveal some tech debt that forces me to take large pivots. It seems the speed is sexy but the results are questionable, models may have hit a limit in my workflow and I think harness engineering is more important than anything. Would love to hear feedback on this take and if others have experienced similar things and what they did to overcome this. (Context would be solo founder bootstrapping greenfield work with full autonomy and sometimes more room for rapid iteration)

- Default: Codex with Terra Max (because it's crazy cheap)

- Preferred: Claude Code with Opus 5 Medium

Qwen 3.8 flash-next or Gemma 26B because I care about responsible and sustainable usage of LLM's.

Glm-5.3-flash for me. I like faster models because I stay involved all the way through. I don't delegate full control to the agent because it's harder to understand the end result that way.

I used to main Claude, but I can't stand how it writes. I feel like I'm wasting too much time trying to decipher the output. Adding writing rules does not seem to work. Now I use GPT 5.6 Terra high fast mode, with Luna for everything else. I might consider using Sol for planning. I can't stand using Sol or smarter models for coding, because they will eventually try to rewrite everything in the codebase.

I also don't want to use the Claude Code and Codex agent harnesses. The good thing with Codex subscription is that it can be used in other harnesses, unlike Claude. As far as I know, only Anthropic has this restriction.

I'm using a lot of gpt-5.6-Luna and glm-5.3-flash. Astra is really fantastic but it's too expensive. I average about 50-90B/tok/mo.

Muse Spark 1.3 Contribs

unbeatable price/intel ratio per M tokens:

$0.10 (input)

$0.20 (output)

$0.002 (cached-input)

I was using various open weights models until glm-5.3-flash came out recently. It's incredibly capable and cheap, even if it's very verbose and not the fastest. I've assigned it to all my agents across my harness and it's getting the job done. Still needs a good steer every now and then, but a great work horse.

I use gemma 4 12B and Qwen 3.8 9B

Sol High-Xhigh, and Opus 5.

Granted, yesterday I threw a few tasks to Astra which the former 2 botches; it produced clean, correct solutions quickly, so pending further eval, this may take over.

IMO unless it's a mechanical tasks, it's worth it to use carefully -crafted queries on the more expensive models, than iterate through messier solutions on the cheaper ones.

I use Sonnit for my personal work and Opus for my professional work.

For my personal stuff, I'm on a small $20 plan, so I need to use tokens conservatively. I was very rarely exceeding limits until I built a Dark Software Factory. It's not as efficient at token use. So, I use Sonnit over Opus here.

At work I have a $100 plan that I rarely exceed so I use Opus. I have access to Fable too, and I did use it a lot while it was new, but I don't find it improves most of my work by too much. I do mostly bug fixing across several hundred repositories with hundreds of thousands of lines of code, mostly written by humans over the past 20-years. These projects interact with each other so I run claude from the root of my sandbox (I was nervous to try this but I'm not looking back now).

I also use Sol as a secondary for my personal work. I pay for it because I like to talk to ChatGPT on the web. Since I already have the subscription, I let Sol write plans for me. It does a better job at certain tasks and it saves me some Claude tokens. Maybe I should consider Terra for the task, but I don't run up against my usage limits for the little bit I use it.

I'm trying to use Gemma 4 12b for some workloads but I haven't mastered the model yet. It's still very experimental for me. I can get work from it but it takes a lot of hand-holding. For local models, however, it's all I have the RAM for.

GLM5.3 - I'm on one of the ancient plans, meaning basically unlimited.

...and then sprinkle in some other models when i think a second opinion will help

I like AI more on the short leash, giving it specific agents tasks, one at a time (centaur mode). I found GPT 5.6 LUNA to be astonishingly capable. Plus, it so fast, that it does not block my flow of throughts, like the more capable but slower models often do. And it is so cheap that I do not use Ollama local models anymore. Luna is far more capable, and so cheap that the energy prices here in Germany eat the gains of local hosting ;-)

Reading through this thread, I notice not many people are using models from the Chinese AI labs ...

If I had to pick just one, these days it’s Composer. It is just a cheap, fast, surgical workhorse. My workflow uses other models as well for various phases: Opus for planning, Sonnet for review, etc.

I prefer Cursor at this point just because of Composer. Claude Code is passable but the lack of a good, fast, cheap workhorse sucks. Sonnet and Haiku aren’t it.

I also do like Codex and Sol, Terra, and Luna. They are decent but I don’t find they stand out enough to use over the others.

Additionally, I have not tried Astra and found Fable to really not worth the cost for the tasks I do.

Finally, the latest Grok is actually a beast of a model, but expensive enough to not be a stand out.

ZAI/GLM because a friend has generously shared an API key. They have other subscriptions and also a lot more from their employer.

A day ago I activated Google AI Pro free via Google's tie-up with a local company (I do pay for this company's product though and it's anything but costly). Now I will use this too.

No other reasons to pick these, or not picking anything else.

Gemini is the least stupid/most neurotypical model family IMHO. Works great for me

I feel like the odd man out ITT but Terra has been absolutely amazing for me and totally blows Sonnet out of the water. My company pays for Claude so I use Sonnet in the office but at home I use Terra and I find it far more likely to one shot some very decent code.

Glm 5.3 Flash and DeepSeek 4.1 Flash

Context will be important for answers to be meaningfully comparable:

- LLM expense budget

- What type of dev: work, personal, real time spaceship thrust vectoring, html contact forms for family, etc.

- human in the loop with short as possible turns, software factories that can run for days, or something in the middle

Just off the top of my head. I'm sure there are others I'm missing.

For personal needs I use a local Qwen3.8-Next-Flash setup on a GB10 cluster. For work, Guthub Copilot with either GPT 5 mini or toss up between Opus/Sol depending on the complexity of the task.

Used to pay for a Claude 20x plan and did everything in Opus, but I hate how it talks now and recent events (OAI scooping, Anthropic's spying, third party Chinese model hosts stealing and selling credentials) have really pushed me towards local AI for personal needs. Am not allowed to use Chinese models for work even if self-hosted so not much choice there.

Qwen3.8

Gemini 3.8 flash via Antigravity desktop has been my daily driver for the last few weeks. I also use DeepSeek V4 flash and a Qwen 3.6 MoE for local stuff.

For personal use, Luna-xhigh on the Plus subscription. I only ever hit my limits if I decide to use Sol or Astra for a bit more quality.

For work, given we pay API pricing anyway, I've been happy experimenting with K3-medium as my default, and now I'm planning on trying GLM 5.3 as well. Sol-medium is my fallback for my work purposes but I occasionally use Opus 4.8 as an additional reviewer.

I like the open weight models because much of my time at work is spent on security hardening (specifically, hardening my own service), and Sol/Opus keep snitching on me and blocking my prompts.

Enterprise pay as you go across a few vendors. But still cost conscious.

Defaulting to Sol Medium/Light for most planning and implementation I think is tricky or want more care in.

Luna Extra High for everything else (implementation, tedious take over my browser and do stuff).

I read all of its output tokens and lots of thinking tokens to understand the general flow of things, but only minimally look at code these days. I can't grok what Claude models speak and it's gotten worse. OAI models speak my kind of tech language I guess.

Light human review, some automated review.

Most work is for internal use.

Sonnet. It makes my token budget go further and I prefer to have the LLM churn away on rote work while I do the thinking. So while Opus and Fable are undoubtedly smarter, that doesn't materially affect _my_ workflow.

I'm not as up to date on the other vendors' models, but when I last used Gemini my feelings were similar between Pro and Flash.

Given the tremendous variety of answers here, one has to wonder how much it matters. Reminds me of when you ask a bunch of motorcyclists what oil to use

Terra Medium/High for most implementations, Sol High for tracking down hard bugs and planning more complex systems.

It pains me to read these answers so far. Listen, for 99%+ of your web and mobile tasks, deepseek-v4.1-flash is all you need. It is blazing fast, super cheap, and quite proficient. It acts responsibly, has top-notch vision for evaluating its own UI work, and is far less smug and flowery than any of the Anthropic models.

For what it's worth, here's a take on its speed vs cost vs intelligence: https://artificialanalysis.ai/models/deepseek-v4-1-flash

I can go all day and night with this thing with multiple sessions going, and I spend like $2 per day retail. With opencode-go, that fits within the $10/mo subscription, so that's what it ends up costing in real life.

I've been trying Grok 4.6 (High) alongside Claude Opus 5 (High). Grok is capable, but disappointing compared to Opus. My experience with Grok has also made me a little skeptical of benchmarks, because the two models benchmark similarly but it (subjectively) feels like there is quite a big capability gap between the two.