Join the discussion
Write your take first — we'll ask for email only when you're ready to publish.
- Hacker News
- The biggest issue I've seen with people burning through tokens is using very long sessions, especially starting with plan mode and then "iterating" over extended periods. I was burnt badly by extra usage so now I run on $20 Pro. I ruthlessly create new sessions/agents, always ask to create markdown files first (no plan mode) and minimise context aggressively - for example I have a lot of skills that use lazy loading and a small local MCP for lookups plus openrouter with a local model for image detection and fulltext search. Basically I use Claude Code in pi.dev style.by janpeuker
- > Do that well and you can build what a team of twenty engineers would put out in a month for around a thousand dollars.
What does this look like after 6-12 months? Like, how much code are you trying to write total?
Maybe it just doesn’t click in my mind, but sometimes I wonder about how much work people are trying to do and how they actually have enough to get done so quickly in such a short amount of time.
by geophph - They prefer to work harder and not smarter. Forever hill climbing to nowhere.
I've never worked on a complicated codebase that started out that way until the rest of the business concerns and office politics came into effect. People may not like it, but the bureaucracy is far and away more valuable than the core functionality.
Mature codebases are years of people thinking of all the possible gotchas while solving their acute pain points. This is not fluff, but the living and breathing part of it. Without that code, it's just a machine barely doing stuff in the most obtuse ways possible that nobody wants to pay for.
I would argue that they're putting LLMs to work on that finer detail stuff, but AI is still far too dumb. No, what they're doing is playing with their skinner box.
by sublinear - AI coding at home literally costs $100/month. I'm wondering where $400 is coming from? $100 is more than enough for "coding at home", IMO. I rarely face the limits, and when I do it's just a time for a quick walk anyway.by RomanPushkin
- Man I’m using the $20/month sub and it works just fine for me. Granted, I have a family and house and lots of obligations so by the time I hit the limits some other task is due before I can return to coding. If I hit the limits before I have something else to do then I just code by hand or review what has been generated until I can use the agent again. Reviewing agent code is a good way to learn too, agents have shown me different approaches than what I would have done and they’re definitely worth thinking about. Also, fixing their mistakes has helped me write better prompting although being a team lead for half a decade has taught me how to specify what I want very clearly and cc gets it right most of the time haha
About interruptions, one thing AI assisted coding really helps with is coding with constant interruption. I can leave CC for half an hour and return then tell it I had to step away, catch me up, and proceed. This works well for me.
by chasd00 - Can I run something comparable to Opus 4.6 locally yet? I keep hearing conflicting things. If I can spend 10k to do that I would cancel my subscription. The problem is I don’t wanna spend the money to find out myself.by vadansky
- Best you could do is connect two Mac Studio M3 Ultra 512G RAM each with Thunderbolt. Then theoretically you can run frontier Chinese models (but not Deepseek v4 Pro yet). That would be about $20k.
But - good luck finding them. Apple discontinued the model a few months ago. And more recently, even 256G model was discontinued. Big AI really really does not want people to get off their needle.
by atemerev - i spent 8k and get close to a 2-3x slower sonnet. running 2x spark deep seek v4 flashby captaintobs
- Some benchmarks have shown Kimi K2.6 within error-bar distance of Opus 4.6, and you can run it on eight RTX6000s. Right now it's not possible to set up a machine like that from scratch for less than $100K... but right now it's also hard to put a price on autonomy.by CamperBob2
- > Can I run something comparable to Opus 4.6 locally yet?
Sadly, no. The best comparable thing you can get is about Sonnet 3.7
by als0 - 10k will not get you anywhere near opus or sonnet. It's simply not possible for mere mortals currently.by grim_io
- If you want frontier-level, the economically reasonable option is OpenRouter or a direct sub to frontier-of-your-choice.
The reality is that they do not offer configurations that would allow a consumer to run that much VRAM on a single setup to protect datacenter margins. Apple used to, and they stopped, those devices are going for ~$20k+ each on ebay now.
You can get very, very capable models on a 3090/4090/5090/6000 series card. But if you want 'frontier level' you are investing ~22k at a bare minimum if you go new. Used you can probably build your own server for much cheaper up-front cost but it's likely going to be 4-6x+ electricity usage.
by Catloafdev - For me, investing in hardware seems to be the way to go.
I learned coding nearly 24 years ago and still learning new stuff all the time. At no point in time I had to rely on a subscription model to learn and do new stuff.
If LLM and agents are the default tools for coding and building software, at least for next few years, it seems like a no-brainer to invest $2000-3000 on hardware, like a Halo Strix PC.
by esalman - 3k? Try 10
- $3k isn't getting you frontier model capability. It's barely getting you any capability if that's split into buying an entire PC rather than just GPUs.by hgoel
- Yes and no. Hardware does lock you in. Granted, I am happy with my 128gb of shared memory, but I am mildly concerned that it actually is more expensive now than when I bought mine. It does not bode well for the future; not when combined with recent WH admin moves on Anthropic and the reality that next batch of good models may require more than 128gb to run well.
edit: I am not dismissing local. I am one such user ( though I have subs too ), but one has to be clear eyed about the trade-offs.
- With you here. I'm using my cheapo 16gig vram card I picked up a year or so ago, and I'm like -- yes, I percieve that you can pay for way more tokens per second that I can do at home.
But that feels like measuring productivity in lines of code. For what I'm doing, I'm not seeing the benefit in any subscription.
Sure, I can't one-prompt a whole new boring CRUD app, but oh well.
by jrm4 - I wondered if there might be a no brainer "free" option on discarded hardware.
I have a GTX1080ti which i think is circa 2018, it's unused, more than paid for itself over the years, owes me nothing at this point so the hardware is free.
It runs Gemma e4b multimodal, qwen 3.5 8b or the qwen 4b embeddings models well enough (40+ t/s for the LLMs).
The machine consumes 350 watts at the wall when under load (3 watts when sleeping, 80w at idle). Electricity costs me £0.035GBP/kwh which is cheap for the UK (load shifting via house battery).
144k output tokens for around 1pence (and takes an hour to do that in theory).
It's only JUST cheaper to use than the far more capable deepseek v4 flash model despite the free hardware and ~10x cheaper than normal electricity.
by CraigJPerry - I invested about $4,000 in an NVIDIA DGX Spark several months ago. 128 GB of unified RAM, and the NVIDIA GB10 chip. With the RAM, the several CPU cores, and the 4 TB NVMe SSD, it's a very capable ARM64 Linux computer even without the GPU, and so far I've mostly been using it as such. But I wonder, what's the most capable model, specifically for coding, that can run well on that hardware?by mwcampbell
- Qwen 3.5 122B can fit with context at a pretty high quant (Q6). That's an excellent model.by anon373839
- If you don't know that already and "using it as such" ... your post should start with "I blew off $4k on a toy several months ago".by ozim
- DeepSeek V4 Flash is a very capable coding model that runs well on the hardware you described. Look up the optimized version specifically designed for local use.by znnajdla
- Deepseek v4 flash is shockingly strong for its size and reportedly runs well on that hardware.by morganastra
- https://www.canirun.ai/?status=tight might answer that questionby Yoric
- I'm currently working through research and testing for an article on Ars about the Spark and what things one might do with it, and I've kind of stumbled into a two-LLM agentic setup with Qwen3.6-35B-A3B (via nvidia/Qwen3.6-35B-A3B-NVFP4) as the planning agent and the FP8 version of Qwen3-Coder-30B-A3B-Instruct (Qwen/Qwen3-Coder-30B-A3B-Instruct-FP8) as the coding agent that the planner delegates tasks down to. I'm sticking with vLLM as the inference engine, and I've got it wired together into a 2-agent loop with Opencode.
The Qwen3.6-35B-A3B planner hums along at 50-55 tokens/s, and the Qwen3-Coder-30B-A3B-Instruct coder does 30-35. With both agents up and ready to work, RAM consumption sits at about 112 of 128GB.
It's pretty okay. I'm faffing around with having it disassemble old MS-DOS games from the 1980s, which is a task that lends itself well to the setup. It's not the fastest thing in the world, but with the planner's context window at 256k tokens and the coding agent at 128k, they chew through pretty long task lists handing things back and forth without complaint. The only real issue is that even with really tightly scoped prompts, the coding agent tends to hallucinate like it's on LSD. But the planning agent appears to be quite good at spotting the hallucinations and re-parceling work back to the coder.
It's neat. I'm going to be sad when I have to return the review unit in a couple of months.
edit - I also have been fiddling with Deepseek v4 Flash via Antirez's setup (https://github.com/antirez/ds4), and it's pretty fantastic (and fantastically easy to get running). It's pretty pokey on the Spark, though, at 14-ish tokens/sec. And unless you have a second Spark, it's going to be the only model you run at one time, as it eats alllll the rams.
by lee_ars - What are people doing at home? I have like 5 different apps I code on the $20/month Claude plan and like sure I can hit rate limits but - What are people doing to burn through $3k in tokens?by mikgp
- > What are people doing to burn through $3k in tokens?
The short answer is: they are doing slop. Most of the coding can be done quickly with a keyboard, intelisense and maybe some code generation templates.
But people became dependent on AI doing everything for them and tech bros now started to squeeze. Like a drug dealers.
by Random09 - Same, but I suspect that I don't have any issues with hitting caps because I actually still do plenty of the thinking myself, and just use the AI to help accelerate some of the boring stuff I don't want to do myself. This has been especially nice for my personal projects at home. It has made me much more likely to actually want to work on my side projects when I don't have to deal with some of the tedium after working on my company's tedium all day.
I suspect that most of these people who are burning through thousands of dollars worth of tokens at home are largely producing big ol' piles of slop.
by hackeradam17 - Same for me. $20/mo is just fine and I use it to code daily.
I suspect the people that burn through tokens have several subagents and 50 skills loaded and 40 MCP tools. All those load up the context on every single turn.
by binarymax - YMMV but automations eat through the $100-$200 plans, which burn thousands in tokens alone.
I have hourly automations for root cause analysis on customer support issues, daily automations for eg log analysis, weekly & monthly automations for KPI tracking & actioning.
I will say, when I was building side projects that were 1) fairly well defined in scope and 2) without users/need for automations it was much easier to stay under $20/mo plan limits. Now I regularly hit weekly limits and need multiple Max plans