Im canceling my claude subscription - Numerator's Blog

Im canceling my claude subscription

Once upon a time i was out of tokens. again. so i launched opencode to keep working. but this time i picked DeepSeek V4 Flash free, and i wasnt ready for what happened.

deepseek finished the work faster. for my workflow — figure out how something should work, give the bot a list of instructions how the software should work and how to create it — deepseek gave better results, faster. claude sonnet could finish too, but it used more tokens and the result was usually more buggy.

and i can hear you screaming in the comments — why sonnet and not opus or even fable? the reason is simple, token budget. but my point stands: the smaller model, at least from my day to day work, outperforms the bigger one. but only when im providing the instructions, i think. this is just how i work — i have a vision how software should work and i want my agent to execute it.

lets get technical

model<br>parameter count<br>ram usage

Sonnet 5<br>1T-2T (rumors+leaks)<br>640gb-1.2Tb

DeepSeek V4 Flash<br>284B (open source weights)<br>85gb-570gb (depends on quantization)

and the main deal for me is that this model can run locally on consumer hardware.<br>im not saying you should cancel your claude sub. this works especially for my workflow.

Real world tests

so lets rewrite my old software — twitch viewer alert. it works like this: it watches the viewer count, and when the number changes, it fires a notification. theres some settings too, but i dont remember much more. the task for the agents is to convert this into a chrome extension. originally it was just a script you paste into the browser console — dont judge me, i was 19 when i wrote it.

each model was tested once in autonomous mode. the metrics are: token usage, time, and how many times i had to step in to keep things moving.

Prompt

hi i have this older project https://github.com/wallach-game/twitch-viewer-alerter and i wanna convert it into chrome addon, btw do not push to this repo, keep it as local cloned repo

DeepSeek V4 Flash

time: 4 minutes

tokens: 46,796

cost: $0.000004

interventions: 0

notes: added 7 new features and an icon

features added:

viewer count tracking: every 5s → every 3s

alert box: replaced with browser notifications

notifications with buttons (open stream / mute for 30m)

on/off switches in popup

sound alert

cooldown (10s, configurable)

spa navigation support via mutationobserver

offline stream detection

viewer count badge on toolbar

robust selectors with 3 strategies and fallbacks

purple play triangle icon

ui:

notification working:

Sonnet 5

time: 2 minutes

tokens: ~562,910 (mostly cache)

cost: $0.40

interventions: 0

notes: the ui has less stuff than the original extension

verdict: imo it just yanked the old code and removed unnecessary stuff. but it figured out that browser notifications are not possible anymore — good catch.

ui:

notification:

Opus 4.8

time: 1:37

tokens: ~148,500

cost: $0.30

interventions: 0

notes: this version doesnt have a settings ui. kinda odd. and the notification doesnt even work — it created some localstorage nonsense and literally told me at the end it might not work. also, it never figured out that browser notifications dont work without https in chrome extensions.

Verdict

so heres the thing. deepseek did it in 4 minutes, added 11 new features, and cost basically nothing. sonnet did it faster (2 minutes) but the output was worse — stripped features, yanked old code. opus was the fastest at 1:37 but the result literally didnt work.

the numbers speak for themselves. deepseek v4 flash cost $0.000004, sonnet cost $0.40, opus cost $0.30. for my workflow — give clear instructions and let the model execute — the cheap model wins.

but in the end sonnet did what the prompt said — convert into an extension. and it did it. next time ill try a better prompt, something like “copy the original behavior first, then add new features and polish”.

am i saying you should cancel your sub? no. this is what works for me. but if you work like i do — plan first, then let the ai build it — give deepseek v4 flash a shot. you might be surprised.

and also deepseek screwed me over a few times, especially paired with opencodes auto exec command. its idea of fixing a running server app was deleting it without backup. but honestly, this can be blamed on my prompting.

Final words

im aware deepseek pricing is probably not sustainable at this level, but imo the same can be said about claude subscription. but i dont wanna get political here. the fact is that deepseek outperformed claude with a model that can be run locally on consumer hardware. thats the part that matters.