How I code with AI on a budget/free
Last updated: July 2025
How I Code with AI on a budget/free
My Browser Setup: The Free AI Buffet
First things first, I have a browser open loaded with tabs pointing to
the free tiers of powerful AI models. Why stick to one when you can get
multiple perspectives for free? My typical lineup includes:
-
At least 2-3 tabs of
z.ai’s GLM 4.5
– free on web, and seems as good or better than Claude 4! no joke. -
1 or 2 tabs of
Kimi K2.. another Claude or Opus-like model. Free to use on website. -
Qwen3 Coder and other new ones at
chat.qwen.ai. -
2 tabs of Kimi K2 at
kimi.com. This one was fixing hard bugs multiple times a day before GLM
seemed even better at it possibly. -
At least one tab of
OpenAI Playground. If you set your account’s data settings to allow OpenAI to use your
data for model training, you get free tokens to use on GPT-4.5, o3,
and other models. -
At least one tab but more like three of Google
Gemini AI Studio
(Gemini 2.5 Pro/Flash are often free and unlimited here).
Also, try
Google Gemini 2.5 Pro
(different than AI Studio, has better image generation and deep
research; I always have a couple tabs of this along with a couple tabs
of AI Studio). -
Couple tabs of
Poe.com
usually set to Claude 4 or o4-mini for its free daily credits on
premium models. -
Several tabs of
OpenRouter, set to several models, some free models, some not. -
At least one tab for
ChatGPT
(the free version is still useful). -
At least one tab for
Perplexity AI, especially good for research-heavy questions. -
At least one tab for
Deepseek
(v3 and r1 are free on their web interface, though watch the context
limit). -
One tab for
Grok.com. Good, free and seemingly unlimited for general use and deep
research/image editing. I mainly use the deep research feature,
similar to perplexity. -
Phind
is another free one, it tries to show you flowcharts/diagram visuals. -
lmarena.ai
offers free access to Claude Opus 4 and Sonnet 4 and others. Free Opus
4 is so good.
Claude.ai
– Free but sometimes so limited it’s annoying to use, so I use other
sites/ways to access Claude like Cody extension, Copilot, etc.
Grok offers free compute and uncensored image generation, which can
be useful when other models’ safety systems interfere with
legitimate tasks. However, it is run by someone who may be
manipulating the company to promote Nazi-adjacent views and
misinformation, possibly attempting to sway users in that direction.
Reports indicate that Grok has been instructed to lie about
historical atrocities, including topics such as genocide in Africa.
While the misinformation appears to be mostly on X, if you keep in
mind to restrict usage to coding or be cautious knowing it might be
programmed with questionable motives, it can occasionally be useful.
A smarter, cheaper workflow: Focused Context
When you use AI in web chat’s (the chat interfaces like AI Studio,
ChatGPT, Openrouter, instead of thru an IDE or agent framework) are
almost always better at solving problems, and coming up with solutions
compared to the agents like Cline, Trae, Copilot.. Not always, but
usually.
When you use things like Cursor, Cline, Roo Code for everything, they
are sending tons of text to the AI about how to use their tools, how to
use or activate MCP server stuff, edit files, etc it “dumbs it down” too
much. It gets confused. People end up paying for the most expensive best
models to do everything and even that isn’t enough to get over the
dumbing down effect from the AIs getting tons of unneeded information
unrelated to your problem.
So when that happens, I use my tool to generate the right context to
solve my problem. Then I paste it into one of the many AI web chat’s
(sometimes more than one, since they sometimes give different answers)
and just ask it questions or ask it to code review, to try to figure out
why x is happening when y is happening…etc then when it figures out a
solution.. i have it write a prompt for Cline or another agent type
thing to do the actual file edits. GPT 4.1 can handle this just fine and
I have unlimited. No reason to be wasting Claude credits to edit files.
No reason to be sending Claude a bunch of crap it doesn’t need making it
dumb. I can use Claude to plan out anything or fix really hard problems,
cheap, using Openrouter web chat then just paste it back in Cline and
let it run.
After doing this for a while, you really get a feel for which models
excel at which types of tasks.
How AI Code Prep Helps (Example Prompt Structure):
Can you help me figure out why my program does x instead of y?
Then,
AI Code Prep GUI
(for Windows, Mac, Linux, and web) steps in. It recursively scans your
project folder (subfolders, sub-subfolders, you name it) and grabs the
code, formatting it nicely for AI like this:
The context block generated by AI Code Prep looks like this:
Can you help me figure out why my program does x instead of y?
fileName.js:
... the contents of the file..
nextFile.py:
import example
...etc
Can you help me figure out why my program does x instead of y?
It writes it twice if you have that option enabled, which helps get
the AI to focus better on your question/prompt. You can choose to have
it on top, bottom, or both. OpenAI claims this helps, I haven’t really
tested to see if that’s true but it seems logical.
On Windows, you just right-click somewhere inside your project folder
(or on the folder itself) and select “AI Code Prep GUI” from the
context menu (look at the screenshots on the site). A GUI window pops
up, usually with the right code files pre-selected. It smartly tries
to skip things you likely don’t need, like node_modules,
.git, etc. If its guess isn’t perfect, you can easily
check or uncheck files.
This is super useful when your project is huge and blows past an AI’s
context limit. You can manually curate exactly what the AI needs to
see.
The problem with many coding agents like
Cline, Github Copilot, Cursor, Windsurf, etc., is that they often send
either WAY too much context or WAY too little. This is why they can
seem dumb or ineffective sometimes. Sometimes, you just gotta do
things yourself, use a tool like mine to select the files yourself,
but it helps auto-select the code files while skipping the stuff you
probably don’t need (but still have the option to add what you want
with the checkboxes) and then dump that curated context into several
AIs (especially the free web ones!).
Sure, there are other context-generating tools, but many are
command-line only, or need a public GitHub repo link. What if your
code is private? What if you want to keep it local? What if you prefer
checkboxes on a GUI? For something like this a GUI makes sense.
Note: This page isn’t updated with all the latest AI Code Prep GUI
features, check
wuu73.org/aicp
for latest major upgrade
