Without tricks google aistudio definitely has limits, though pretty high ones. gemini.google.com on the other hand has less than a handful of free 2.5 pro messages for free
OpenAI offering 2.5M free tokens daily small models and 250k for big ones (tier 1-2) is so useful for random projects, I use them to learn japanese for example (by having a program that list informations about what the characters are just saying: vocabulary, grammar points, nuances).
My experience lines up with the article. The agentic stuff only works with the biggest models. (Well, "works"... OpenAI Codex took 200 requests with o4-mini to change like 3 lines of code...)
For simple changes I actually found smaller models better because they're so much faster. So I shifted my focus from "best model" to "stupidest I can get away with".
I've been pushing that idea even further. If you give up on agentic, you can go surgical. At that point even 100x smaller models can handle it. Just tell it what to do and let it give you the diff.
Also I found the "fumble around my filesystem" approach stupid for my scale, where I can mostly fit the whole codebase into the context. So I just dump src/ into the prompt. (Other people's projects are a lot more boilerplatey so I'm testing ultra cheap models like gpt-oss-20b for code search. For that, I think you can go even cheaper...)
To the OP: I highly recommend you look into Continue.dev and ollama/lmstudio and running models on your own. Some of them are really good at autocomplete-style suggestions while others (like gpt-oss) can reason and use tools.
If you're looking for free API access, Google offers access to Gemini for free, including for gemini-2.5-pro with thinking turned on. The limit is... quite high, as I'm running some benchmarking and haven't hit the limit yet.
Open weight models like DeepSeek R1 and GPT-OSS are also made available with free API access from various inference providers and hardware manufacturers.
I am the person that wrote that. Sorry about the font. This is a bit outdated, AI stuff goes at high speed. More models so I will try to update that.
Every month so many new models come out. My new fav is GLM-4.5... Kimi K2 is also good, and Qwen3-Coder 480b, or 2507 instruct.. very good as well. All of those work really well in any agentic environment/in agent tools.
I made a context helper app ( https://wuu73.org/aicp ) which is linked to from there which helps jump back and forth from all the different AI chat tabs i have open (which is almost always totally free, and I get the best output from those) to my IDE. The app tries to remove all friction, and annoyances, when you are working with the native web chat interfaces for all the AIs. Its free and has been getting great feedback, criticism welcome.
It helps the going from IDE <----> web chat tabs. Made it for myself to save time and I prefer the UI (PySide6 UI so much lighter than a webview)
Its got Preset buttons to add text that you find yourself typing very often, per-project state saves of window size of app and which files were used for context. So next time, it opens at same state.
Auto scans for code files, guesses likely ones needed, prompt box that can put the text above and below the code context (seems to help make the output better). One of my buttons is set to: "Write a prompt for Cline, the AI coding agent, enclose the whole prompt in a single code tag for easy copy and pasting. Break the tasks into some smaller tasks with enough detail and explanations to guide Cline. Use search and replace blocks with plain language to help it find where to edit"
What i do for problem solving, figuring out bugs: I'm usually in VS Code and i type aicp in terminal to open the app. Fine tune any files already checked, type what i am trying to do or what problem i have to fix, click Cline button, click Generate Context!. Paste into GLM-4.5, sometimes o3 or o4-mini, GPT-5, Gemini 2.5 Pro.. if its a super hard thing i'll try 2 or 3 models. I'll look and see which one makes the most sense and just copy and paste into Cline in VS Code - set to GPT 4.1 which is unlimited/free.. 4.1 isn't super crazy smart or anything but it follows orders... it will do whatever you ask, reliably. AND, it will correct minor mistakes from the bigger model's output. The bigger smarter models can figure out the details, and they'll write a prompt that is a task list with how-to's and why's perfect for 4.1 to go and do in agent mode....
You can code for free this way unlimited, and its the smartest the models will be. Anytime you throw some tools or MCPs at a model it dumbs them down.... AND you waste money on all the API costs having to use Claude 4 for everything
If i am not wrong, aider uses repomap to select the context and pack it efficiently for LLM? this coupled with auto copy paste to and from chat web UI of AI provider?
I jump between Claude Sonnet 4 on GitHub Copilot Pro and now GPT-5 on ChatGPT. That seems to get me pretty far. I have gpt-oss:20b installed with ollama, but haven't found a need to use it yet, and it seems like it just takes too long on an M1 Max MacBook Pro 64GB.
Claude Sonnet 4 is pretty exceptional. GPT-4.1 asks me too frequently if it wants to move forward. Yes! Of course! Just do it! I'll reject your changes or do something else later. The former gets a whole task done.
I wonder if anyone is getting better results, or comparable for cheaper or free. GitHub Copilot in Visual Studio Code is so good, I think it'd be pretty hard to beat, but I haven't tried other integrated editors.
What are you using GPT 4.1 with, anything? I find that it will follow your instructions REALLY well, so maybe whatever you are using it in, there is a problem with the prompt. Something might be telling it to do that. It really does do whatever you tell it (but you might not be able to see the full prompt if you are using GPT 4.1 with some kind of other tool/not raw).
Using it in Cline (free VS Code extension) it does exactly what it is told to do, I will have a bigger/smarter model figure out plans for it. It will edit files and just do stuff but its not really the model to think or plan etc.
> 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.
I completely agree with this!
While I understand that it looks a little awkward to copy and paste your code out of your IDE and into a web chat interface, I generally get better results that way than with GitHub copilot or cursor.
I think there’s huge potential for a fully local “Cursor-like” stack — no cloud, no API keys, just everything running on your machine.
The setup could be:
• Cursor CLI for agentic/dev stuff (example:https://x.com/cursor_ai/status/1953559384531050724)
• A local memory layer compatible with the CLI — something like LEANN (97% smaller index, zero cloud cost, full privacy, https://github.com/yichuan-w/LEANN) or Milvus (though Milvus often ends up cloud/token-based)
• Your inference engine, e.g. Ollama, which is great for running OSS GPT models locally
With this, you’d have an offline, private, and blazing-fast personal dev+AI environment. LEANN in particular is built exactly for this kind of setup — tiny footprint, semantic search over your entire local world, and Claude Code/ Cursor –compatible out of the box, the ollama for generation. I guess this solution is not only free but also does not need any API.
But I do agree that this need some effort to set up, but maybe someone can make these easy and fully open-source
I bet it's crazy to some people that others okay with giving up so much of their data for free tiers. Like yeah it's better to selfhost but it takes so much resources to run good enough LLM at home that I'd rather give up my code for some free usage, anyway that code eventually will end up open source
Wow, there's a lot here that I didn't know about. Just never drilled that far into the options presented. For a change, I'm happy that I read the article rather than only the comments on HN. ;)
And lots of helpful comments here on HN as well. Good job everyone involved. ;)
As the post says, the problem with coding agents is they send a lot of their own data + almost your entire code base for each request: that's what makes them expensive. But when used in a chat the costs are so low as to be insignificant.
I only use OpenRouter which gives access to almost all models.
Sonnet was my favorite until I tried Gemini 2.5 Pro, which is almost always better. It can be quite slow though. So for basic questions / syntax reminders I just use Gemini Flash: super fast, and good for simple tasks.
65 comments
[ 2.9 ms ] story [ 77.5 ms ] threadFor simple changes I actually found smaller models better because they're so much faster. So I shifted my focus from "best model" to "stupidest I can get away with".
I've been pushing that idea even further. If you give up on agentic, you can go surgical. At that point even 100x smaller models can handle it. Just tell it what to do and let it give you the diff.
Also I found the "fumble around my filesystem" approach stupid for my scale, where I can mostly fit the whole codebase into the context. So I just dump src/ into the prompt. (Other people's projects are a lot more boilerplatey so I'm testing ultra cheap models like gpt-oss-20b for code search. For that, I think you can go even cheaper...)
Patent pending.
It's my goto copilot.
Open weight models like DeepSeek R1 and GPT-OSS are also made available with free API access from various inference providers and hardware manufacturers.
Every month so many new models come out. My new fav is GLM-4.5... Kimi K2 is also good, and Qwen3-Coder 480b, or 2507 instruct.. very good as well. All of those work really well in any agentic environment/in agent tools.
I made a context helper app ( https://wuu73.org/aicp ) which is linked to from there which helps jump back and forth from all the different AI chat tabs i have open (which is almost always totally free, and I get the best output from those) to my IDE. The app tries to remove all friction, and annoyances, when you are working with the native web chat interfaces for all the AIs. Its free and has been getting great feedback, criticism welcome.
It helps the going from IDE <----> web chat tabs. Made it for myself to save time and I prefer the UI (PySide6 UI so much lighter than a webview)
Its got Preset buttons to add text that you find yourself typing very often, per-project state saves of window size of app and which files were used for context. So next time, it opens at same state.
Auto scans for code files, guesses likely ones needed, prompt box that can put the text above and below the code context (seems to help make the output better). One of my buttons is set to: "Write a prompt for Cline, the AI coding agent, enclose the whole prompt in a single code tag for easy copy and pasting. Break the tasks into some smaller tasks with enough detail and explanations to guide Cline. Use search and replace blocks with plain language to help it find where to edit"
What i do for problem solving, figuring out bugs: I'm usually in VS Code and i type aicp in terminal to open the app. Fine tune any files already checked, type what i am trying to do or what problem i have to fix, click Cline button, click Generate Context!. Paste into GLM-4.5, sometimes o3 or o4-mini, GPT-5, Gemini 2.5 Pro.. if its a super hard thing i'll try 2 or 3 models. I'll look and see which one makes the most sense and just copy and paste into Cline in VS Code - set to GPT 4.1 which is unlimited/free.. 4.1 isn't super crazy smart or anything but it follows orders... it will do whatever you ask, reliably. AND, it will correct minor mistakes from the bigger model's output. The bigger smarter models can figure out the details, and they'll write a prompt that is a task list with how-to's and why's perfect for 4.1 to go and do in agent mode....
You can code for free this way unlimited, and its the smartest the models will be. Anytime you throw some tools or MCPs at a model it dumbs them down.... AND you waste money on all the API costs having to use Claude 4 for everything
If i am not wrong, aider uses repomap to select the context and pack it efficiently for LLM? this coupled with auto copy paste to and from chat web UI of AI provider?
Claude Sonnet 4 is pretty exceptional. GPT-4.1 asks me too frequently if it wants to move forward. Yes! Of course! Just do it! I'll reject your changes or do something else later. The former gets a whole task done.
I wonder if anyone is getting better results, or comparable for cheaper or free. GitHub Copilot in Visual Studio Code is so good, I think it'd be pretty hard to beat, but I haven't tried other integrated editors.
Using it in Cline (free VS Code extension) it does exactly what it is told to do, I will have a bigger/smarter model figure out plans for it. It will edit files and just do stuff but its not really the model to think or plan etc.
I completely agree with this!
While I understand that it looks a little awkward to copy and paste your code out of your IDE and into a web chat interface, I generally get better results that way than with GitHub copilot or cursor.
While the ai has less context, you have more context using the limited chat window. You know what you need from the ai.
The setup could be: • Cursor CLI for agentic/dev stuff (example:https://x.com/cursor_ai/status/1953559384531050724) • A local memory layer compatible with the CLI — something like LEANN (97% smaller index, zero cloud cost, full privacy, https://github.com/yichuan-w/LEANN) or Milvus (though Milvus often ends up cloud/token-based) • Your inference engine, e.g. Ollama, which is great for running OSS GPT models locally
With this, you’d have an offline, private, and blazing-fast personal dev+AI environment. LEANN in particular is built exactly for this kind of setup — tiny footprint, semantic search over your entire local world, and Claude Code/ Cursor –compatible out of the box, the ollama for generation. I guess this solution is not only free but also does not need any API.
But I do agree that this need some effort to set up, but maybe someone can make these easy and fully open-source
And lots of helpful comments here on HN as well. Good job everyone involved. ;)
I only use OpenRouter which gives access to almost all models.
Sonnet was my favorite until I tried Gemini 2.5 Pro, which is almost always better. It can be quite slow though. So for basic questions / syntax reminders I just use Gemini Flash: super fast, and good for simple tasks.