Notes

AI coding prices got stupid this week.

Prices are per million tokens.

Muse Contributor is the cheapest of the bunch. Ten million input tokens and two million output tokens cost $1.40. The trade is permission for Meta to train on your prompts and completions, so I would keep work code far away from it.

DeepSeek V4 Flash is nearly as cheap without needing a special contributor tier, and its cached input price is a frankly stupid $0.0028 per million tokens.

Luna then went from cheap to silly. OpenAI dropped the direct API price to $0.20/$1.20, and OpenRouter is temporarily selling it at another 50% off ($0.10/$0.60).

Subscriptions are getting dragged into the fight too. OpenCode Go is $5 for the first month and $10/month after that, with Luna currently getting 2x usage limits.

Luna is still king for me, though.

By my rough usage-limit math, Luna Max on the $20 ChatGPT Plus plan works out to around 1.5 billion tokens a week. The exact number depends on how credits are counted, but the allowance is large enough that I have stopped thinking about individual requests. Max reasoning when I need it, fast mode almost everywhere else.

This is basically TPS over Quality again, except the cheap fast model is now also very good.

ChatGPT Plus. Luna Max. Fast mode. That is all I need.

Begun, the price wars have.

Having used Claude Code extensively at work and Codex + Antigravity for personal projects, I can confidently say the Codex $20/month Plus plan is the best value AI coding plan out there right now.

The main reason: frequent session resets, banked reset credits that carry forward, and the fact that it runs gpt-5.5 as the top model. I've burned through close to $600 API equivalent in the last ~20 days across two accounts — and paid $40 total.

The pricing isn't an accident. Plans like this are the gateway drug — cheap enough to get you hooked, generous enough to make you dependent, and carefully designed so you can't imagine going back to writing code without an agent sitting next to you. And it's working.

Enjoying it while it lasts — these things have a way of getting quietly squeezed once the dependency is locked in.

My local LLM post blew up a little this week.

Not "quit your job and become a full-time blogger" blew up. More like "wait, people are actually reading this?" blew up. The post is at around 14k views now, which is probably a rounding error for people who do this professionally, but feels pretty good for a hobby site I keep poking at after work.

The site also crossed 5k users. Google Analytics says 6.6k active users over the last 7 days, 9.2k events, 20k views, and 6.6k new users. I know analytics numbers are always a bit cursed, but still. This is more attention than this little corner of the internet usually gets.

The funny thing is that the post that did it was exactly the kind of thing I like writing: too much detail about local inference, a bunch of practical notes, and enough benchmarking to make the post useful to the version of me that would have been Googling this at 1am.

So yeah. Small milestone. Big enough for a note.

PS: Sorry for using Google Analytics lol. I will move this to Simple Analytics soon.

Lost my Copilot quota today. Stopped using Claude when they started testing on "prosumers". Finished Codex Quota for the site redesign. Then finished Antigravity and Gemini CLI Pro models until I had nothing left.

I had to rely on Gemini 3 Flash. Turns out, it has no right to be as good as it is for repeated tasks (skills) or refactoring. Insanely good model, but the best thing is the speed. On top of that, I cant seem to run out of quota either.

Unlike what you'd think, when you factor for quality (graphics settings in games, basically) vs TPS (FPS, effectively), it's actually more fun to work with a fast model. It responds insanely quickly and the iteration loop becomes super fast.

I think I might just stick with this for a while longer.

If you're reading a technical blog, there's a high chance you use some sort of personal knowledge management system-like Obsidian, Notion, or Logseq. And if you're anything like me, you frequently copy interesting snippets, setup guides, or entire articles into your own vault for future reference.

The problem? Selecting text and hitting Ctrl+C often grabs the rendered HTML. When you paste it into your Markdown-native PKM, the formatting breaks: code blocks lose their syntax highlighting, links get mangled, and lists become a mess of weird indentation.

To solve this friction, I've added a "Copy as Markdown" and "View Raw" button to all technical content on this site (posts, snippets, TILs).

This serves two crucial purposes:

1. Frictionless PKM: Saving tutorials (like the Unsloth example above) directly into your knowledge base without re-formatting code blocks and mangled links.
2. LLM Context: Quickly copying an entire guide or snippet into ChatGPT, Claude, or local LLMs to ask questions, debug setups, or request modifications.

It uses a small Eleventy template to generate a .txt file containing the raw, unparsed .md content. A simple Alpine.js component fetches that file and copies the pure Markdown directly into your clipboard. No more re-formatting text just to save a useful tutorial.

Every technical blog that expects to be referenced by its readers should adopt this pattern. It respects the reader's time and acknowledges how developers actually consume and store knowledge today.

Every piece of content on this site now carries an authorship badge - a small indicator of how it was made:

• Human - Written entirely by me, no AI assistance
• AI Assisted - Written with AI as a creative partner; ideas and direction are mine
• AI Generated - Primarily generated by AI with my oversight and editing
• Unclassified - Not yet tagged (working through the backlog)

I added this because transparency matters, especially as AI writing becomes indistinguishable from human writing. Most of my recent posts are AI-assisted - that's just how I work now. I'd rather be upfront about it than pretend otherwise.

This note itself is AI-generated. Fitting.

Zed's AI settings - a simple toggle to disable all AI features. Every app should have this. (Zed Blog)

Honestly, everything on the web needs a toggle like this - a reader-controlled switch to filter by authorship. Your feed, your rules.

On the flip side, I'm also building aidar - a tool that scans websites and measures stylistic signals associated with AI-era writing. The badges here are self-reported; aidar is the attempt to detect the same thing from the outside.

Its interesting how i've replaced all my social media activity, bit by bit with my own site. Be it microposts (stupid name for tweets), photos, or even small notes, its all here. i think this speaks to the state of social media today moreso than my desire to vibe code.

I know no one reads it, but as some wise person once said - "If a tree falls in a forest and no one is around to hear it, does it make a sound?" - well, if I write something on my own site, and no one reads it, does it matter? To me, yes. It does.

Just implemented so many small changes to the site. Using Copilot and Codex coding "agents" to just crank through a TODO list. Maybe I'll write about it later.