GPT-5.6 Luna Max Is the New King of Grunt Work

I just tried using the post-price-cut GPT-5.6 Luna MAX for grunt work, and it’s fucking good.

I handed it two jobs. One: use the browser to go to every bank and payment platform (after I logged in for it) and pull down the statements I was missing — deliberately no API, no MCP, no CLI. Two: sync the harness from the Claude Code side over to Codex (a SKILL already exists for this). I ran /goal on both. The first one finished cleanly in 22 minutes; the other has been running for an hour and is still going, in good order. So far it’s burned 1% of the weekly quota — converted to API pricing by tokens, that’s roughly 3.14 USD.

At this price-performance ratio, it’s not even a question of comparing it to A＼ anymore. I think DeepSeek’s users might all get pulled away. On token efficiency, on agent ability, on computer/browser use, DeepSeek is going to have a hard time going up against GPT-5.6. I’m calling it: GPT-5.6 Luna is the new king of grunt work.

What it’s best at is brute-force browser scraping, going after sites with no API that love to block crawlers. Before, you either had the AI download a pile of packages and scripts and hack away at it, or you went and paid a third-party API for the data. But sometimes I just want to scrape something once. I don’t want to force it and risk getting banned, and I don’t want to pay a third party for the data either. So I fire up Luna Max + goal, let it run all night, wake up down only 5% HP, job done.

For this kind of thing it drives your real browser directly. Tokens, fingerprint, speed, frequency all look human. I scrape every day and haven’t been 403’d yet. Just leave it running overnight. Whatever tier of GPT 5.6 you use, its browser control and computer control are far ahead of Claude. On accuracy and efficiency of browser control, Claude really loses to GPT by a wide margin.

LUNA is for grunt work specifically. Take the jobs that traditionally eat tokens: sites with no API that block crawlers, where you have no choice but to scrape through a browser MCP. Or being out of the house and remotely driving your computer with computer use. Luna is the only choice.

A few scenarios where I use LUNA most. One: scraping social media posts to monitor sentiment. Reason: I don’t want to risk going head-to-head with anti-scraping, I only check occasionally, and I don’t want to specially buy third-party API data. Two: having it find award seats on airline sites. Mechanically clicking through dozens of date combinations, glancing at whether a seat exists, copying down the availability and flight info for me to make the final call, and there’s no API that does this. Three: having it compare prices across accommodation platforms, even filing best-price-guarantee refund claims. Same reason as above, plus it’s now smart enough to understand best-price-guarantee rules and terms. Four: being out somewhere and telling LUNA from my phone: go open LINE/Messenger, download the file the other person wants, have Sol/Terra process it, then send it back to them on LINE. I’ve already gotten a fair amount done remotely this way. Same goes for filtering job listings on 104, checking hotel prices across a pile of platforms. All my favorite Luna Max scenarios. Feels like I should shoot a video demoing this? (Update: I did, EP.34 https://youtu.be/8Sxw1Mdcl8Q)

Thank you GPT 5.6 Luna Max. Two days ago it snagged me two business class award seats, Bangkok to Tokyo. Only people who’ve used ANA know how bad their award system is and how complicated the rules are.

People who sit around trashing some model “across the board” just have too narrow an imagination, and want to look knowledgeable by putting models down. Every model has use cases it fits. It depends on whether you have the imagination. In my eyes, even Gemini has its place.

I tell everyone GPT 5.6 Sol is a mad dog. Give it the /狗死 (dog-die) command, and I usually play LISA’s “dog-diedog-die” alongside it, and it clamps onto the target and won’t let go until the job’s done.

Lately, besides 5.6 Sol, my favorite is 5.6 Terra high. It’s the top pick for grunt work. Still very accurate at driving the browser and the computer, and it even handles complicated award-seat searches for me. The key thing is how little quota it uses, so I can save Sol for more important things. After using both GPT’s and Claude’s computer control MCP, anyone would be surprised at how big the gap has gotten: one operates smoothly and checks its work carefully, the other yanks your mouse around and gives up easily.

For social sites like X and Threads, direct WebFetch from Claude and Codex often gets blocked and you have to route through a browser, while Gemini can just read them. Reddit is especially unfriendly to Claude; Codex and Gemini go through fine. Models aren’t only competing on how smart they are. It also comes down to what data they can get in the door.

Postscript: A Twenty-Minute Video With Luna Max

Sticking to the principle that every model is good for something, I kept throwing different kinds of work at Luna Max. Eventually I made a twenty-minute long-form video entirely with Luna Max plus HyperFrames.

For work like that, Luna Max is passable if you pair it with a pipeline whose spec is nailed down hard. The catch is that the video spec you wrote will drift in certain scenes and you have to go fix it. Same weakness shows up elsewhere: even with a solid spec, LUNA drifts easily. It really only suits data-collection work, the mechanical browser-driving kind.

After finishing, my read is this: next time, pair it with a higher-tier Terra or Sol model for the final review and it should go smoother.

Postscript: Two Commands Cover Most of It

/goal and /loop. Those two commands solve most problems.

Postscript: The God of Video Post-Production

DaVinci Resolve Studio wired up to MCP and driven by Luna Max is the god of video post-production in this new world. For grunt work like this, Luna Max is genuinely first in the world.