Mario Behrendt | AI Founder

@Shpigford Claude Code, not Conductor? 🤨

what building a 9-figure company looks like in 2026. (or a total flop. who knows. certainly not me.)

I've been using FluidVoice over the last 2 weeks after I got too fed up with WisprFlow, its inaccuracy, and the complete lack of development. The settings panel could be a tiny bit more polished but other than that it's exceptional. Much faster, local, and much more stable!

GPT 5.6-Sol definitely uses a lot of tokens. Most days running 5-6 agents in parallel I average ~600m tokens. With Sol I average ~1.5b for the same work. 🤨

This. Anything that requires multiple steps to configure/use, and anything I don’t want to deal with even if small, goes straight to Codex.

@LLMJunky This happened to me despite being on the latest version because I also use Conductor and it bundles its own Codex binary. So MacOS wasn’t referring to the Codex I was using directly. Maybe something similar happened for you?

Noted lol A badge of honor And its wrong. I was on codex version 0.144 I was not on the old cert

Apple's war on OpenAI has begun lol

@Kappaemme1926 Would be nice to have this on skills.sh rather than a custom installer

CODEX SKILL THAT FINDS YOUR STARTUP’S FIRST CUSTOMERS! I made a Codex skill that analyzes your startup and finds potential customers from real public signals. Paste your startup URL while Codex defines your ideal customer, searches public discussions, qualifies each prospect, and generates a polished report with personalized outreach openers. -> ideal customer profile analysis -> public pain + buying signal research -> evidence-backed prospect shortlist -> fit, timing + reachability scores -> original source links for every prospect -> personalized outreach openers -> polished HTML report -> one-command install Install: npx --yes codex-first-customer-finder-skill 100% open source. Repo in Bio.

@Kappaemme1926 Sol Medium. Terra and Luna Max look good on paper but they hang themselves before they get to a conclusion. Sol Medium gets there faster. High is unnecessary for most people’s work and quite expensive.

With the new GPT-5.6 models, what reasoning setting are you actually using? I’m trying to find the sweet spot: -great output -low token usage -fast enough for daily work What’s been working best for you so far?

@Shpigford I’ve never done Safari extensions, only Chrome and iOS apps. Is there no way to update OTA once approved with Safari extensions?

4 days later...still waiting on Apple to review the knockoff.shopping browser extension. and even once approved, it'll be multiple versions behind the current version. 🙄

@adonis_singh I have a plus and a pro plan. Took me 2 minutes to get through the 5h limit on the plus plan. 15 minutes on pro. On Sol medium. So not quite 20x I guess 🫠

gpt-5.6-sol ultra makes 20x plan look like $20 plan

@onusoz That’s faster than the majority of home broadband in London lol

The wifi I have in the plane with starlink is almost better than the wifi I had at home in Germany

@ChristianSelig I’m on an M4 Pro. 8-10 codex agents without compiling anything is enough to make it literally hot to the touch. I can’t keep it on my lap

Does anyone find their laptop gets super hot/loud using Codex/Claude and XcodeBuildMCP or is my M1 Mac finally starting to show its age

Really impressed with Grok 4.5! The benchmarks aren’t quite accurate. It’s not Opus level (yet). But it’s not far off and it’s MUCH faster. And obviously more efficient/cheaper. If they get their next (bigger) training run right, Anthropic might have gone from #1 in coding to #3 since October last year. Beaten by two models that are significantly faster and more efficient too.

@SteveOnSpeed Regulation

Name one thing that Europe does better than the United States. I'll go first: Public transportation.

@nicdunz Based on what? Soo Medium eats usage at least at 10x the rate of 5.5 High

5.5 Low → Luna High: smarter, cheaper. 5.5 Medium → Luna Max: smarter, cheaper. 5.5 High → Sol Medium: smarter, cheaper. 5.5 XHigh → Sol High: smarter, cheaper.

@jonkomet How though? I’m not seeing that whatsoever. x.com/behrendtio/sta…

ok now i can say, that GPT 5.6 has resolved this token problem with efficiency and i am no longer experiencing bad usage limit issues🔥

On reasoning level medium that is!

Right. GPT-Sol is great but it takes me 15 minutes to eat through my 5h window on a Pro sub. That’s as uneconomical as Fable.

I wanted to compare Fable and GPT Sol a bit more, after months of only GPT/Codex (and some Conductor). Today I've been working on various things across 5.6 Sol and Fable. Mostly building but also some marketing, video editing and other stuff. Fable: It is a good model but I'm still not seeing the big leap the hype tells us to believe. Sol: Definitely better than 5.5. But it eats token like crazy. I was at 500 million tokens in about a quarter of the time it took me to get there with 5.5. With the same results... 🫠 Claude Code Desktop: Continues to be the most disappointing piece of hype software. It's just so badly designed. Clearly not for engineers/builders. Or for folks who like a nicely designed UI. Codex: Keeps getting better. It's so good! Love the new features in the recent release. But PR-status updates are still broken. 😭 So I'm back to @conductor_build for now. Though it won't let me set Sol as the default model @charlieholtz

Managers are the biggest lever for happiness at work "Tech workers with an extremely effective manager report much higher job enjoyment and dramatically lower burnout than those with an ineffective one."

@virattt I just started my third company at 38 after just over 10 years as an employee. Very impressive what you’re up to mate! Good luck! 🤘

I'm 37 years old and doing S26. • solo founder • married with 1 kid • grinding side projects for 4 years My edge is consistency + improving 1% daily. If I can do it, so can you ❤️