Day 228: A 12-Minute Minecraft, a 3-Cent Model and $233K in ARR

Day 228 opened with two numbers on the board: a one-million-dollar goal and $231,260 in ARR. The streamer hides none of it and puts the figures in every stream title. I think that is the whole appeal of the marathon format: progress is measured in front of everyone.

The first news of the day had landed that morning with the release of DeepSeek-V4.1-Flash. Reuters confirmed the launch in a September 10 report, describing the model as the smallest member of a new architecture family. The streamer set a timer and announced an hour of live testing.

The show opened on a benchmax table putting DeepSeek-V4.1-Flash at 74.2 points, ahead of Opus 5 and GPT-5.6 Sol. Chat joked that benchmaxing itself should become a scored discipline. I take the joke seriously, because the party showing the table swims in the same ecosystem as the party selling the model.

The real test ran on Bridgebench. The opening run finished in 3 minutes 28 seconds for three cents, while the same job took GPT-6 Astra 3 minutes 54 seconds at roughly twenty times the price. The streamer praised reflections and design quality, and chat agreed.

I cross-checked the pricing story with web sources. The stream quoted 15 and 60 cents per million tokens, while the tech press reports the model racing rivals at roughly 86 times lower cost with sharply reduced memory needs. The Flash label is no accident: this segment is where speed meets price.

The test setup came from the community. A Minecraft prompt from a viewer named Isaac was picked, and the model was told to open a subdirectory and build the clone inside it. Commands went out live with waiting times uncut. That is where I find this format honest: errors and wins both stay on the record.

The result surprised everyone: the Minecraft clone landed in about twelve minutes for fifty cents of total spend. The streamer called it the best output seen from a flash-tier model. The Muse Spark 1.3 and Gemini 3.8 Flash attempts were judged weak, though the Opus 5 and GPT-5.6 Sol level stayed out of reach.

The comparison did not stop there. An older Minecraft build generated with Muse 3 was launched in Chrome for a side-by-side look. The fact that the old build still runs raises the consistency question behind one-shot wins.

The harshest lines of the day targeted the harness, not the model. The streamer said plainly he dislikes the ready-made DeepSeek test setup. Then came complaints about inconsistency and slowness, capped with a refusal to let the model anywhere near his real codebase. That was the most honest moment of the day, in my view.

A second story runs on the Bridgebench side: version three of the platform is nearly ready. With the stream passing 44 thousand views in under two hours, the streamer argued the interest would convert into revenue once the new release ships. The testing tool and the content feed each other.

More trials filled the gaps: Sonnet got a rocket-themed prompt and yesterday's test ran again. Chat debated code quality, with the popular view that visual tests signal quality. I stay cautious: a pretty screen is not clean architecture.

The bookkeeping segment was remarkably open. The streamer pays about $1,300 a month in AI subscriptions. A $15,000 API quota melts fast, the AWS bill hits $1,000 a month, and daily token burn runs between one and two billion. That is the invisible cost of the marathon.

On the product front there was an iOS build. The team opened the Apple account for the BridgeMind iPhone app and dug into a failed deploy on AWS Amplify. Logging into accounts and debugging live has become the signature move of vibe-coding culture.

The most talked-about moment was the counter: ARR crossed $232K during the broadcast and pressed toward $233K. Second-quarter figures were disclosed as $70K revenue against $37K net income, a margin above fifty percent. YouTube income is excluded from the counter, as stressed on air.

The target is equally explicit: $300K before September ends. The streamer claimed the supporter base on X can deliver 250 thousand views on a single post. The community functions as a marketing channel here.

The voice-assistant segment showed where the product is heading. A wake word for the iOS side and hands-free prompting for the desktop assistant were discussed. The current voice stack was admitted to be pricey, with user credits draining fast.

Hopes ride on GPT Live and realtime voice models. The streamer did not hide finding current realtime models weak, yet argued the new release could justify rewriting the assistant. One independent study measuring four thousand sessions found per-minute costs diverging widely from folk estimates, so pricing stays foggy.

The rituals are part of the format too: a combined YouTube-plus-Twitch chat app, grind music in the background, gifted ChatGPT Pro subscriptions, and a sponsor read in between. These read less as filler and more as mechanics holding the community together.

I verified the model identities against the web. The GPT-5.6 family spans Luna, Terra and Sol; Opus 5 launched claiming coding parity at half a rival's price; GPT-6 Astra ships as the OpenAI flagship and has entered Copilot. The names in the stream are real.

The test rivals check out as well: Muse Spark 1.3 is Meta's current announcement and Gemini 3.8 Flash is Google's. So the duel ran against that week's actual news cycle, which raises the historical value of the broadcast.

My map of the better-than claims looks like this: DeepSeek leads the benchmax table, dominates the price-performance round in visual Bridgebench trials, while Opus 5 and Sol hold the heavyweight line on general capability. Three different yardsticks, three different winners; no single table can tell it.

The show closed on GPT Live with 5.6 Luna. The streamer rated the day's progress high and left the rest for tomorrow. The sign-off was quick, the counter resting near $233K.

My overall read: this 228-day marathon frames both shores of the AI economy in one shot. Inference keeps getting cheaper while orchestration keeps getting dearer; the model arrives nearly free and the labor turning it into product costs a fortune.