I opened my feed this week and nearly spilled my coffee. Four brand new state-of-the-art models dropped in a matter of days, and my brain was fully melting trying to keep up. Luckily, the creator behind Future Tools did the drinking-from-the-firehose part for us and boiled it all down, so I get to pass the good stuff straight to you.
Here’s the wild part. This industry pro didn’t just read the benchmarks. He actually tested each model and found the scoreboards don’t match real life. I love when someone does the hands-on work instead of trusting the leaderboard.
The four new drops
- 🧠 Claude Fable 5.1: crushed the benchmarks on day one, but the author found it’s the priciest per task and took 18 minutes to make one SVG image.
- ⚡ Gemini 3.8 Flash: cheap, fast, and shockingly good at coding. He clocked it around 58 cents per task versus Fable’s $3+.
- 🎲 Muse Spark 1.3 (Meta): tops the coding charts and it’s free right now, yet the tester’s game clone came out as plain cubes shooting cubes.
- 👑 GPT-6 Astra: still rolling out, but he says its output looked the most polished and complete.
The twist
Here’s what got me. This industry pro ran the same “build a Megabonk clone” test on all four. The model ranked number one for coding (Muse Spark) gave him the ugliest result. The models ranked lower gave him the best-looking games. That gap made him question the benchmarks he’s trusted for years, and honestly, it made me rethink them too.
His quick take on the game test:
- Fable’s version looked great but cost him over $120 and ate his whole daily usage.
- Gemini’s version looked way better than expected and barely used any credits.
- GPT-6’s version was the most aesthetically pleasing, generated in about 12 minutes.
Rapid-fire finds he flagged
The creator packed in a ton more, so here’s the skim:
- ChatGPT now lets you connect more than one Google account.
- Google added voice chat inside Gmail, Docs, and Keep for paid plans.
- World Labs Atlas turns a few photos into a full 3D environment you can move around in.
- Runway Solaris lets you drag, drop, and edit video in real time.
- Nvidia acquired Hugging Face, which the author reads as a big bet on open-weight models.
- Two new transcription models landed from Meta and Microsoft in the same week.
The one that made me laugh
The person who shared this pointed out Dyson built a $500 AI toothbrush called the Camerajet. It has a camera and a water jet that uses AI to floss between your teeth while you brush. He admits he’s tempted, and weirdly, so am I.
A heads-up worth keeping
This contributor also noted your ChatGPT chats aren’t private in court, and New York just barred AI use in K-8 classrooms so kids learn the fundamentals first. Small notes, but good to know.
My honest reaction? The lesson I’m taking from his testing is simple: don’t pick a model off a leaderboard. Run your own real task and judge the actual output. Benchmarks are a starting point, not the verdict.
Want the full side-by-side breakdowns and the game demos? Check out the original video, it’s packed with detail I couldn’t fit here. 🚀