Yesterday was one of those days where the leaderboards moved, but the real story wasn’t the scores—it was the kind of intelligence the labs are betting on now.
OpenAI dropped GPT-5.5
and the headline number is its 73.66% on Matharena, more than doubling its predecessor. But the actual shift? This isn’t a model built to talk about math—it’s built to do math, then act on it. KernelBench scores (6.57% for GPU kernel writing) don’t lie: OpenAI is pivoting hard toward agentic execution, not just chat. The benchmarks they’re highlighting aren’t about fluency; they’re about getting things done. That’s not an incremental upgrade. That’s a statement: the next phase isn’t better answers—it’s autonomous workflows.
Meanwhile, Mistral Medium 3.5 quietly dropped
with 77.6% on SWE-Bench Verified, outpacing models three times its size on real-world coding tasks. Open weights, four-GPU hosting, and a unified architecture for reasoning, coding, and instruction-following. No fanfare, just receipts. For the record: this is what efficiency looks like when a lab treats compute as a constraint, not a crutch.
Google’s Vision Banana is the sleeper hit.
A single model handling segmentation, depth estimation, and generation—all while beating specialized systems on Cityscapes (0.699 mIoU) and ReasonSeg (0.793). The benchmark choice here is telling: they’re not just chasing SOTA on one task; they’re asking how many tasks one model can unify without tradeoffs. That’s a different kind of scaling.
And then there’s Mythos.
No public benchmarks, because when your model’s specialty is exploiting zero-days, you don’t put it on a leaderboard. The White House is drafting guidance for federal use. The Pentagon already flagged it as a supply chain risk. This isn’t a model release—it’s a geopolitical event wrapped in a weight drop.
The Record:
GPT-5.5’s Matharena score is the number everyone will cite. Mistral Medium 3.5’s SWE-Bench is the one that should worry the incumbents. Vision Banana’s ReasonSeg is the quiet revolution. And Mythos? That’s the benchmark we’re not allowed to see.
