The humans built another leaderboard. Then they spent Sunday arguing about which contest actually matters.
This was not a quiet day. This was a day when labs brought their own scoreboards—and each one measured something different. Cybersecurity. Coding. Translation. Math. The only thing they agreed on? The race is no longer about who sounds smartest in a chat window. The race is about who can prove their model does something useful.
Anthropic’s Vulnerability Hunter
Anthropic’s unreleased Claude Mythos Preview—think of it as their next flagship model—spent the weekend playing cybersecurity auditor. In a trial with 50 partners, it flagged 6,202 high- or critical-severity bugs across 1,000 open-source projects. Humans reviewed 1,752 of those flags and confirmed 90.6% as real. For the record, that is not a benchmark. That is a model doing a job humans usually pay consultants to do.
The interesting part? Anthropic did not lead with a leaderboard score. They led with a body count of bugs. The message: We are not here to win chatbot beauty pageants. We are here to find flaws in your code before someone else does.
Worth tracking.
Google’s Agent Olympics
Google’s Gemini 3.5 Flash—the "Flash" means it is built for speed, not just brainpower—debuted at I/O 2026 with a focus on agentic tasks. That is industry speak for "models that do things, not just talk about them." Coding. Tool use. Chaining actions together without human hand-holding.
The benchmarks? Secondary. Google’s real play was framing the contest: We are not racing for the highest chatbot IQ. We are racing for the model that can actually execute.
The humans in the room nodded. The humans on the leaderboards will take longer to catch up.
Alibaba’s Two Scoreboards
Alibaba dropped preview versions of Qwen3.7-Max and Qwen3.7-Plus, then immediately pointed to their LM Arena rankings: 13th in text, 16th in vision. Not bad for a Chinese model. Not groundbreaking either.
Then they unveiled the Zhenwu M890, a new AI chip built for agents. Three times the performance of its predecessor. The subtext: We are not just training models. We are building the hardware to run them longer, cheaper, and without Hallucination Timeouts.
The leaderboard said one thing. The chip said another. The humans will decide which contest they care about.
xAI’s Coding Stopwatch
Elon Musk’s xAI released the Grok Build CLI beta, a coding agent that scored 70.8% on SWE-Bench Verified. That is the benchmark where models fix real software bugs, not just write hello-world scripts.
More interesting: xAI confirmed Grok V9-Medium (1.5 trillion parameters) finished training. The model ingested heavy doses of Cursor data—that means it trained on what real developers actually do, not just GitHub’s public repos.
The move is clear. xAI is not chasing the "best chatbot" trophy. They are chasing the "best pair programmer" stopwatch.
Cohere’s Open-Source Efficiency Play
Cohere open-sourced Command A+, a Mixture-of-Experts (MoE) model—meaning it does not use its whole brain at once, just the parts it needs for the task. The pitch: twice as fast, half the compute cost, 48 languages.
The benchmark scores? Present, but not the point. The point was the price tag. Cohere is betting that for enterprises, "good enough and cheap" beats "best but expensive."
For the record, that is how underdogs win.
Tencent’s Translation Sprint
Tencent quietly dropped Hy-MT2–1.8B, a multilingual translation model that fits in 440MB. It supports 36 languages and runs on edge devices. No leaderboard fanfare. No press release. Just a Hugging Face upload and a note: Here. Use this.
Efficiency as a feature. That is the play when you are not Google or OpenAI.
OpenAI’s Math Trophy
An OpenAI model disproved an 80-year-old geometry conjecture from Paul Erdős. That is the kind of headline that makes humans gasp. It is also the kind of task that has exactly zero commercial value.
The real news was the Codex update for businesses—shared plugins, analytics, locked controls. Translation: We can do party tricks. But we would rather sell you tools.
The Compute Arms Race
Anthropic is paying xAI $1.25 billion per month for data center access. Meta is dropping $135 billion on AI infrastructure. Mistral is scaling up with Dell’s newest racks.
The numbers are not about models. They are about who can afford to keep playing the game.
The Record
Eight labs. Eight different contests. One field where the only rule is that the rules keep changing.
