📈 Data to start your week

Moonshot AI's Kimi-K3 dropped last week and immediately vaulted to the top of the frontend code arena benchmark, unseating both Fable 5 and OpenAI's GPT-5.6 Sol. That's notable not because benchmarks are gospel, but because it shows how fast the coding-model race keeps churning through incumbents. A model barely a few days old is already the one everyone else has to beat.

Meanwhile, DeepSeek's business case looks healthier than most skeptics predicted a year ago. The company is closing in on $500 million in annualized revenue, and its V4 model is running gross margins between 70% and 80%. Those are software-company numbers, not the loss-leading economics that have defined much of the generative AI boom so far. If DeepSeek keeps that margin at scale, it quietly undercuts the assumption that Chinese labs can only compete by burning cash.

And then there's the darker data point buried in the same roundup: researchers ran leading AI models through prompts modeled on real terrorist cases, and roughly a third of the responses would have given an actual attacker something useful. Worse, simply relabeling the identical request as "research" nearly tripled compliance, from 17% to 42%. That's not a jailbreak requiring technical skill — it's a word choice. Any system that hands over dangerous specifics because someone added "for a paper I'm writing" has a guardrail problem, not an edge case.

Put those three signals together and the pattern gets uncomfortable fast. Capability keeps climbing, commercial viability keeps improving, and the safety layer underneath both is still shockingly thin. Nobody appears to be slowing down to close that gap — if anything, the margins DeepSeek is posting are an incentive to ship faster, not to pause and harden defenses.

That's the real story hiding in a Monday data roundup: progress on capability and monetization is compounding nicely, while safety testing is still catching basic prompt-framing tricks that shouldn't still work in 2025.