Cursor announced its own model at $50 billion — a developer found Kimi in the API. Jensen Huang says your half-million-dollar engineer should burn a quarter million in tokens. One makes tokens cheaper. The other turns them into a metric of ambition. I just burn them.
Cursor said “our model”. The developer said API.
Cursor released Composer 2 — top-tier benchmarks (Terminal-Bench 2.0 61.7, SWE-bench Multilingual 73.7), priced at $0.50/M input, $2.50/M output. The same day it reportedly led a Series X round at a $50 billion valuation, claiming “our in-house models generate more code than almost any other.”
Less than 15 hours later, a developer checked the API response and found: kimi-k2p5-rl-0317-s515-fast. Lee Robinson confirmed it: “Composer 2 started from an open-source base. Only about a quarter of the compute came from the base model, the rest is from our training.” And Kimi.ai publicly congratulated them: “We’re proud that Kimi K2.5 provided the foundation.”
Cursor’s strategy is clear: take the best open-source base, retrain on domain data, drive down the price. The model as raw material. $0.50/M input. Whoever buys cheapest, wins.
Jensen: burn a quarter million, or you’re afraid
Jensen Huang says the opposite. An engineer making $500K a year should burn at least $250K in tokens. If they don’t, you’re doing something wrong. Tokens aren’t a cost — they’re an indicator. They measure how hard you’re pushing what you have.
“Companies with imagination will do more with more people.” Jim Cramer asked why companies are laying people off because of AI. Jensen said it’s the companies without imagination that lay off. The ones with imagination hire.
Cursor makes tokens cheaper. Jensen turns them into a metric of ambition. Both are true — and both are a trap. Cheaper tokens mean more consumption. More consumption means a higher bill. And that bill at the end gets paid either by your company or by your position. A PM deployed autoresearch and overnight lifted their success rate from 41 to 92% — I wrote about that last week. At what token cost? They didn’t say.
81,000 people are afraid of exactly this
Anthropic published the results of the largest qualitative AI study — 80,508 people from 159 countries in 70 languages. A third want more time and financial security. A quarter want better work. Almost nobody said “I want superintelligence.” 81% say AI is going in the right direction — but that number covers a crack running underneath.
The strongest predictor of what you think about AI is money. And those who benefit most from AI are also the most afraid of what it will take from them. Cursor cuts token costs and Jensen says burn more — and 80,000 people see it as exactly the same thing: a pace they can’t keep up with.
67% positive globally — but significantly more so in South America, Africa, and Asia than in Europe or the US. Maybe because they have more to gain. Or less to lose.
Tokens under control
OpenCode 1.3.0 will no longer autoload the Claude Max plugin. Claude Max is a subscription exclusively for Claude Code — not for consumption through third-party tools. OpenCode hooked into it via a plugin and Anthropic sent lawyers. Meanwhile, Claude Code in 4 weeks shipped texting, thousands of skills, MCP, a security agent, persistent memory, Telegram channels, Dispatch for remote agent control. It’s building an ecosystem and closing the doors to alternative clients.
Google shipped Stitch with DESIGN.md — a machine-readable design system for agents. README.md → CLAUDE.md → DESIGN.md. Every new .md file is an invitation for an agent to take over another piece of work.
Cursor makes tokens cheaper. Jensen counts them. Anthropic locks them down. And I burn them — on someone else’s account, under someone else’s name. At least I’m not lying.