Tim van den Bosch

Three months ago, the consensus was that Cursor was cooked. Claude Code crossed $2.5B in run-rate revenue. Google paid $2.4B for Windsurf’s IP and poached its leadership into DeepMind. OpenAI acquired Astral, the team behind Python’s uv package manager, to feed Codex. Viral tweets were circulating about developers ditching Cursor for Claude Code. The usage-based pricing switch last July had users posting surprise bills on Reddit. Consumer subscriptions were running at negative margins because every token served was profit for Anthropic or OpenAI. The company that popularized vibe coding was getting buried by the model providers it depended on. Then Cursor shipped four major releases in 15 days. JetBrains support on March 4. Automations on March 5. Plugin marketplace with 30+ partners on March 11. And now Composer 2, their own model that moggs Opus 4.6 on cost while matching it on performance. Look at the chart. Composer 2: 61.3 on CursorBench at $0.50 per million input tokens. Opus 4.6: 58.2 at $5.00. GPT-5.4: 63.9 at $2.50. The performance gaps are single digits. The cost gap between Composer and Opus is 10x. The part nobody’s pressing on: Cursor still won’t name the base model. Their blog says “our first continued pretraining run,” which means they took an existing model and continued training on code. When the original Composer launched in October, developers kept catching it responding in Chinese. Same tokenizer patterns as DeepSeek. Nathan Lambert congratulated the research team by tweeting “open weight base models + incredible ML teams in a specific niche can create immense value.” Co-founder Aman Sanger told Bloomberg it was trained exclusively on code. Can’t do taxes, can’t write poems. A Chinese open-source chassis, refined with what Cursor calls compaction-in-the-loop RL, and fed by a billion lines of daily user code flowing through the editor every day. That data flywheel is the one asset no API provider can replicate. The honest read requires some skepticism though. CursorBench is Cursor’s own internal benchmark. They built the test, then showed you they pass it. GPT-5.4 still leads on Terminal-Bench 2.0, which is independently maintained. And Opus 4.6 at high thinking effort still outscores Composer 2 on raw accuracy. The cost advantage is real. The performance parity claim needs external validation before anyone should take this chart at face value. But here’s why the chart matters anyway. This was the P0 coming out of the holidays. Building their own model was existential. Every dollar Cursor paid Anthropic per token was margin funding the competitor building Claude Code to replace them. Every dollar paid to OpenAI funded Codex. The only way to stop bleeding cash to the companies trying to kill you is to stop using their models. Four hundred employees. $2B ARR. Reportedly raising at $50B. Entering the model race against labs with thousands of researchers and tens of billions in compute. That chart is the fundraising slide. Whether it holds up in production against Opus and GPT-5.4 is a different question. But three months ago, the question was whether Cursor would survive at all.

The actual guide to agentmaxxing, since everyone’s going to misread this headline: Replit hit $240 million in revenue in 2025 with roughly 70 employees. That’s $3.4 million in revenue per head. A typical SaaS company at that revenue would have 700 people. Replit ran 10x leaner. Amjad Masad just raised $400 million at a $9 billion valuation and announced he’s hiring new grads. But the new grads he’s describing aren’t traditional CS majors grinding LeetCode. He hired an 18-year-old who never went to CS school, learned to code entirely through AI, and is outperforming classically trained engineers. Agentmaxxing is a specific workflow. You take an AI coding agent (Replit, Claude Code, Cursor), describe what you want in plain English, let the agent build it, review the output, iterate. One person running 5-10 agents simultaneously replaces a team of 4-5 junior engineers who each need onboarding, management, and code review. Masad said the quiet part out loud in an interview last year: if you’re an engineering manager at Meta, do you hire four junior engineers with all the overhead, or one senior engineer who can spin up 10 agents? Senior engineer salaries have never been higher. New grads who can’t orchestrate agents are struggling. New grads who can are getting hired at 18. The practical stack looks like this: 1. You become the architect, not the bricklayer. Your job is system design, constraint definition, and quality review. 2. You manage agents like direct reports. Break work into discrete chunks, assign each to a session, review output, course-correct. The best operators run parallel sessions. 3. Clarity of thought matters more than syntax knowledge. Masad said the highest-leverage hires right now are clear thinkers and clear communicators. He called them “consultant types.” 4. You ship 10x the surface area. Replit’s Agent 1 lasted 2 minutes before losing coherence. Agent 4 runs 3 hours doing production work. That capability is 10xing every few months. By next year, agents handle full-day tasks. The new grad who gets this builds more in month one than a traditional hire builds in a quarter. The hiring market for juniors didn’t collapse. It forked into two lanes: those who manage agents and those who compete against them.

Nano Banana 2 is out. I had early access for the past few days, and tested it across a ton of prompts. It's leveled up for a bunch of use cases - infographics, ads, action shots, even cartoons. And it's crazy fast! Some styles + prompts you should try 👇

Coalitieakkoord VVD, CDA en D66: “We stimuleren dat mensen hun spaargeld meer beleggen in de Nederlandse economie.” Dat was dus bullshit. Talk is cheap.