What's new in DevEx - June 18

Days after its IPO, SpaceX signed an all-stock deal to buy Cursor at a ~$60B equity value, closing Q3 pending approval. The most-used AI coding tool will now sits inside a frontier lab, with a jointly trained Grok-Cursor model in the pipeline. For the millions who code in it daily, independence and roadmap just became open questions.

Days after Fable 5 shipped, US ordered Anthropic to block it for any foreign national worldwide, citing a "jailbreak" that let the model hunt security flaws. Anthropic pulled it for everyone. But as Anthropic has pointed out in their announcement, that "jailbreak" was just asking the model to fix code with known CVEs.

Z.ai dropped MIT-licensed GLM-5.2 (753B total / 40B active) — now the top open-weights model on the Artificial Analysis Intelligence Index (51), at 81.0 on Terminal-Bench 2.1, within four points of Opus 4.8. It beats GPT-5.5 on long-horizon coding at a sixth of the cost. Right after Anthropic showed a cloud model can vanish overnight, a self-hostable frontier coder lands that can't be pulled.

At Ship 2026, Vercel open-sourced eve: a filesystem-first framework where each agent is just a directory of files, with durable execution, sandboxing, human-in-the-loop approvals, and evals built in. Companion launch Vercel Connect hands agents short-lived, task-scoped credentials instead of long-lived secrets when they hit Slack, GitHub, or any MCP endpoint. Agents now write over half of Vercel's commits, up from 3%.

GitHub detailed harness-level changes that quietly cut Copilot's bill: deferred tool loading (tools fetched via a search call, not front-loaded) trims tokens ~18% for the median user; cache-aware routing keeps requests on warm models; and OpenAI prompt-cache retention now runs 24 hours, hitting ~94% cache rates on Anthropic models. The payoff: 16–19% faster first tokens and a smaller bill.

Hugging Face shipped agent-eval, a harness that scores not whether an agent solved a task but how much work it took. On transformers it caught a trap: an agent CLI + Skill sped up big open models but broke small ones — Qwen3-14B's sentiment task fell 100% → 0% when it mistook the docs for a callable tool and gave up. The lesson: test agent APIs across model sizes.

🌉 AI Engineer World's Fair is coming to SF on June 29 – July 2.

The largest technical AI conference in the world — 6,000+ AI engineers, founders, and VPs of AI, 300 speakers across 29 tracks, and 100+ expo partners including Anthropic, OpenAI, Google DeepMind, Braintrust, Modal, and PostHog. Hotel room blocks close today and are expected to sell out fast — so if you're going, book now. Grab tickets here.

Till next time,