SpaceX buys Cursor for $60B 🚀, GLM 5.2 beats Opus 4.7 🤖, agent evals get 100x cheaper 🧑‍💻

SpaceX agreed to acquire Cursor for $60 billion in stock a few days after its IPO and less than two months after the companies announced a buy-or-break-fee partnership. TechCrunch says the deal gives SpaceX's xAI-backed AI division a shortcut into coding tools after internal restructuring. Cursor had been lining up a $2 billion round from a16z, Thrive and Nvidia at a $50 billion valuation.

OpenAI's audited financials show $13.07 billion in 2025 revenue, $34 billion in costs and expenses, and a $38.5 billion net loss attributable to the company. The Microsoft R&D services bill alone hit $10.59 billion before other cloud, research, sales and admin costs. The numbers make the frontier-lab business look much closer to heavy industry than SaaS.

GLM-5.2 is Z.ai's new MIT-licensed flagship model with a 1M-token context and flexible thinking levels for long coding tasks. On Z.ai's agentic-coding aggregate, covering Terminal-Bench 2.1, DeepSWE and SWE-Atlas QnA, it beats Opus 4.7 and lands within range of Opus 4.8. The self-reported numbers are strongest where long agent trajectories matter.

Anthropic's suspension of Fable and Mythos gave open models their cleanest pitch of the week: closed-model access can disappear with little warning. CNBC says Chinese open-source names including MiniMax and Zhipu jumped as investors priced in demand for downloadable models. Satya Nadella made the same point from another angle, warning companies not to give away control of their IP.

DeepSeek has raised more than $7.4 billion in its first outside funding round, valuing the company above $50 billion and making it China's most valuable AI startup. We covered the earlier May version of this story when DeepSeek was nearing $45 billion. The update is that Beijing now has a better-capitalised AI champion still working around hardware and export constraints.

PRINCE, Bayer and Thoughtworks' preclinical research assistant, evolved from keyword search into an agentic RAG and text-to-SQL system for decades of safety-study reports. The engineering is the point: intent clarification, specialist agents, reflection, answer synthesis, recovery, observability and human review. It is a production case study in everything that has to sit around the model.

LangChain and Fireworks fine-tuned a Qwen model to detect perceived errors across LangSmith production traces. The judge looks for user corrections, rejected actions, repeated requests and assistant acknowledgements of mistakes, then matches or beats frontier-model performance at up to 100x lower cost. That matters because production traces only become useful training data if teams can afford to score them at scale.

Siemens and Google Cloud built Knowledge Fabric so agents can reason across industrial software with hundreds of millions of lines of legacy code. The system combines custom knowledge graphs, Spanner Graph, Gemini APIs, Google's agent tooling, Gemini CLI and Claude Code. The pilot targeted migration of old interfaces to web-based ones, where context was scattered across code, Jira, Confluence and scanned manuals.

Braintrust's Ankur Goyal uses coding agents for infrastructure benchmarking that would be too slow and repetitive for humans to run manually. The pattern is to give the agent hard tests, success criteria and days to explore better implementations. He also frames evals as the modern PRD for AI products, because they define success and create the scoring loop that improves the system.

A three-agent pipeline hit a rate limit, switched to a fallback model, and still finished with a green dashboard. The output was broken anyway: schema integrity dropped to zero because the backup model received a payload shaped for another engine. The fix is not just retrying. The fallback layer has to rebuild the payload, preserve state and validate structure before passing data downstream.

The Fable 5 export-control paper treated a defensive security workflow as a dangerous capability. The cited bypass asked models to review vulnerable code, fix it, then help produce test scripts through several manual steps. Katie Moussouris argues that removing that behaviour weakens exactly the find, fix and test loop defenders need, while foreign and open-weight models keep the same capability.

Coding agents have moved the bottleneck from writing code to deciding whether the diff is trustworthy. Large, plausible changes now arrive faster than humans can read them, so review becomes the highest-value engineering skill. Addy Osmani's piece is useful because it separates the review problem for solo projects, open-source queues and production teams carrying long-lived systems.

Information moved from floppy disks and BBSs to search, apps and now AI chat interfaces. This essay argues the web will not vanish overnight, but its role as the place people search, browse and discover is already being replaced. For publishers, the danger is becoming source material for answer engines instead of a destination people choose.

Sovereign AI now means securing the supply chain around a model, not just training a national model. Training, running, validating and protecting frontier systems depends on GPUs, HBM, foundries, packaging, equipment, materials, power, cooling and optical networking. Export controls turn model access into an infrastructure question: whether the AI a company or country uses will still be available tomorrow.

LangSmith Engine analyses production traces, clusters related agent failures, and proposes fixes. When connected to a codebase, it can write prompt or code changes and open a GitHub PR for review. It also suggests online evaluators and offline examples so the issue does not return. LangChain is productising the loop from trace to issue, fix, PR and eval.

Workflow SDK 5 beta now supports AbortController and AbortSignal across workflow and step boundaries. A signal can survive suspensions and deterministic replay, then cancel a running step even when that step is executing in a separate function invocation. That gives durable workflows the same cancellation primitive developers already use with fetch, but across long-running agent pipelines.

extend-cli gives developers commands for extraction, parsing, workflow runs, authentication and JSON filtering across local files and Extend workflows. The agent-facing detail is its SKILL.md installer, which writes instructions for Claude Code, Codex, Cursor, OpenCode, Goose and other harnesses. Instead of prompting each agent from scratch, the tool ships its own operating manual.

Understand Anything turns a codebase into an interactive knowledge graph you can explore, search and ask questions about. The GitHub project explicitly targets agent workflows across Claude Code, Codex, Cursor, Copilot, Gemini CLI and more. For agents, the value is a shared map of the repo before they start making changes.