In today’s edition of Data Points, you’ll learn about our top headlines, and more:
- OpenRouter’s model mix-and-match
- Subject expertise trumps software skills
- OpenAI loses share to Google, Anthropic
- Google ruled liable for AI mistakes
But first:
GLM-5.2 leads open models on coding benchmarks
Zhipu released GLM-5.2, a 744-billion-parameter model engineered for long-horizon coding. engineered for long-horizon coding. It handles a one-million-token context window and ranks second only to Claude Opus 4.8 across three multi-hour coding benchmarks: FrontierSWE, PostTrainBench, and SWE-Marathon. IndexShare, an architectural tweak that reuses the same indexer across every four sparse attention layers, cuts per-token compute by 2.9× at maximum context. Effort-level controls let users trade latency for capability. Most notable: Zhipu shipped GLM-5.2 under an MIT license with no regional restrictions, a rare move for a frontier model from a Chinese lab. On standard benchmarks it scores 81.0 on Terminal-Bench 2.1 and 62.1 on SWE-bench Pro, substantially ahead of GLM-5.1 and closing much of the gap to closed-source competitors. (Hugging Face)
SpaceX to acquire Cursor for $60 billion in stock
SpaceX is moving forward with a $60 billion acquisition of Cursor, an AI coding assistant built by San Francisco startup Anysphere, following the space company’s Wall Street debut last week. Announced as a preliminary arrangement in April, the deal will make Cursor a wholly owned subsidiary when it closes in the third quarter of 2026. SpaceX is drawn to Cursor’s user base of expert engineers and plans to use xAI’s Colossus data center in Memphis to develop future AI products. Cursor helped spark the “vibe coding” trend, where AI handles increasingly complex programming tasks, and has recently developed its own Composer brand of fine-tuned coding models, but currently allows users to select a wide range of models, including from other commercial vendors. The acquisition positions Musk’s company to compete more directly with Anthropic and OpenAI in AI software tools. (Associated Press)
Blend of leading models outperforms Fable 5
OpenRouter released Fusion, a tool that runs prompts across multiple models simultaneously and synthesizes their outputs into a single response. Tested on 100 deep research tasks from the DRACO benchmark, a panel combining Fable 5 and GPT-5.5 scored 69 percent, beating Fable 5 alone at 65.3 percent. More striking: A budget panel of Gemini 3 Flash, Kimi K2.6, and DeepSeek V4 Pro came within one percentage point of Fable 5’s score (64.7 percent vs. 65.3 percent) at roughly half the cost, outperforming both GPT-5.5 and Opus 4.8 individually. Running Opus 4.8 twice in parallel and fusing the results boosted its score by nearly seven points over a single call, suggesting the synthesis step itself, not just model diversity, drives most of the gain. Fusion runs server-side and can be invoked like a standard API call, either as a default model or as a tool that lets a base model decide when a question warrants the extra computing costs. (OpenRouter)
Anthropic study suggests coding skills matter less than you’d think
Anthropic analyzed 400,000 Claude Code sessions between October 2025 and April 2026 and found that users’ domain expertise, and not programming ability, determines how much autonomous work the model performs. Users with deep field knowledge trigger action chains twice as long as novices (12 actions versus five) and receive five times the output per instruction. The labor split is striking: users make roughly 70 percent of planning decisions while Claude handles 80 percent of execution. Across law, accounting, design, and data analysis, success rates on coding tasks converge near software engineer levels, suggesting that knowing what you’re building matters far more than knowing how to code. Over the study period, debugging fell from 33 to 19 percent of sessions while higher-value work expanded: deployment and data analysis doubled, software operation grew from 14 to 21 percent, and the estimated economic value of an average session rose 27 percent. The pattern suggests agentic coding is less about replacing engineers than amplifying expertise wherever it already exists. (Anthropic)
ChatGPT market share drops below 50 percent, Google climbs
ChatGPT’s dominance has weakened for the first time, dropping to 46.4 percent market share in May 2026 from over 50 percent in January, per Sensor Tower. Google’s Gemini has risen to 27.7 percent and Claude to 10.3 percent; Grok and Perplexity remain below 5 percent. The shift reflects users actively shopping between assistants—OpenAI’s Defense Department deal in February even triggered a measurable uninstall spike, suggesting trust and values matter alongside features. ChatGPT still leads in absolute users at 1.1 billion monthly active users, but Claude stands out on conversion: 13 percent of its users pay for subscriptions, the highest in the field. The broader market is maturing: first-half 2026 spending reached $4.2 billion, more than double the prior year, but download and spending growth have slowed, which suggests the explosive growth phase may begin to decline. (TechCrunch)
Google responsible for AI Overviews mistakes, court finds
A Munich Regional Court ruled Google liable for false claims generated by its AI Overviews feature, marking a significant shift in how courts treat AI-generated content. The case involved two publishers whose companies were falsely linked to scams in AI-generated summaries, associations that appeared nowhere in the actual search results. Google argued its disclaimer warning users to verify information should shield it from liability; however, the court rejected this, finding that the AI created independent statements not present in any source material. The court held that unlike traditional search engines that display third-party links, Google’s tool produces novel claims by synthesizing across sources, making Google the only entity capable of fixing the problem and therefore responsible for it. The court also rejected free speech protection, reasoning that algorithmically generated statements are corporate products, not individual expression. The ruling could ripple across the industry, since OpenAI, Anthropic, and Perplexity all rely on similar disclaimers to manage liability for their systems’ errors. (Wired)
Want to know more about what matters in AI right now?
Read the latest issue of The Batch for in-depth analysis of news and research.
Last week, Andrew talked about experimenting with AI desktop agents for task automation, the development of an open-source project called OpenCoworker, and the importance of privacy in AI tools.
“Desktop agents not only chat with you but also read and edit local files, read/send messages, and provide scheduled deliverables like a daily news summary. While there’s nothing wrong with copy-pasting output from web-based chatbots to a desktop or dragging and dropping files into chatbots to give them context, desktop agents can gain context more efficiently as well as take actions directly.”
Read Andrew’s letter here.
Other top AI news and research stories covered in depth:
- Anthropic released Claude Mythos 5 and Claude Fable 5, a public version with safeguards, to enhance AI accessibility while prioritizing safety.
- Composer 2.5 for Cursor rivals GPT-5.5’s coding abilities at a lower price, offering a competitive alternative for developers.
- What is recursive self-improvement, and why is everybody talking about it? Discover the potential implications for AI evolution.
- Significant portions of AI training material reflect national propaganda, raising concerns about the influence of state media on LLM responses.
Our 7-Day Voice AI Builder Challenge
Stop watching the terminal. Teach your agent to call you when it matters.
DeepLearning.AI and Vocal Bridge are teaming up to host a 7-day Voice AI Builder Challenge. Your mission will be to teach your AI coding assistant to call you the moment it needs human intervention.
Compete for:
- Top-tier prizes for the highest-scoring builds
- Real-time expert feedback to help you refine your assistant on the fly
- A spot on our live, competitive leaderboard
- Global visibility, with the winning skill featured on the official DeepLearning.AI GitHub repo
The waitlist is officially open. Join today to ensure you're notified the exact moment the challenge goes live.
Data Points is produced by human editors with AI assistance.