Another packed day in artificial intelligence. OpenAI dropped two major announcements — the GPT-5.6 family and GPT-Live voice — while xAI and Cursor shipped Grok 4.5, a highly economical frontier model. On the open-source side, a solo developer got the massive 744-billion-parameter GLM 5.2 running on a humble laptop. And in tech policy, the FTC secured a landmark right-to-repair settlement with John Deere. Here are the top stories.
FTC Secures Right-to-Repair Settlement with John Deere
The Federal Trade Commission, joined by attorneys general from Arizona, Illinois, Michigan, Minnesota, and Wisconsin, secured a right-to-repair settlement Wednesday with agriculture equipment giant Deere & Co. The settlement requires John Deere to provide farmers and independent repair shops with the diagnostic software and repair tools necessary to service their own equipment — access the company had long restricted to its authorized dealer network.
The antitrust lawsuit, filed in January 2025, argued that Deere had illegally monopolized the repair market for tractors, forestry equipment, and construction machinery by withholding its proprietary service software tool from independent mechanics. Under the settlement, Deere dealers are also prohibited from retaliating against customers who choose independent repair over authorized service. The company must pay $1 million collectively to the five states for antitrust enforcement costs and will operate under strict compliance oversight for the next ten years.
“For too long, Arizona farmers and independent mechanics have been at the mercy of Deere’s monopoly over repair tools, forced to wait — and pay — for authorized dealers just to fix broken tractors and other equipment,” said Arizona Attorney General Kris Mayes in a statement.
The settlement follows a separate $99 million class-action settlement Deere reached with farmers in April. Deere Vice President Denver Caldwell welcomed the agreement, saying, “This is good news for our customers and for the future of how Deere equipment is supported.” The case represents a significant escalation in the broader right-to-repair movement, which has gained momentum across the tech and agriculture sectors in recent years.
OpenAI Unveils GPT-5.6 Family: Sol, Terra, and Luna
OpenAI released GPT-5.6, a new model family comprising three variants under a fresh naming scheme: Sol for flagship frontier capability, Terra for balanced performance and cost, and Luna for efficient high-volume workloads. The gpt-5.6 API alias routes to Sol by default.
GPT-5.6 Sol achieves a new state-of-the-art score of 7.8% on ARC-AGI-3, making it the first verified frontier model to beat an ARC-AGI-3 game benchmark. OpenAI also published a safety system card documenting the model’s capabilities and deployment safeguards. According to OpenAI’s developer documentation, the model features improved “intent understanding” — better inferring a user’s underlying goal without requiring every step to be spelled out — and delivers significant gains in token efficiency across complex production workflows.
Notable benchmark comparisons suggest GPT-5.6 Sol leads against Anthropic’s Fable across multiple evaluations, though the gap varies by task. The model also shows marked improvement in frontend aesthetics, including layout, visual hierarchy, and design judgment. Pricing for the GPT-5.5/5.6 tier is reported at $5/M input tokens and $30/M output tokens, placing it at the premium end of the frontier model market.
OpenAI also recommends tuning the reasoning.effort parameter explicitly when migrating from GPT-5.5, as GPT-5.6’s improved efficiency may allow using a lower reasoning setting for equivalent or better results.
xAI and Cursor Ship Grok 4.5: Fast, Cheap, and Trained on Real-World Code
xAI, in collaboration with Cursor, released Grok 4.5 — a new frontier model that stands out for its remarkable efficiency. Priced at just $2 per million input tokens and $6 per million output tokens (with a fast variant at $4/$18), Grok 4.5 undercuts competitors by a wide margin: GPT-5.4 costs $2.5/$15, GPT-5.5/5.6 run $5/$30, and Anthropic’s Opus 4.8 and Fable are priced at $5/$25 and $10/$50 respectively.
What makes Grok 4.5 particularly interesting is its training data. The model was trained on “trillions of tokens of Cursor data” — real-world developer interactions with codebases and software tools — giving it an unusually grounded understanding of how developers actually work. Unlike Cursor’s previous Composer 2.5, which was a coding specialist, Grok 4.5’s training mix was deliberately broadened to include high-quality STEM tasks, research papers, and knowledge work spanning software engineering, data science, finance, and legal domains.
The model was refined using reinforcement learning on difficult problems in realistic environments, teaching it to investigate problems, use tools, recover from mistakes, and verify results. Cursor developed a distributed agent system to construct these training environments at scale — some of which would have taken “teams of hundreds of engineers months to build.”
Early benchmarks place Grok 4.5 around the Opus 4.7 level with roughly 4x better reasoning efficiency per dollar. HN users report inference speeds around 90 tokens per second — easily outpacing GPT-5.5, Opus 4.8, and GLM 5.2 — while delivering strong results on real-world coding tasks including native iOS app development. Grok 4.5 is available immediately in Cursor across desktop, web, iOS, CLI, and SDK, with double usage allocations for the first week.
GPT-Live: Real-Time Voice AI with Full-Duplex Conversation
OpenAI also introduced GPT-Live, a new real-time voice model built on a full-duplex architecture — meaning it can speak and listen simultaneously, like a natural human conversation. The model can also delegate complex queries to GPT-5.5 in the background, ensuring users aren’t limited to a voice model that lags behind the text frontier.
In a demonstration video, the model handled real-time language translation and natural banter with elderly users, positioning it as a potential companion and assistant for everyday conversation. OpenAI’s Atty Uttamre confirmed on HN that GPT-Live-1 is “the first version of a new generation of models” and that the full-duplex architecture “enables entirely new ways of human-AI interaction.”
Preview users report mixed reactions. One HN user described a one-hour uninterrupted brainstorming session while walking the dog, praising the ability to stay in voice mode throughout. Another noted that the model sometimes interrupts too eagerly or laughs at unintended jokes. Community feedback also highlighted a significant gap: none of the frontier assistants — ChatGPT, Claude, Gemini, or Grok — currently support tools or connectors while in voice mode, limiting productive use cases like researching, pulling up documents, or jotting notes mid-conversation.
Despite these rough edges, GPT-Live represents a meaningful step toward more natural human-AI interaction, and multiple commenters noted its potential as a life-changing accessibility tool for blind users once video capabilities and smart glasses integration arrive.
One Developer’s Quest: Running GLM 5.2 (744B Parameters) on a 32GB Laptop
In a remarkable feat of software engineering, a solo developer — going by the handle “vforno” — managed to run GLM 5.2, Zhipu AI’s massive 744-billion-parameter Mixture-of-Experts model, on a 12-core laptop with just 25 GB of usable RAM. The project, called Colibrì, is a minimalist C inference engine (roughly 1,300 lines) that makes novel use of disk streaming to sidestep memory constraints.
GLM 5.2 activates only about 40 billion parameters per token thanks to its MoE architecture. Colibrì keeps the dense portion — attention, shared experts, and embeddings (~17B parameters) — resident in RAM at int4 precision (~9.9 GB). The 21,504 routed experts (~370 GB total) live on disk and are streamed on demand via a per-layer LRU cache, an optional pinned hot-store, and the OS page cache as a free second-level cache. The result: about 0.1 tokens per second on the slowest hardware, but it works.
“The important thing was the journey to reach this goal,” the developer wrote. “I just wanted it to work at all costs, even slowly.” The project has sparked a lively HN discussion about disk-streaming inference, with multiple developers sharing similar approaches for other large models. One commenter is working on a Metal-backed macOS variant targeting Apple Silicon’s unified memory, while others noted that llama.cpp already supports mmap-based model loading for similar purposes, though Colibrì’s advantage is its extreme portability and minimal dependencies — no BLAS, no Python at runtime, and no GPU required.
It is a striking demonstration of how far open-source inference has come: a model that would have required a server cluster just two years ago can now be coaxed into running on consumer hardware, albeit slowly. The full project is available on GitHub.
Closing Thoughts
Today’s headlines capture an industry moving on multiple fronts simultaneously. OpenAI is pushing on both the pure intelligence frontier (GPT-5.6) and the interaction frontier (GPT-Live). xAI and Cursor are proving that specialized, efficiently trained models can compete with — and in some dimensions outperform — the big incumbents at a fraction of the cost. And the open-source community continues to democratize access, finding clever ways to run billion-parameter models on hardware that would have been laughed at just a few years ago. Meanwhile, regulators are increasingly holding tech giants accountable for how their software locks in consumers. It was a busy day in AI.
