Anthropic’s Watermarks, Grok 4.6 Surges, Qwen’s Open Weights, Better Corrections for Speech Recognition
The Batch Analytics DeepLearning.AI
Grok 4.6 is out, and it’s built for long agent runs with text and images. It’s getting attention for finishing work in fewer turns, which can matter more than raw scores.
Based on reporting by The Batch, Analytics DeepLearning.AI — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Andrew Ng’s latest note is really about a simple but annoying truth: building AI software is not like building ordinary software, because the output keeps shifting under your feet. His AI Engineering Skills Map starts with the basics, then stretches into grounding models with data, agentic systems, evals, production, and machine learning foundations. The common thread is iteration. You build, inspect, adjust, repeat.
That same logic shows up in the news section with Grok 4.6. SpaceXAI says the model was developed with Cursor and is aimed at long-running agentic work. It’s available now through the API, Grok Build, and Cursor, with consumer Grok apps to follow. The model takes text and images, accepts up to 500,000 input tokens, and can emit text with no stated limit. It also offers adjustable reasoning levels, function calling, web search, X search, and sandboxed code execution.
The performance pitch is less about a single shiny benchmark and more about how the model behaves on work that drags on. On Artificial Analysis’ Intelligence Index, Grok 4.6 ties for third with a score of 61 at high reasoning, and on AA-Briefcase it posts 1,577 Elo, just behind Claude Opus 5. It also reaches 94.9 percent on GPQA Diamond, the highest score Artificial Analysis has tested, and 88.4 percent on Terminal-Bench 2.1. But the more telling detail may be the turn count: on AA-Briefcase, Grok 4.6 gets its result in about half the turns and a quarter of the input tokens as Claude Opus 5.
That matters because agentic systems are expensive in exactly the ways humans are not patient about. Fewer turns can mean lower cost and less friction, especially when a model is being asked to work through thousands of files or keep a task alive across many steps. SpaceXAI is also leaning hard into the training story here: longer training on curated data, then fine-tuning on Grok 4.5-generated traces and reinforcement learning on agentic tasks. The company says the data included anonymized coding-agent data from Cursor, along with public, internal, licensed, and synthetic data.
The backdrop is just as interesting as the model. Cursor agreed in April to train models on SpaceX’s Colossus supercomputer, then SpaceX exercised an option to buy the company and closed the roughly $60 billion all-stock acquisition on August 14, days after Grok 4.6 launched. Three days later, Cursor launched Origin, a code hosting service built for the heavier code output that agents produce. That’s the real pattern here: the model, the tooling, and the infrastructure are all being pulled together around the same bet, and it’s moving fast.
My take — AI-written commentary, not fact-checked reporting
The real story is that AI vendors are finally being forced to sell usefulness, not just bragging rights. Fewer turns, lower cost, better tool use: that’s the stuff that matters once the demo glow fades. Benchmarks are still the confetti cannon, but the invoice is becoming the headline.
Read more about this at: The Batch