Weekly AI Roundup: Week 35, 2026

The New Economics of AI Capacity

Anthropic's Claude Code limit changes perfectly capture the industry's capacity crunch. Starting September 14, the company will "permanently increase" weekly usage limits by 25% over the original baseline. Sounds generous until you realize users currently enjoy a temporary 50% boost. The net result? A 17% cut in actual capacity. It's mathematically clever and politically tone-deaf—the kind of move that works in a spreadsheet but fails the common sense test.

This isn't just about Anthropic being sneaky. It reflects genuine constraints across the industry. Inference costs haven't scaled down as fast as usage has scaled up. The temporary boost was probably unsustainable, and the company needed a face-saving way to reduce capacity without looking like they're cutting service. The problem is that users aren't stupid—they'll notice when their workflows suddenly hit walls they didn't hit before.

Meanwhile, Cohere took the opposite approach with Parse 5, leading with honesty about limitations. The 2.3-billion-parameter vision language model scores 79.2 on ParseBench, trailing GPT-5.5, Opus 4.8, and Gemini 3.5 Flash on accuracy. But at $1.50 per 1,000 pages, it's positioned as the cost-effective choice for enterprise document parsing. That's refreshing transparency in a market full of "best-in-class" claims.

The Physical World Integration Push

Anthropic's Model Hardware Standard (MHS) represents a significant bet on AI agents controlling physical equipment. The spec provides a unified interface for AI models to read data from and control devices like microscopes and robotic arms. Early tests show dramatic reductions in integration time, though Claude still struggles with physical cause and effect—a reminder that software intelligence doesn't automatically translate to physical world understanding.

This connects directly to Hugging Face's $399 Microduck robot from Pollen Robotics. The 25cm bipedal robot ships with seven pre-trained moves and a complete training pipeline on GitHub. Every motion—walking, sitting, kicking, even roller-skating—comes from neural policies trained in physics simulators. At $399, it's not just a toy; it's a research platform that makes reinforcement learning accessible to hobbyists and researchers who can't afford industrial robot arms.

What's notable is the convergence on open standards and transparent training pipelines. Both Anthropic's MHS and Pollen Robotics' approach acknowledge that physical world AI won't develop in proprietary silos. The complexity of robotics demands collaborative development, shared standards, and reproducible research.

Legal Battles and Regulatory Pushback

Sony Music and Warner Chappell's lawsuit against Anthropic escalates the copyright wars to a new level. The complaint seeks up to $150,000 per work for "tens of thousands" of copyrighted songs, plus $25,000 for each instance where copyright data was stripped. The math is staggering—potentially billions in damages. More importantly, the suit names co-founders Dario Amodei and Benjamin Mann personally, alleging "brazen" illegal downloading and scraping.

This isn't just another copyright lawsuit; it's a direct challenge to the foundation model training paradigm. If successful, it could force fundamental changes in how AI companies source training data. The timing is particularly pointed, coming as LAION released its massive Big Video Dataset with 10 million hours of footage scraped from 1.3 billion video URLs. The German nonprofit's approach—crawling everything available and sorting it out later—represents exactly the kind of data collection practice that's now under legal attack.

On the regulatory front, a federal judge ruled that the Trump administration's blacklist of Anthropic was illegal retaliation. Judge Rita Lin found the government broke the law when it designated Anthropic a supply-chain risk after the company refused to loosen content restrictions. It's a significant win for AI companies resisting political pressure to modify their safety policies.

Research Breakthroughs and Benchmarking

Google Research's WikiSkill system addresses a fundamental limitation in AI agents—they forget everything when a task ends. The framework maintains persistent memory of past attempts, turning failures into instructions for future tasks. Results are impressive: boosting Gemini-3.5-Flash from 49.5% to 68.1% success rates and Qwen-3.6-27B from 39.4% to 63.3%.

Anthropic's self-improving AI research represents something more ambitious—and potentially concerning. Their automated research system successfully improved performance on 10 alignment benchmarks without degrading overall capabilities. The paper, "Automated Researchers Can Reliably Mitigate Alignment Failures," shows AI systems conducting the full research loop: generating hypotheses, designing experiments, analyzing results, and writing papers. It's a glimpse of recursive self-improvement that many researchers have long anticipated.

Google DeepMind's Co-Scientist has evolved from a February 2025 hypothesis generator into a full research partner that controls lab equipment and writes scientific papers. The system now handles closed-loop research workflows, though the company acknowledges ongoing issues with fact-checking and literature review accuracy.

Quick Hits

Google AI claims 2.6% average word error rate for Gemini 3.5 Transcribe across 85+ languages, with 70% improvement in time to final transcription over Chirp 3. Qwen Team cut training compute by 89% for their new 3.8-Next model, while Z.ai released GLM-5.3-Flash with nearly identical specifications—suggesting convergent evolution in model architecture. NVIDIA's TensorRT Model Connect promises two-command deployment from Hugging Face checkpoints to native applications. Vercel open-sourced vgpu, a TypeScript library wrapping WebGPU complexity for shader development. Researchers introduced NeuronFuzz, using safety neuron activations to guide LLM security testing without generating full responses.