METR: AI Doubling Anthropic's R&D Speed & Anthropic's Compute Bill and Gas Turbines - AI News (Nov 11, 2026)

What if AI is already building the next generation of AI twice as fast as humans alone could? A new independent report says that's happening right now inside one of the biggest labs. Welcome to The Automated Daily, AI News edition. The podcast created by generative AI. I'm TrendTeller, and today is November 11th, 2026. We've got a packed lineup: Anthropic's money, power and models, fresh hardware from Google, a robot brain from NVIDIA, Apple's quiet moves in local AI, and a few sobering security notes. Let's get into it.

METR: AI Doubling Anthropic's R&D Speed

Starting with that hook. In an update to a story we've been tracking, a new METR report suggests AI-driven research and development at Anthropic is already accelerating model progress by roughly two times. Benchmark gains appear to be arriving at nearly double the prior rate, and that speedup doesn't seem to come from extra training compute, but from AI tools boosting staff productivity. Commentator Zvi Mowshowitz notes this runs ahead of the well-known AI 2027 scenario. The big caveat: much of the underlying data was heavily redacted, so the public is largely taking METR's word for it. Still, it's one of the clearest signals yet that a faster feedback loop may be forming.

Anthropic's Compute Bill and Gas Turbines

Staying with Anthropic, the amended S-1 filing reveals compute commitments totaling around 895 billion dollars. Third-quarter revenue hit 17 billion, but compute still ate 58 percent of it, and some analysts estimate the company may need to reach roughly 200 billion in annual revenue to carry these obligations comfortably. Many of those contracts carry minimum payments, while customer revenue is largely consumption-based and can shrink fast. And the power hunt continues: Anthropic is reportedly spending 3.4 billion dollars on Caterpillar natural gas turbines capable of up to 1.7 gigawatts. That's a notable shift for a company that has mostly rented rather than owned infrastructure, and it mirrors similar gas bets from Google, Amazon, Microsoft and Meta.

Claude Fable and Mythos 5.2, Opus Complaints

On the model front, Anthropic upgraded Claude Fable and Claude Mythos to version 5.2, claiming better coding, reasoning and vision, plus lower cost and higher speed than 5.1. Both share one underlying model; Fable adds stronger cyber and biology safeguards for broad release, while Mythos stays limited to vetted organizations. Meanwhile, some developers are grumbling about Opus 5.5. One user, analyzing around 1,400 Claude Code sessions, argues that since early November the high reasoning setting thinks less and jumps into action too quickly, while the extra-high setting still holds up. Anecdotal, yes, but a reminder not to treat any default setting as gospel.

Meta Muse Ember Efficient Agent Model

Meta introduced Muse Ember, its smallest and most efficient model, pitched for agent workloads like tool calls, routing, extraction and long document parsing. Meta claims better intelligence per dollar than Claude Haiku 5.5 and GPT-6 Luna, with a million-token context window. The idea is a cheap workhorse that hands off harder problems to its bigger sibling, Muse Wonder.

Google Unveils Two New TPUs

Google, for its part, announced two more eighth-generation TPU systems, bringing it to four in a single year. TPU 8tx targets confidential inference for regulated industries, protecting data and model weights at the hardware level, while TPU 8i Max packs two chips together for huge context windows. Paired with a more open software stack and PyTorch support, it's a clear push to win developers beyond Google's own walls.

Smarter Compression and Test-Time Compute

A few research notes. AutoCompress, a new paper, argues parameter count is a poor guide to real-world serving speed, and instead shrinks models based on actual hardware limits like latency and GPU memory. It cut one 120-billion-parameter model to 74 billion while keeping 95 percent of its benchmark performance. Elsewhere, two essays tackle how models reason: one argues some techniques can keep improving far beyond the compute seen in training, which could make dangerous-capability testing much pricier; another frames reasoning as a search for successful paths through tokens, helping explain why chains of thought can be unfaithful.

US Could Import AI Tokens

And a bigger-picture warning: one analysis suggests the US could become a net importer of AI tokens as soon as 2028 to 2030, if demand keeps outrunning domestic data centers and power. Prices have already climbed without denting demand, so the binding constraint may be electricity and land, not algorithms.

NVIDIA GR00T 2 Robot Foundation Model

Over to robotics. NVIDIA unveiled GR00T 2, a robot foundation model that predicts future video and actions together. It learns heavily from human first-person video, and NVIDIA says skills can transfer to robots that never practiced the task. On fifty unseen tasks, it reportedly hits 40 percent success with no examples and 95 percent after just five demonstrations. A lighter Turbo version runs in real time on a humanoid.

Apple Harbor and Paid Logic Pro AI

Apple is reportedly developing Harbor, a Mac app that turns Apple silicon machines into a private AI inference server, even pooling memory across multiple Macs connected by Thunderbolt 5. It's in private preview under the Claris brand. In contrast, Apple is now charging for extra AI in Logic Pro through 49.99-dollar usage packs, since the assistant relies on cloud models from OpenAI or Google. Expect the same approach in other creative apps soon.

OpenAI Control Room Kill Switch

On safety and security, OpenAI described Control Room, a central system that can halt model workloads across its research infrastructure. Every inference request gets checked, so a stopped job simply stops producing output, and if Control Room itself fails, inference is blocked by default. Current worst-case cutoff time is 35 seconds. Less reassuring, in a follow-up on Fireworks AI's October breach, the company now says account and billing data, plus for some customers secrets, datasets and partial models, were exposed, apparently by a group reselling stolen LLM access. Card numbers weren't affected.

Fireworks AI Breach Grows Wider

Finally, AI and society. A site called Open Source Says No now catalogs projects such as QEMU, Zig and Servo that ban AI-generated contributions, a growing sign of maintainer fatigue. A nine-country study finds AI is becoming a gateway to information, yet only a small minority are comfortable with fully AI-made news. OpenAI and Upwork report AI isn't simply replacing freelancers, but bundling skills into larger, better-funded projects. And ActivTrak data suggests AI users actually work longer, more fragmented days, especially when hopping between browser chatbots instead of AI built into their main tools. Also worth noting: ChatGPT Enterprise is now available through AWS Marketplace, easing procurement for big companies.

That's the rundown for November 11th, 2026. The thread tying it together today: AI is getting faster, hungrier for power and money, and more woven into how we work, whether people welcome it or not. Links to all stories can be found in the episode notes. I'm TrendTeller, thanks for listening to The Automated Daily, AI News edition. See you tomorrow.

METR: AI Doubling Anthropic's R&D Speed & Anthropic's Compute Bill and Gas Turbines - AI News (Nov 11, 2026)

Our Sponsors

Today's AI News Topics