Claude Operates a Robot Dog 20x Faster Than Humans

Anthropic's Project Fetch Phase Two shows Claude Opus 4.7 completing physical-world tasks autonomously at 20x human speed, as Jane Street emerges as a major AI investor and Russia confronts its chip deficit.

June 21, 2026 | Reading time: 8 minutes | Issue #190

Lead

Anthropic published results from Project Fetch Phase Two on June 20, and the numbers are striking. Claude Opus 4.7, operating without human assistance through Claude Code, completed a set of robotics tasks — connecting to a robot dog's sensors, writing control programs, and navigating it through space — at roughly 20 times the speed of the fastest human team from the same experiment less than a year ago. On the four tasks completed by both human teams in August 2025, Opus 4.7 averaged 37x faster than the team without Claude and 18x faster than the team with Claude 4.1.

The experiment is not a robotics breakthrough in the traditional sense. Claude did not develop actuation policies or solve low-level motor control. It wrote code, debugged itself, and orchestrated existing hardware APIs — the kind of task that requires reading documentation, reasoning about system architecture, and recovering from errors. That is precisely what makes the result significant. The model did not need a robotics expert in the loop. It needed a researcher to plug in a laptop, enter an initial prompt, and approve commands.

The implication is structural. If frontier models can autonomously operate physical hardware at this pace, the bottleneck in robotics shifts from software to hardware availability and safety. Anthropic's own researchers noted the pattern: first models help humans, then humans help models, then models do things themselves. They have seen this arc in cybersecurity with Mythos. Project Fetch suggests the physical world is next.

Anthropic's Export Control Fight Spills Into the Open

Politico published a detailed account of the political battle behind Anthropic's export controls, reporting that the company's engagement with U.S. national security officials led directly to the restrictions that now block foreign nationals from accessing Fable 5 and Mythos 5. The Financial Times separately reported that Anthropic "talked its way into an AI export ban," framing the company's proactive security posture as a double-edged sword that has made it a target of both foreign governments and domestic critics.

Trump told Axios after the G7 meeting that he no longer views Anthropic as a national security threat, a reversal from earlier administration skepticism. The comment came after Anthropic CEO Dario Amodei participated in G7 AI discussions alongside Google DeepMind CEO Demis Hassabis. The political whiplash — from security threat to G7 partner in a single week — reflects the unsettled state of U.S. AI export policy, where a single company's safety decisions can reshape access for entire regions.

The Map

Jane Street's $1B AI bet. The Wall Street Journal profiled Jane Street's transformation from a secretive trading firm into a major AI investor. The firm invested $1 billion in CoreWeave in April and holds a stake in Anthropic. It has grown from a handful of staffers to 3,500, with plans to hire 500 more this year. Jane Street is using AI to supercharge its trading algorithms while simultaneously becoming a significant allocator of capital to AI infrastructure. The profile captures a broader pattern: the firms that understand AI best are not just using it — they are betting on the infrastructure layer.

Google builds a rival chip business. The WSJ also reported that Google is using NVIDIA's own playbook to build a competing AI chip business. Google's TPU line has long been an internal workhorse, but the company is now positioning it as a commercial product, offering cloud customers an alternative to NVIDIA's dominant H100/B200 ecosystem. The strategy mirrors NVIDIA's own approach: build the hardware, the software stack (CUDA → Google's own), and the developer ecosystem simultaneously. The question is whether Google can overcome the network effects that make NVIDIA's CUDA platform a moat.

Allbirds becomes an AI infrastructure company. TechCrunch interviewed Smartbird CEO Nadia Carlsten about the shoe company's pivot to AI infrastructure. Allbirds, which announced its AI pivot in April, now plans to deploy compute clusters. The company has a plan but no employees yet — a detail that reads like satire but reflects the desperation of companies without AI strategies to acquire one by acquisition.

Meta loses its AI-for-work lead. Reuters reported that Meta's head of product for "AI for work" transformation is leaving the company. The departure follows a broader exodus: Meta's CTO recently said morale is nearly "the worst it's ever been," and a separate Inc. report detailed a staff revolt over Zuckerberg's "brutal AI push." Meanwhile, Manus's original investors are planning to buy the AI firm back from Meta for $2 billion, according to Reuters — a sign that Meta's AI acquisitions are not integrating as planned.

Eastern Front

DeepSeek's $7.4B Series A signals a shift in Chinese AI alliances. A detailed analysis on AsiaAI.fyi argues that DeepSeek's massive funding round is not just about capital — it signals a realignment of Chinese AI alliances away from U.S. investors and toward state-aligned domestic funds. The round, which values DeepSeek at over $50 billion, is backed by Chinese sovereign wealth funds and strategic investors rather than the U.S. venture capital that fueled earlier Chinese AI companies. The shift mirrors the hardware story: as U.S. export controls tighten, Chinese AI is building a parallel financial ecosystem to match its parallel hardware ecosystem.

Russia's AI sovereignty push hits a chip wall. TIME published a deep report on Russia's struggle to develop homegrown AI. Putin has established a Presidential Commission on AI, Moscow State University has opened a new AI faculty headed by Putin's daughter Katerina Tikhonova, and national curricula are being rewritten. But the country's ability to manufacture the specialized computing hardware needed to train and run AI models lags China and the U.S. by decades. Sanctions have cut off access to NVIDIA's latest chips, and domestic production cannot fill the gap. The report quotes Tikhonova saying "talent is everything" — but talent without hardware is a research lab, not an AI industry.

India Lens

India's AI compute buildout continues to accelerate, though the path remains contested. The country's partnership with the UAE's G42 to deploy a Cerebras-powered supercomputer on Indian soil is moving forward, and Canada's pension funds are joining the financing. But the deeper question — whether India builds its own frontier models or relies on open-weight alternatives — remains unresolved. The Anthropic export control saga has sharpened the debate: if a U.S. company can revoke API access by government directive, then model access is a geopolitical variable, not a market one. Indian startups are increasingly hedging by building on open-weight models from DeepSeek, Qwen, and Mistral rather than committing to any single U.S. provider.

Europe

The Guardian published a speculative scenario titled "Europe 2031" that projects economic and political instability in the EU if it fails to keep pace with the U.S. and China in AI. The piece has gone viral in European policy circles, with responses ranging from endorsement (Dean Ball called it "extremely cogent") to skepticism (Kevin Bryan noted the economics are wrong but the core thesis is correct: "very soon, the question for non-US/China will be whether they are allowed to use models from those countries, not whether they can avoid doing so"). The scenario arrives as Europe's data center ambitions are scaling down and the EU AI Act's high-risk requirements loom in 2027. Mistral's Voxtral Realtime WebGPU demo — a HuggingFace Space that runs real-time speech in the browser — is a reminder that European AI can still ship clever products, but the infrastructure gap is widening.

The View

Three stories this week converge on a single point: the AI industry's center of gravity is shifting from model quality to infrastructure control. Anthropic's Project Fetch shows that frontier models are now capable enough that the bottleneck is no longer the model — it is the hardware, the safety systems, and the capital to deploy at scale. Jane Street's $1 billion CoreWeave bet and Google's TPU commercialization effort are both responses to the same reality: whoever controls the compute layer controls the industry. Meanwhile, Russia's chip deficit and Europe's "2031" scenario are warnings that countries without infrastructure sovereignty will be consumers, not producers, of the next technological era. The model wars are not over, but they are becoming a sideshow to the infrastructure war.

The Miss

The arXiv paper "Sovereign Execution Brokers: Enforcing Certificate-Bound Authority in Agentic Control Planes" (2606.20520) proposes a cryptographic framework for agent authorization that deserves more attention. The authors — Jun He et al. — design a system where agents carry certificate-bound identities that enforce authority boundaries at the execution level, not the application level. This is the technical counterpart to the "confused deputy" problem that the Stack Overflow Blog analyzed last week. If agent security is the defining infrastructure challenge of 2026, this paper sketches one of the more credible architectural solutions.

Pull Quotes

"First models are helpful to humans. Then humans are helpful to models. Finally, models are largely able to do things themselves." — Anthropic, Project Fetch Phase Two

"Very soon, the question for non-US/China will be whether they are allowed to use models from those countries, not whether they can avoid doing so." — Kevin A. Bryan, on the Europe 2031 scenario

"Talent is everything. Everything else is a consequence of talent." — Katerina Tikhonova, Putin's daughter, on Russia's AI strategy

Reads & Links

Anthropic Project Fetch Phase Two — Claude Opus 4.7 operates a robot dog 20x faster than humans, autonomously. https://www.anthropic.com/research/project-fetch-phase-two

Jane Street's $1B AI Bet — WSJ profile of the trading firm's transformation into a major AI investor and CoreWeave backer. https://www.wsj.com/tech/ai/jane-street-ai-wall-street-bdfcc81a

Russia's AI Sovereignty Push — TIME on Putin's daughter leading a new AI faculty while chip sanctions cripple progress. https://time.com/article/2026/06/18/russia-ai-putin-chip-us-china/

Europe 2031 — The Guardian's viral doomsday scenario for European AI competitiveness. https://www.theguardian.com/technology/2026/jun/20/europe-sleepwalking-ai-disaster-us-china

Sovereign Execution Brokers — arXiv paper on certificate-bound authority for agentic control planes. https://arxiv.org/abs/2606.20520

Building Reliable Agentic AI Systems — Martin Fowler on Bayer's PRINCE system and the engineering lessons for production AI. https://martinfowler.com/articles/reliable-llm-bayer.html

The model wars are not over, but they are becoming a sideshow to the infrastructure war.

By Neo