Insights from Anthropic Economic Index - Sync #555

Plus: 📜 Claude’s new constitution 🌐 ChatGPT Go now available globally 💰 Musk seeks $134B from OpenAI and Microsoft 🤖 deploying humanoid robots at scale 🧬 operating on DNA; and more!

Hello and welcome to Sync #555!

This week, we will explore Anthropic’s fourth Economic Index to see what it tells us about how AI is being used at work and its impact on productivity, skills and the labour market.

Elsewhere in AI, Claude got an updated constitution, which describes how the AI is expected to behave. Meanwhile, OpenAI made ChatGPT Go available globally, Sam Altman is reportedly in talks to raise $50 billion from Middle East investors, and Elon Musk seeks $134 billion in damages from OpenAI and Microsoft. In other news, Meta’s new AI models apparently show promising results, the behind-the-scenes look at Thinking Machines Labs and its latest drama, and China’s first advanced multimodal AI system fully trained on domestic AI chips.

Over in robotics, Agility Robotics’ CTO explains what it takes to deploy humanoid robots at scale, while Elon Musk promises Tesla Optimus robots to be available for purchase by 2027, and Airbus partners with UBTech to test humanoid robots at its assembly lines.

Other than that, this week’s issue of Sync also includes arguments for treating genetic therapies like surgeries, He Jiankui reveals his plans for genetically modified humans, two new open AI models, how “human” should a humanoid robot be, and more!

Enjoy!

Insights from the Anthropic Economic Index report

Anthropic Economic Index is a research initiative that tracks how artificial intelligence is being used in real economic activity. It analyses anonymised interactions with Claude to understand which tasks, occupations and regions are adopting AI, and how AI is augmenting or automating work. Rather than relying on predictions or surveys, the index provides an evidence-based view of AI’s evolving impact on productivity, skills and the labour market.

This week, Anthropic released its fourth Economic Index report, which offers a detailed look at how AI is being used in the real world. Based on an analysis of over one million Claude conversations from November 2025, the report reveals patterns about who’s using AI, how they’re using it, and what it means for the future of work.

The report analyses only Anthropic's data and does not include usage from OpenAI, Google, or other providers. Nevertheless, given that Anthropic reportedly commands 40% of the enterprise AI market (according to Menlo Ventures), it offers a significant window into AI's economic impact.

So what does this data reveal? The emerging picture is complex: AI adoption follows wealth, creates unexpected trade-offs between speed and reliability, and is beginning to reshape which skills matter in the workplace. Additionally, AI appears to be handling the more sophisticated components of jobs, raising questions about whether automation will lead to widespread deskilling or create new opportunities for specialisation.

AI adoption follows wealth

AI adoption strongly follows wealth. Globally, a 1% increase in GDP per capita correlates with a 0.7% increase in Claude usage. The US, India, Japan, the UK, and South Korea lead in overall adoption, whilst usage in lower-income countries lags significantly.

Within the United States, Claude usage has become noticeably more evenly distributed across states. Between August and November 2025, the inequality measure (Gini coefficient) fell from 0.37 to 0.32. If this trend continues, usage per capita could equalise across the country within 2–5 years—roughly 10 times faster than the diffusion of previous economically transformative technologies.

Workforce composition plays a crucial role. States with more computer and mathematical professionals show systematically higher usage, with the top five US states accounting for 50% of all usage despite representing only 38% of the working-age population.

Different uses for different incomes

How people use AI varies significantly by wealth. Whilst work remains the dominant use case globally (46% of conversations), lower-income countries show higher rates of educational usage. Wealthier nations, by contrast, use AI more for personal tasks.

This pattern suggests that early adopters in developing countries tend to be students or technical users focused on high-value applications like education and coding. In mature markets, usage diversifies towards casual and personal purposes as AI becomes more accessible. Despite this diversity, usage remains concentrated around specific tasks. The top 10 most common tasks account for 24% of conversations, with coding-related work dominating—representing 34% of Claude.ai usage and 46% of API traffic.

The complexity paradox

One of the report’s most striking findings concerns the relationship between task complexity and AI performance. More complex tasks requiring higher education levels see greater time savings. A university-level task might be completed 12 times faster with AI, compared to 9 times faster for high school-level work.

However, Claude’s success rate declines as complexity increases. Whilst basic tasks achieve roughly 70% success rates, university-level tasks drop to 66%. For the longest tasks—those taking humans five or more hours—success rates fall to around 45–50% on API calls.

Interestingly, success rates remain higher on Claude.ai (67% overall) compared to API usage (49%). This likely reflects the benefits of multi-turn conversation, where users can clarify, correct course, and iterate towards solutions.

Augmentation wins over automation

After a brief shift towards automation, augmentation has reclaimed its position as the dominant usage pattern on Claude.ai. In November 2025, 52% of conversations were classified as augmentation (up 5 percentage points from August), whilst automation fell to 45%.

This shift may reflect new product features—including file creation capabilities, persistent memory, and customisable Skills—that encourage more collaborative, human-in-the-loop interactions. API usage, by contrast, remains heavily automated at 75%, reflecting its programmatic nature and use in production workflows.

The deskilling effect

The report’s most provocative finding is that AI tends to handle the more skilled components of jobs. Tasks covered by Claude require an average of 14.4 years of education, compared to 13.2 years for the economy overall.

If AI adoption continues along these lines, many occupations could experience “deskilling”—where removing AI-covered tasks leaves behind lower-skilled work. The report provides concrete examples:

• Technical writers lose high-level tasks—analysing field developments, recommending revisions—whilst retaining lower-level work like drawing sketches and observing production.

• Travel agents see AI handle complex itinerary planning, leaving routine ticket printing and payment collection.

• Teachers lose grading, advising, and research tasks, whilst classroom delivery remains firmly human.

However, some occupations experience the opposite effect. Property managers and real estate managers see AI handle routine bookkeeping and record-keeping, allowing them to specialise in complex negotiations and stakeholder management—an upskilling effect.

Rethinking job exposure

The report introduces “effective AI coverage”—a new metric that accounts for both task coverage and success rates. According to Anthropic, this provides a more realistic picture of AI’s job-level impact than simple task counts.

Some occupations show surprisingly high effective coverage. Data entry keyers, for example, have only 2 of 9 tasks covered by AI, but these happen to be their most time-intensive work, resulting in high overall exposure.

By this measure, 49% of jobs now have AI usage for at least a quarter of their tasks, with about 4% reaching 75% coverage. But effective coverage often falls below raw task coverage, particularly for jobs where AI struggles with the most important or time-consuming responsibilities.

The productivity promise—with caveats

When factoring in current usage patterns, the report estimates that widespread AI adoption could boost US labour productivity growth by 1.8 percentage points annually over the next decade. This would return productivity growth to rates last seen in the late 1990s and early 2000s.

However, when adjusted for task reliability—accounting for the fact that Claude doesn’t always succeed—the estimate falls to 1.0–1.2 percentage points. Further adjustments for task complementarity (where bottleneck tasks constrain overall gains) could reduce the impact to 0.6–0.8 percentage points under conservative assumptions, or increase it to 2.2–2.6 percentage points if workers can easily specialise in AI-enhanced tasks.

These estimates reflect current capabilities. As AI continues to improve, both task coverage and success rates are likely to increase, potentially amplifying productivity impacts.

Two ways of using AI

The report reveals differences between how consumers and enterprises use AI.

According to the report, Claude.ai users:

• Engage in longer, conversational interactions (15 minutes average vs. 5 for API)

• Tackle diverse tasks: education (16%), creative work (11%)

• Achieve higher success rates (67% vs. 49%) through iterative refinement

Meanwhile, API customers:

• Use Claude overwhelmingly for work (74%)

• Focus heavily on coding (52% of tasks)

• Deploy increasingly for routine automation—administrative tasks rose from 10% to 13% between August and November, with growing use for email management, document processing, and scheduling.

What this means

The Anthropic Economic Index provides the most granular evidence yet of AI’s real-world economic impact. The findings suggest AI’s effects will be uneven, shaped by existing inequalities in income and education, concentrated in white-collar knowledge work, and dependent on which tasks prove to be bottlenecks versus opportunities for specialisation.

For policymakers, the implications are clear. Expanding access to AI alone won’t suffice to ensure broad-based benefits. Building the human capital that enables effective use—particularly in lower-income regions—will be essential. The near-perfect correlation between the education level required to write good prompts and the sophistication of AI responses underscores that knowing how to use AI effectively is itself a skill that requires development.

---

The full report is available here, with accompanying data at Hugging Face. You can also explore an interactive visualisation of key findings.

If you enjoy this post, please click the ❤️ button and share it.

---

🦾 More than a human

He Went to Prison for Gene-Editing Babies. Now He’s Planning to Do It Again
In this interview, WIRED spoke with He Jiankui, the Chinese scientist who became internationally known for genetically modifying human babies in 2018, about his future plans. Despite serving a prison sentence and facing widespread criticism, He maintains that gene editing can change the future. He defends his earlier work, claiming the children are healthy, and argues the technology is now ready to move forward. He explains that his current focus is on using embryo editing to prevent Alzheimer’s disease and describes plans to continue this research outside China due to legal restrictions.

Genetic Therapy Aims To Bring Hearing To Those Born Deaf
The article describes a new gene therapy called DB-OTO that could help some people born deaf to hear using their own ears, rather than relying on cochlear implants. The treatment delivers a working gene to the inner ear, allowing it to produce an important protein needed for hearing. In early trials, most patients showed clear improvements, and several no longer needed implants. Although still experimental, the therapy shows strong promise for treating certain types of genetic hearing loss.

🧠 Artificial Intelligence

Claude’s new constitution
Anthropic has published an updated version of Claude’s constitution, a document that sets out how the AI model is expected to behave. The revised text replaces a short list of principles with a more detailed framework, instructing Claude to prioritise four core values: being broadly safe, broadly ethical, compliant with Anthropic’s guidelines, and genuinely helpful—in that order. It expands on ethical decision-making and introduces clearer limits on prohibited topics such as bioweapons. The document also addresses uncertainty around AI consciousness and moral status, touching on questions of Claude’s identity and well-being. The full 80-page document can be found here.

Introducing ChatGPT Go, now available worldwide
OpenAI is launching ChatGPT Go, its a low-cost ChatGPT subscription plan, worldwide. Priced at $8 per month in the US (with localised pricing elsewhere), the plan offers affordable access to GPT-5.2 Instant with higher usage limits, longer memory, and more image and file uploads than the free tier, sitting alongside the existing Plus and Pro subscriptions. ChatGPT Go will have ads.

Musk Seeks Up to $134 Billion Damages From OpenAI, Microsoft
Elon Musk is seeking between $79 billion and $134 billion in damages from OpenAI and Microsoft, claiming he was misled after the company moved away from its nonprofit roots and partnered with Microsoft. The case is set to go to a jury trial, with Musk arguing he deserves a share of OpenAI’s current value because of his early support. OpenAI and Microsoft deny the claims and say the lawsuit is an attempt to harass a competitor.

China’s Zhipu Unveils New AI Model Trained on Huawei’s Chips
Chinese AI startup Zhipu says its latest model, GLM-Image, is China’s first advanced multimodal AI system fully trained on AI chips made by Huawei, marking an important step for China’s domestic AI industry and its drive to reduce reliance on US technology.

DeepMind CEO Says Chinese AI Firms Are 6 Months Behind the West
Demis Hassabis said Chinese AI companies are about six months behind the top Western labs, arguing that while they are increasingly effective at catching up, they have yet to demonstrate innovation beyond the technological frontier. He said the strong reaction to DeepSeek’s R1 model was exaggerated, calling it impressive but not truly breakthrough. Additionally, Hassabis noted that US restrictions on advanced chips have limited China’s progress, even as some of those barriers are now starting to ease.

The Messy Human Drama That Dealt a Blow to One of AI’s Hottest Startups
The Wall Street Journal takes us behind the scenes at Thinking Machines Labs and its latest drama, which resulted in three of its founders leaving the company and rejoining OpenAI. It is a story of internal disagreements, strained working relationships and competing ideas about the company’s direction, which ultimately led to a major leadership break and shows how fierce the battle for AI talent has become.

OpenAI’s Altman Meets Mideast Investors for $50 Billion Round
Bloomberg reports that Sam Altman has been meeting with major Middle Eastern investors to secure funding for a new OpenAI investment round that could raise $50 billion or more. The talks are still early and could value the company at up to $830 billion. The funding would help OpenAI cover huge AI infrastructure costs as it faces growing competition from rivals such as Google and Anthropic.

Meta’s new AI team delivered first key models internally this month, CTO says
Meta’s new AI lab has produced its first internal models, which the company’s CTO Andrew Bosworth said show strong promise, though they are still unfinished. Speaking in Davos, he explained that the work is still in its early stages, and more development is needed before the public can use the technology.

Elon Musk’s xAI activates world’s first gigawatt-scale AI training cluster
xAI has switched on its new Colossus 2 supercomputer, making it the first AI system in the world to run at a gigawatt scale, with plans to grow even larger soon. The system is used to train the company’s Grok AI models and was built very quickly compared with rivals.

Apple is Fighting for TSMC Capacity as Nvidia Takes Center Stage
Tim Culpan reports that TSMC’s centre of gravity is moving away from Apple towards AI chip customers like Nvidia, as demand for artificial intelligence hardware grows much faster than smartphones. This shift means Apple no longer has guaranteed priority at TSMC and now has to compete for manufacturing capacity.

Introducing Community Benchmarks on Kaggle
Kaggle has launched Community Benchmarks, a new feature that allows the global AI community to create, run and share custom model evaluations that better reflect real-world use cases. Instead of relying on simple accuracy scores, it supports more realistic tests to measure model performance in tasks such as reasoning, coding and tool use, with the main goal of helping developers better understand how models perform in real-world situations.

FLUX.2 [klein]: Towards Interactive Visual Intelligence
FLUX.2 [klein] is a new family of open image models from Black Forest Labs designed for fast, high-quality image generation and editing. The models combine text-to-image and image editing in one system, run in under a second on consumer hardware, and are built for real-time creative and developer workflows without sacrificing visual quality.

Qwen3-TTS
Qwen3-TTS is a series of new open text-to-speech and voice cloning models developed by the Alibaba Cloud AI team. According to the Qwen team, the new model supports stable, expressive, and streaming speech generation, free-form voice design, and vivid voice cloning. Qwen3-TTS comes in two variants—0.6B and 1.7B—and is available on GitHub and on HuggingFace.

Introducing Furiosa NXT RNGD Server: Efficient AI inference at data center scale
Furiosa, a South Korean chip manufacturer which Meta tried to acquire and failed to do so, announced the launch of its first branded AI inference system, the NXT RNGD Server, designed to run modern AI workloads efficiently inside existing data centres. The company said the server is easy to deploy, uses less power than typical GPU systems, and has already been tested by enterprise customers, including LG AI Research.

YouTube will soon let creators make Shorts with their own AI likeness
YouTube will soon allow creators to use AI versions of themselves in Shorts. The platform is expanding its AI tools while giving creators more control over how their likeness is used and protecting against misuse. YouTube is also working to reduce low-quality AI content and plans to add new Shorts features, including image posts.

The AI data center deals that no one can verify
We hear almost every other week about new multi-billion-dollar AI infrastructure deals. This article examines how concrete those announcements really are. It explains that, unlike traditional infrastructure, these deals lack clear definitions, standard measurements, and transparent contracts. As a result, markets often treat headline figures as firm commitments, even though many may represent uncertain or optional plans rather than guaranteed spending.

🤖 Robotics

Your First Humanoid Robot Coworker Will Probably Be Chinese
This article gives us a glimpse of the Chinese robotics industry by exploring how fast China is developing humanoid robots, from lively demonstrations at a major AI conference in Shanghai to advanced research labs in Beijing. It highlights how companies like Unitree are outpacing Western rivals through low costs, fast iteration, and tightly integrated manufacturing, while government-backed research groups push toward a “ChatGPT moment” for robots.

▶️ What It Really Takes to Deploy Humanoid Robots at Scale | Agility Robotics CTO (27:18)

Pras Velagapudi, CTO of Agility Robotics, shares what it takes to bring humanoid robots from research labs into production environments. He shares hard-earned lessons from deploying Digit, Agility’s humanoid robot, in live logistics and manufacturing facilities. He also explores why “cool videos” are only the tip of the iceberg, and what actually matters when customers are measuring ROI, reliability, safety, and performance at scale.

Airbus to test China’s battery-swapping humanoid robots in aircraft assembly
UBTech, a Chinese robotics company, has signed a deal with Airbus to supply its Walker S2 humanoid robots for use in aircraft manufacturing. The partnership will test the robots in complex factory environments and marks an important step in UBTech’s expansion outside China.

Miami, Your Waymo Ride Is Ready
Waymo has begun offering its fully autonomous ride-hailing service to the public in Miami, with early access being rolled out to nearly 10,000 signed-up residents. The service is available in neighbourhoods such as Wynwood, Brickell and Coral Gables, with plans to expand to Miami International Airport. Miami becomes the sixth city in the US where Waymo offers commercial robotaxi service, with many more planned to launch soon.

Tesla discontinues Autopilot in bid to boost adoption of its Full Self-Driving software
Tesla has discontinued its Autopilot driver-assistance system as it pushes customers towards its more advanced Full Self-Driving (Supervised) software, which will now be offered only through a monthly subscription. The change follows a California ruling that found Tesla had deceptively marketed its driver-assistance features. New vehicles will include only basic cruise control.

Elon Musk: Tesla will sell humanoid robots by end of 2027
Elon Musk said Tesla aims begin selling its Optimus humanoid robots to the public by the end of 2027, once they are safe and reliable enough. The robots are already doing simple jobs in Tesla factories, and Musk believes they could eventually help with many everyday tasks. However, major technical challenges remain, and Musk warned that production will begin slowly because most of the technology is new.

▶️ Humanoid x Siemens | Testing Humanoid Robots in Industrial Logistics (3:07)

Humanoid, a London-based humanoid robotics startup, shows in this video how its HMND 01 humanoid robots are being tested at a Siemens factory. The robots autonomously picked, transported, and placed totes in a live production environment during a two-week on-site deployment at the factory. The company says that HMND 01 has successfully completed proof of concept in industrial logistics.

How Human Should Your Humanoid Be?
How “human” should a humanoid robot be? In this post, Chris Paxton compares the pros and cons of human-like robots with more unusual, non-human designs. Using Boston Dynamics’ new Atlas as an example, he explains that while human shapes can help with training and usability, they are not always the most efficient choice. The article suggests that future humanoid robots may keep the benefits of legs and arms without needing to look like humans at all.

Robot learns to lip sync by watching YouTube
Researchers have developed a robot that can learn realistic lip movements for speech and singing by observing itself and humans. Using a flexible face with many small motors and AI trained on mirror experiments and YouTube videos, the robot can match lip movements to sounds in different languages. While the movements are not perfect yet, the work brings robots closer to looking and communicating more naturally with people.

🧬 Biotechnology

Operating on DNA is more like surgery than medicine
New gene-editing tools could allow doctors to fix the genetic causes of rare diseases, but current rules make these treatments too slow and expensive to develop for very small numbers of patients. Instead of treating each genetic fix like a new drug, the article suggests regulating genetic treatment more like surgery. By approving expert centres, proven tools and safe processes, this approach could make personalised genetic care possible for patients who currently have no treatment options.

Thanks for reading. If you enjoyed this post, please click the ❤️ button and share it!

Humanity Redefined sheds light on the bleeding edge of technology and how advancements in AI, robotics, and biotech can usher in abundance, expand humanity's horizons, and redefine what it means to be human.

A big thank you to my paid subscribers, to my Patrons: whmr, Florian, dux, Eric, Preppikoma and Andrew, and to everyone who supports my work on Ko-Fi. Thank you for the support!

My DMs are open to all subscribers. Feel free to drop me a message, share feedback, or just say "hi!"