Anthropic modeled your job in 2030

Anthropic Just Modeled What Its Own AI Does to Your Job by 2030

On September 9, Anthropic's economics team published an interactive model of the US economy in 2030 under three AI scenarios. The exercise is remarkable for a simple reason: a frontier AI lab ran the numbers on its own technology's impact on workers — and published what it found, including the ugly parts. In the most extreme scenario, GDP grows by 32.4% and knowledge-worker wages fall by more than 10%. Most Americans surveyed expect something in the middle: the economy grows meaningfully, and knowledge-worker pay stays flat. Growth and wages decouple. The economy gets richer. Workers don't.

Key Points:

1. The three scenarios, plainly stated — The modest scenario: AI adds 1.6% to GDP, disruption is limited, wages hold. The substantial scenario: AI performs about half of knowledge work autonomously, GDP grows 8.3%, knowledge-worker wages stay flat, but workers in non-cognitive jobs see gains. The extreme scenario: AI is more productive than humans at nearly all knowledge-work tasks and performs nearly all of them autonomously — GDP hits $44.4 trillion, unemployment reaches 11.9%, cognitive unemployment reaches 17.9%, and labor's share of national income drops from 60% to 45%. Capital's share of income ends up roughly 81% higher. The economy grows enormously. The gains go almost entirely to whoever owns the machines.

2. The scenario most people expect — Anthropic surveyed 10,980 US adults. The median respondent's expectations map closely to the substantial scenario: about 8–10% GDP growth by 2030 and unemployment around 5%. That's not the apocalypse. It's a slow squeeze — flat wages against rising productivity, where the value AI creates accrues somewhere other than your salary. The key detail from the substantial scenario: non-cognitive workers (electricians, plumbers, healthcare aides) see wage gains. Knowledge workers don't.

3. What Anthropic is doing with this — The company was explicit that these are modeled scenarios, not forecasts. But the Econ Scenario Explorer is interactive — users can adjust assumptions about AI capability, adoption speed, and autonomy, and watch the economic effects change in real time. Anthropic says it plans to use these findings to inform policy proposals aimed at ensuring AI's economic benefits are broadly shared. A company modeling its own product as a wage suppressor and publishing it publicly is doing something unusual. Whether anything follows from it is another question.

Big question: If the scenario most people expect is one where the economy grows and wages don't — who's responsible for closing that gap, and does anyone with the power to act on it actually want to?

Businesses Are Quietly Downgrading From Frontier AI to Save Money

Ramp's monthly AI Index — which tracks real AI spending across thousands of businesses — came out this week with a number that should make the frontier labs nervous: usage of their most powerful models is falling. Frontier models like GPT-6 Astra, Claude Fable, and Opus drove 45% of token share in August, down from a peak of 53%. The models picking up that slack are cheaper mid-tier options like Claude Sonnet and GPT-5.6 Terra. Businesses are voting with their budgets, and right now, they're voting against the top of the line.

Key Points:

1. The numbers behind the slide — Per-employee AI spend dropped nearly 10% in August, falling from $7,976 to $7,205 among the top 1% of AI-adopting firms. Overall AI adoption growth has slowed to less than half a percentage point month-over-month. The effective price per million tokens has dropped 41% since March, down to $0.68 — while frontier models like Astra and Fable still cost $10 per million input tokens. That's a nearly 15x gap between what the market is paying on average and what the best models cost.

2. The explanation companies are giving — "Standard models are still highly performant and also more cost effective," according to Ramp's chief economist. Companies are imposing company-wide defaults that dial back frontier model access. Most day-to-day work — email, summaries, research, drafting — simply doesn't need the most powerful model available. The models that perform best on hard benchmarks aren't the ones most businesses reach for when the bill arrives.

3. What this means for the labs — OpenAI and Anthropic are spending enormously to train frontier models that, by their own admission, cost far more than most use cases require. Meanwhile, open-source alternatives are growing (though still limited to 6.4% of AI-spending businesses). The competitive pressure isn't just from each other — it's from the growing reality that "good enough" is genuinely good enough for most tasks, and good enough keeps getting cheaper.

Big question: If the market is consistently choosing cheaper models for real work, are frontier labs building for the customers they have — or the use cases they hope will eventually justify the cost?

Meta's Giant AI Bet Had a Quiet Launch. The Timing Didn't Help.

Meta launched Muse this week — its personal AI agent app, designed to understand your goals and work on things for you 24/7, with its own isolated Linux computer, a browser, memory, and a "Sentinel" agent watching everything it does. It's Meta's biggest consumer AI bet since the metaverse pivot. It hit number two on the US App Store. And by the numbers that actually matter, it had a quieter debut than almost any major AI launch in recent memory.

Key Points:

1. The download gap — Muse pulled 83,000 US iOS downloads in two days, which sounds impressive until you stack it against comparisons: Threads hit 4.3 million US downloads on launch day. The Meta AI app hit 108,000 downloads on its debut. ChatGPT topped half a million US installs in under a week. Muse took twice as long as ChatGPT to reach that figure. The Android side is worse — sitting at #338 in Google Play's Productivity category.

2. The timing problem — Muse launched days after Meta agreed to an $18 billion multistate settlement over social media harms to children. The app asks users to hand Meta significantly more personal information than a typical social app — goals, credentials, daily context — and to let a Meta-run agent act on their behalf around the clock. Asking for deeper access to people's lives immediately after settling the largest child safety case in social media history is a difficult ask, and users appear to be answering cautiously.

3. The competition is sharper than expected — Everyone has an agent app now, but the most interesting competition isn't other AI labs — it's Instinct, a text-message-based AI agent recently valued at $2.5 billion. This week alone, Instinct gave every user their own email address and announced that one person's agent can now coordinate directly with their friends' agents to plan shared activities. If agents are eventually as social as social media, a network built on who people actually text might matter more than Meta's friend graph.

Big question: Meta built the world's largest social network by giving people a reason to share more of themselves — but is there a version of that playbook that still works when people have learned to be suspicious of what happens to what they share?

OpenAI Stopped Taking $200/Month Subscribers. Astra Broke Their Servers.

Three days after GPT-6 Astra launched, OpenAI did something companies almost never do: it stopped accepting new customers. The $200/month Pro plan — OpenAI's highest tier — is now paused for new sign-ups, because demand for Astra is so intense that the company can't serve new Pro users without degrading the experience for existing ones. It's a strange kind of problem to have. It's also a revealing one.

Key Points:

1. What happened — OpenAI product lead Thibault Sottiaux announced the pause, describing demand for Astra as "genuinely unprecedented" and saying he'd "not seen anything like it" despite having been through very steep growth before. The company raised usage limits for Codex users just last month — and still ran out of runway. The Pro tier carries the most system load because it's the plan with the fewest restrictions on the most powerful model.

2. The AGI framing drove it — OpenAI didn't just launch Astra as a new model. President Greg Brockman called it a "generational leap" and suggested it could be seen as the arrival of AGI. Sam Altman said it was a "new capability level" that had changed his own workflows. That framing — intentional or not — created a wave of demand that the infrastructure couldn't absorb. The hype served its marketing purpose and then immediately created an operational problem.

3. What it signals — A company turning away customers at $200/month is leaving real money on the table. It also confirms that the ceiling on what people will pay for genuinely better AI is higher than most pricing models assumed. OpenAI hasn't said how long the pause will last — which itself signals they don't fully know how long it takes to build enough capacity to serve what they've already promised.

Big question: If OpenAI can't serve the demand its own marketing created for its most expensive product — is that a capacity problem, a pricing problem, or a sign that the gap between what frontier AI promises and what it can actually deliver is still wider than the launch events suggest?