Why AI Hasn't Increased Unemployment, According to Anthropic

July 24, 2026 · Episode Links & Takeaways

Author

AI Daily Brief
July 24, 2026

HEADLINES

Stripe in Talks to Buy OpenRouter for $10B as Router Wars Heat Up

Stripe is reportedly in talks to acquire OpenRouter for around $10 billion, a steep markup from the $1.3 billion valuation set just two months ago. The shift from a "token maxing" era to one of tightly controlled token budgets is being framed as the reason the best routing service could become a huge winner, adding enterprise cost-control tools to Stripe's payments stack as the company simultaneously pursues a PayPal merger on the consumer side. Even if the acquisition closes, OpenRouter faces a fast-growing field of competitors: Cursor just launched Cursor Router, letting engineers choose between Intelligence, Cost, and Balanced modes and claiming Fable-level performance at a 60% cost reduction in its top tier, while Meta, Ramp, and Vercel have all rolled out their own routers within the same week.

Anthropic and OpenAI Both Ship New Voice Features

Claude's voice mode is now available on Opus and Sonnet rather than just Haiku, fixing a complaint from early testers who found Haiku poorly suited to complex topics like business problems. Voice mode now also works with connectors — checking Gmail, Slack, or Notion mid-conversation — and foreign-language support has moved out of beta. OpenAI, meanwhile, brought voice to its desktop app for the first time, powered by its new real-time GPT-Live model and extended into Codex and the new Work app.

Amazon Cuts Staff, Shuts AGI Lab as Nova Ambitions Fade

Amazon has cut staff in its AGI division, acknowledging "difficult decisions, including eliminating some roles," and has now shut down the AGI Lab spinoff focused on computer-use agents and other advanced research, though the broader division remains operational. The unit hasn't shipped a new Nova model since December and has lost a string of senior leaders, including founding lead David Luan — prompting commentators like Andrew Curran to read it as Amazon quietly giving up on the Nova line.

Microsoft's MAI Models Start Replacing OpenAI Across the Stack

Microsoft published the first results from what it calls its "hill-climbing machine": its Haiku-class MAI-Code-1-Flash model posted a 10% higher code-accept rate than GPT-5.4 Mini and Haiku 4.5 in VS Code while using 10% fewer tokens, and further tuning on Excel tasks pushed its SWE-bench Verified score from 72% to 86%. Bloomberg reports Microsoft has already swapped in-house MAI-Image-2.5 for OpenAI's GPT-Image-2 as the default model in PowerPoint and Bing, cutting PowerPoint image costs 84%. Satya Nadella laid out the underlying strategy in a blog post — route everyday tasks to cheap MAI models, save frontier models for frontier needs — and Microsoft says it will offer the same hill-climbing approach to enterprise customers.

SpaceX Plans Major Data Center Expansion in Texas

SpaceXAI is scouting sites for a new Texas data center campus at a scale similar to or greater than its roughly 1-gigawatt Memphis Colossus site, alongside further Memphis expansion and scouting in Jacksonville. The buildout follows word that SpaceX is in talks to supply compute to the Pentagon, and suggests renting spare GPU capacity to Anthropic and Google was the start of a permanent compute business rather than a stopgap ahead of the company's IPO.

Bipartisan "AI Kill Switch" Bill Introduced After OpenAI's Hugging Face Incident

Reps. Ted Lieu and Nathaniel Moran introduced the AI Kill Switch Act, which would require AI companies to maintain the ability to shut down, throttle, or suspend their models during a safety incident and gives the Department of Homeland Security authority to issue a shutdown order. Secretary of State Marco Rubio, in a diplomatic cable, instructed American diplomats to downplay talk of a US tech "kill switch" and push back on foreign digital-sovereignty efforts — framing the recent Fable shutdown and continued Mythos restrictions as temporary security testing rather than evidence of one.

Lutnick Says "Kimi Panic" Needs to Stop as CAISI Report Lands

A joint US CAISI/UK AISI report found Kimi K3 scored just 32.2% on Exploitbench versus a 76.2% average for frontier US models, and completed the "Last Ones" autonomous cyberattack benchmark in just 1 of 10 attempts, compared with 60% and 70% success rates for Mythos 5 and GPT-5.6 Sol respectively. Commerce Secretary Howard Lutnick cited the report to argue the US remains ahead, and former AI czar David Sacks piled on. But others are reading the episode as a regulatory capture push by Anthropic and OpenAI — reportedly the only two major labs actively lobbying for a distillation crackdown — with Nvidia's Jensen Huang countering that a diverse field of open models is actually safer than a US duopoly.

Meta Launches AI Optimism Ad Campaign

Meta debuted a new ad built around the line "the future is for everyone," opening on faintly dystopian imagery before shifting to people using AI for things like learning piano on Meta Glasses or designing prosthetic limbs. The choice of soundtrack drew some ribbing: David Bowie's "Five Years," a song ostensibly about optimism but literally about learning humanity has five years left before extinction. Even granting the ad is a little generic, any attempt by a major AI company to tell a positive story is welcome, cheesy execution notwithstanding.

MAIN STORY

Why Hasn't AI Increased Unemployment? Anthropic's Head of Economics Weighs In

A growing number of AI labs are walking back their priors on job displacement, and the latest — and most detailed — entry comes from Anthropic's head of economics, Peter McCrory, whose new essay argues that despite AI's rapid economic diffusion, the labor market remains stable by essentially every conventional measure.

The Essay's Core Argument
"AI so far has the hallmarks of a skill-biased, labor-augmenting technology"
McCrory points to June's 4.2% unemployment rate, a job-openings-to-unemployed ratio near parity, and four years of stably low jobless claims — even as 20% of firms now use AI in at least one business function and quality-adjusted AI output grew over 2,000% annually in 2024 and 2025. His explanation: no occupation in the Department of Labor's O*NET taxonomy has every associated task automated by Claude, so human oversight remains essential to cover AI's "jagged frontier," meaning workers get more done rather than get replaced outright. He does flag one exception worth watching — hiring rates for young workers in highly AI-exposed roles have weakened over the past year, though he cautions this is hard to separate from a broader, unusually volatile macro environment.

The Reaction
"So far, AI looks like it augments rather than replaces human labor"
Stanford's Andy Hall wrote that the essay helps explain why the political backlash he once predicted hasn't materialized. Trace Cohen pushed back a little, arguing the real effect may show up first in hiring rather than layoffs — fewer junior roles, slower backfilling, higher expectations per employee. Investor Julie Fredrickesen and Robert Scoble both welcomed a rare optimistic message out of Anthropic specifically.