The price of running AI on long documents and multi-step tasks fell sharply this week, and the mechanism behind the cut is worth understanding because it rewards a specific way of working. Elsewhere: Uber eliminated a tenth of its workforce and said the money is going to driverless cars, Waymo and Tesla spent the weekend arguing about what makes a self-driving car safe two days before Tesla's launch event, an AI reader caught fifteen liver cancers that radiologists had missed, and three of the largest companies in America put a quarter of a billion dollars into training electricians and welders.
AI got 25% to 45% cheaper, if you reuse your context
Anthropic released Claude Fable 5.1 and Claude Mythos 5.1. The headline number is not the model quality — it is the bill. Anthropic cut the price of cached context by 75%, from $1.00 to $0.25 per million tokens. Everything else stayed put: $10 per million input tokens and $50 per million output, with batch processing at half those rates. The company says the net effect is roughly 25% lower cost for typical workloads and as much as 45% for heavily agentic ones, where the same background material gets read over and over.
That distinction is the whole story. Caching means the system stores the material you already sent — a policy manual, a codebase, a customer history — so it does not have to be re-processed every time you ask a new question about it. Cached content now costs 2.5% of the normal input price. If your AI work involves asking many questions against the same large body of reference material, your costs just dropped substantially. If you send a fresh, small prompt every time, they did not move at all. The pricing change quietly favors businesses that have organized their information into something stable enough to reuse.
The capability gains are concentrated in long, tool-using tasks. On Terminal-Bench 4.0, which measures whether a model can carry out multi-step work in a command line, Fable 5.1 scored 55.8% against Fable 5's 42.0%. On Terminal-Bench-Science 0.1 the jump was larger, 52.6% from 24.7%, and on AutomationBench it went from 17.1% to 31.4%. Both models handle up to a million tokens of context. Fable 5.1 adds mid-conversation effort control, letting an operator dial how much work the model puts into a given step.
Mythos 5.1 shares the same underlying architecture but ships with fewer restrictions, and is available only to vetted organizations in cybersecurity and life sciences through a restricted-access program. Alongside the models, Anthropic introduced Enterprise Frontier Safeguards, which lets a company keep its own monitoring data inside its own AWS, Azure or Google Cloud environment under its own encryption keys, at no additional charge — a direct answer to the objection that safety monitoring means handing a vendor your logs.
Uber cuts 3,300 jobs and names the reason
Uber is laying off 3,300 people, about 10% of its global workforce and its largest reduction since 2020. CEO Dara Khosrowshahi framed it as a structural problem rather than a financial one: the company had "built new products, expanded into new businesses," but that growth "brought complexity: more layers, more coordination, more fragmented ownership."
The specifics are unusually blunt for a restructuring memo. The number of managers drops by roughly 20%, with some moving into individual-contributor roles. Teams consisting of only one or two people are being cut by half. No employee will sit more than seven layers from the CEO. Engineering, science and delivery divisions are consolidating, and remote work is being restricted to under 1% of staff. The stated destination for the freed-up spending is ridesharing, delivery and the robotaxi division.
What makes this notable is not the layoff — it is the sequencing. Uber is not cutting because driverless cars replaced its workers; its drivers are contractors, not the people being let go. It is cutting corporate overhead now in order to fund an automation bet that has not paid off yet. That pattern, thinning the middle of the organization to finance a capital-intensive technology transition, is the one worth watching, because it does not require the technology to work before the jobs go.
The model race narrows around coding and cost
Google is preparing to release Gemini 3.8 Flash, and internal testers reportedly prefer it to Anthropic's Opus model for coding work. Flash is Google's smaller, cheaper, faster tier, which makes the comparison significant — a budget model matching a frontier one on a core commercial task changes the buying decision. Google has meanwhile fallen months behind on its Pro series, repeatedly scrapping internal candidates that were not meaningfully better than Flash.
Google also shipped agentic video understanding in Gemini, which it says cuts token consumption by up to 88% and costs by up to 66% while improving quality by as much as 7%. Rather than processing every frame of a video, the model decides which portions to actually examine. For anyone processing security footage, recorded meetings or inspection video, the economics of that work changed more than any model announcement this week.
Separately, World Labs unveiled Atlas, an omni world model for spatial intelligence that works natively across text, images, video and 3D data, generating camera-controlled video up to a minute long at 1440p and reconstructing 3D scenes from as few as one to three input images. It is in early access with selected partners, with no published pricing.
An AI reader found cancers the radiologists missed
A study published in Nature Medicine on August 19 deployed an AI system called LiON, for Liver DiagnOsis Network, as an additional reader alongside radiologists working through more than 10,000 real patient CT scans during routine hospital care. The system flagged 51 liver lesions that had been overlooked in the initial reports. Fifteen of those turned out to be cancer.
The work was led by Shengjing Hospital of China Medical University and hospitals affiliated with Zhejiang University, in collaboration with Alibaba's DAMO Academy, King's College London, and the French research institute EURECOM. The design detail that matters is the phrase "additional reader" — LiON did not replace anyone. It reviewed scans that human radiologists had already read and surfaced candidates for a second look. That is a materially different proposition from autonomous diagnosis, both clinically and legally, and it is the deployment pattern most likely to survive regulatory scrutiny in other fields too.
Security, policy and the cost of agents that act
OpenAI designated its Astra model as reaching the "Critical" tier for cybersecurity capability under its own preparedness framework, and gated it at launch, citing the model's ability to identify previously unknown vulnerabilities and develop exploitation methods with minimal human guidance. A frontier lab restricting its own release on capability grounds is still rare enough to note, and it arrives alongside Anthropic's decision to put Mythos 5.1 behind vetted access.
The European Commission designated ChatGPT a Very Large Online Search Engine under the Digital Services Act, the first time a generative chatbot has been placed in that category. The trigger was scale: roughly 159 million average monthly users in the EU against a 45 million threshold. The classification brings obligations around risk assessment, transparency and researcher data access that were written for search engines and are now being applied to an assistant.
On the practical side, AIR Security raised $50 million in seed funding from Sequoia and Greenoaks to secure AI agent supply chains, and reports that roughly 27% of publicly available add-ons fail its security checks. Separately, Aesto Health confirmed a breach affecting 9.5 million individuals, stemming from unauthorized AWS access that occurred in December 2025 and was not confirmed until May 2026 — a five-month detection gap.
Power, chips and the people who build the buildings
Fervo signed an agreement to supply Google with nearly 400 megawatts of clean electricity from its geothermal project in southwest Utah, which would be the largest geothermal system in the world when completed in 2028. Fervo expects test power from a 33-megawatt unit in the fourth quarter, and the deal underwrites a second 400-megawatt phase. Nvidia put $125 million into Spanish startup iPronics, which builds photonic chips for AI networking, and Tencent-backed Enflame's Shanghai IPO was oversubscribed more than 6,000 times as it raised roughly $900 million for domestic alternatives to Nvidia silicon.
The most interesting infrastructure story is about labor. Meta, Google and BlackRock have committed more than $265 million combined to train skilled trade workers rather than AI engineers. Meta put $115 million into America's Workforce Academy, a five-week program with a daily stipend, free training, travel and housing, and a guaranteed job with one of Meta's construction partners for every graduate. Google committed $50 million through Google.org toward training more than 300,000 skilled trades workers across more than 20 states, and BlackRock's $100 million Future Builders initiative targets 50,000 workers over five years. In July, BlackRock, Google, Ford and Carhartt launched an Alliance for America's Skilled Trades to coordinate the effort. Every data center needs miles of high-voltage cable, industrial cooling and backup power, installed by people. The AI boom's most durable job creation may be in trades nobody files under "AI."
Physical AI
The robotaxi argument got loud this weekend, two days before it gets tested. Waymo went public with a direct critique of Tesla's approach ahead of Tesla's September 3 Cybercab launch event. Srikanth Thirumalai, VP of Waymo's driving software, argued that "cameras are incredible, but they aren't enough," and that safe autonomy at scale requires a mix of sensors. He characterized pure end-to-end neural systems — those converting raw pixels directly into steering commands — as risking "black box failures," noting that "even the best AI models with trillions of parameters still hallucinate." Waymo's position is backed by deployment: roughly 4,000 robotaxis across more than 14 U.S. cities, about 500,000 paid trips a week, and over 200 million real-world driverless miles. Tesla's Cybercab is a two-seater with no steering wheel or pedals, targeted at more than 125,000 units annually, and the company has been registering vehicles with the Texas DMV ahead of launch while running modified Model Y fleets in Texas and Florida, recently removing safety monitors from most of them.
The capital behind all of this reached a record. Robotics and physical AI startups raised $16.3 billion across 492 deals in Q1 2026, the strongest quarter on record, lifted by megadeals for Shield AI, Saronic and Neura Robotics. By late June the year-to-date total stood at $18.8 billion globally, already ahead of the $15 billion raised in all of 2025 and the $14.1 billion of 2021's venture peak. The largest individual rounds tell you where the conviction is: Saronic's $1.75 billion Series D in March, Neura Robotics' $1.4 billion Series C in June, Skild AI's $1.4 billion Series C in January, and Apptronik's $520 million Series A extension in February. Agility Robotics has agreed to merge with Churchill Capital Corp. XI, a special purpose acquisition company, which would make it the only U.S.-listed pure-play humanoid company with active commercial deployments.
Deployments remain narrow and industrial. GXO Logistics is running humanoids from Agility, Reflex Robotics and Apptronik, with Agility's bipedal Digit lifting containers onto conveyor belts at a GXO-operated facility. BMW has deployed Figure's robots at its Spartanburg, South Carolina plant and plans to move to the newer Figure 03. Meta acquired San Diego-based Assured Robot Intelligence, folding the team into its Superintelligence Labs unit to speed up training of a foundational physical AI model. The consistent shape here is a fixed station, a repeated motion, and a known environment — the same conclusion the honest humanoid reporting reached last week.
At the consumer end, IFA in Berlin has become the venue for stair-climbing robot vacuums, where brands including Dreame and Eufy solved multi-floor coverage by building a separate carrier unit that ferries the vacuum between levels rather than making the vacuum itself climb. Dreame's CyberX module uses treaded wheels and a triple braking system to handle steps up to 25 centimeters. It is a small engineering lesson with broad application: when the general-purpose version of a problem is too hard, the shipping product is usually the one that split the job in two.
Quick Takes
New York City barred student use of AI through eighth grade, one of the more restrictive district-level policies in the country and a signal of where the K-12 debate is heading.
John Deere launched "JD," an AI chatbot that helps farmers optimize equipment and operations — a mainstream industrial brand shipping a customer-facing assistant.
The Pentagon added Grok for Government to GenAI.mil, alongside the previously announced ChatGPT Mil, for more than three million military and civilian personnel.
Snap CEO Evan Spiegel does not expect the company's $2,195 smart glasses to reach consumers broadly until the end of the decade, an unusually candid timeline from a company shipping the product now.
Meta reported that only about a quarter of the output from its agentic testing loop was worth keeping — and that the approach works anyway, because failing code is discarded automatically.
Android Studio Quail ships with 23 curated skills for its agent mode, with support for custom ones, as Gemma 4 lands in the toolchain.
What This Means for Your Business
Restructure how you send information to AI before you shop for a cheaper vendor. The 75% cut to cached context pricing rewards a specific behavior: sending the same large reference material repeatedly and asking different questions against it. If your team pastes a fresh copy of a policy document, price list or customer record into a new chat every time, you are paying full input price on every request and capturing none of this. Consolidate your recurring reference material into stable documents, and ask whichever tool you use whether it supports prompt caching and whether your usage pattern actually hits the cache. That is a configuration conversation, not a procurement one, and it is where the 45% lives.
Read the Uber memo as an organizational template, not a tech story. The specific moves — cutting teams of one or two people by half, capping the org at seven layers from the top, reducing managers by 20% — are legible at any size. A twenty-person company has its own version of a one-person team that exists because someone's role drifted. The harder lesson is the sequencing: Uber is paying for an automation bet in advance, out of headcount, before the automation works. If you are planning something similar, be honest internally about which part is the efficiency and which part is the wager, because your team will work out the difference on their own.
Audit the agent extensions your team has already installed. AIR Security's finding that roughly 27% of publicly available add-ons fail basic security checks lands in the same week that a health company disclosed a breach it took five months to confirm. Connectors and plugins that let an AI tool touch your email, files or customer data are software supply chain, and most of them were installed by an enthusiastic employee rather than through any review. Make a list this week of every AI integration with access to a business system, note who approved it and what it can reach, and remove the ones nobody can justify. This is a one-afternoon task that gets much harder after an incident.
Copy the "additional reader" pattern from the liver cancer study. LiON's fifteen missed cancers came from reviewing work humans had already completed, not from replacing the humans. That structure — AI as a second pass over finished work, surfacing candidates for review rather than making the call — is the highest-value, lowest-risk deployment available to most businesses right now. Run it over invoices already approved, contracts already signed, quotes already sent, or support tickets already closed. You are not asking the AI to be right; you are asking it to be suspicious, and a human still decides.
Finally, if you employ or hire skilled trades, the competitive picture just changed. Meta, Google and BlackRock are collectively funding training pipelines for hundreds of thousands of electricians, welders and HVAC technicians, with Meta guaranteeing graduates jobs with its own construction partners. If you run a contracting, facilities or industrial services business, the tight labor market you have been navigating is about to get more competitive at the entry level, with well-capitalized employers offering stipends, housing and guaranteed placement. The upside is that these programs will produce trained workers who need somewhere to go after their first placement. Build a relationship with a regional program now rather than competing for its graduates later.