Read the latest
Thursday, September 3, 2026HotTea archive editionVerified 2:37 AM PDT

minutes. Facts before narrative.

Meta cuts the cost of frontier AI tasks

Artificial Analysis put Muse Spark 1.3 in its top tier and measured a lower cost per task. Courts, schools, companies, and labs moved on AI too.

Start the briefingSee how we verify →

Published daily by 3:00 AM Pacific. No forced optimism. No manufactured panic.

Listen to today’s briefing

Meta cuts the cost of frontier AI tasks

The sourced HotTea edition, condensed into a chaptered morning podcast with verified audio and a full transcript.

Follow on Apple PodcastsFollow on SpotifyPodcast home →
01

The lead

The LeadHot

Meta cuts the cost of frontier AI tasks

Meta built Muse Spark 1.3 for long agent and coding jobs. Artificial Analysis gave it a frontier-tier score and measured lower task costs than nearby rivals.

What happened

Meta released Muse Spark 1.3 on Wednesday in Muse Code and the Meta Model API. Meta says the model handles longer agent workflows and follows complex instructions more reliably. Meta engineers also measured about 20 percent fewer tool calls and 25 percent fewer tokens than Muse Spark 1.2. Artificial Analysis gave the available xhigh version a 61 on its Intelligence Index. That put it four points above Muse Spark 1.2 and tied it with several frontier systems. At Meta's unchanged API price, Artificial Analysis measured a $0.55 cost per index task. No other model scoring 59 or higher cost less. The limited-preview max version scored 62, one point higher, but Meta has not published its price.

Why it matters

Meta did not beat the benchmark. It reached the frontier tier at a lower measured cost. Agent products rack up cost when a model reads context, works through a task, calls a tool, and checks its work. A lower bill lets more jobs run in the background. It also makes rivals defend price, not only capability.

What to watch

Watch whether the max version keeps its score after a wider release. Watch what Meta charges for it, and whether Meta ships the promised open-weights release. Real agent completion rates matter too. Meta reports fewer tool calls and tokens. Product teams still need to test that efficiency on their own work.

The caveat

Meta made the usability and safety claims. Artificial Analysis measured benchmark performance and task cost independently. A benchmark still does not prove that a model will hold up inside a specific business workflow.

Meta AI: Introducing Muse Spark 1.3 Artificial Analysis: Muse Spark 1.3: Meta reaches the frontier
Read this story on its own →

Worth knowing

The rest of the morning

Facts, pressure point, next evidence.

02
MoneyHot

Nscale tells investors it has $103 billion under contract

Reuters reported, citing The Information's account, that British AI infrastructure company Nscale showed investors a pitch for a possible 2027 public offering. The pitch forecasts about $2 billion in revenue next year and $15 billion in 2029. It says Nscale has $103 billion in contracted revenue with customers including Microsoft, OpenAI, and Anthropic. That number is a long-term contract claim, not collected sales.

Pressure point Nscale did not respond to Reuters's request for comment. One publication described a private investor pitch. The figures are still a company forecast, not independently verified revenue.

Watch Watch for audited accounts, contract lengths, and the capital Nscale must raise before an IPO. Huge compute commitments matter only if Nscale can pay for and build the data centers behind them.

Reuters ↗
Read article →
03
WorkHot

Uber plans to cut 3,300 jobs as robotaxi rivals close in

Bloomberg reported that Uber plans to eliminate about 3,300 corporate jobs. Reuters and The Guardian cited that report. The cuts equal roughly 10 percent of Uber's global workforce. An internal memo called the move a reset to remove layers and speed decisions. The cuts are Uber's largest since 2023. Uber is also spending to compete in autonomous driving and delivery.

Pressure point Uber did not immediately comment to Reuters. The memo tied the cuts to management structure and speed, not directly to AI. Robotaxi competition is business context here, not Uber's stated reason for the layoffs.

Watch Watch which teams lose staff and whether Uber shifts more spending toward autonomous vehicle partnerships. The sharper test is whether fewer management layers help Uber move faster, or leave it with less capacity.

Reuters ↗The Guardian ↗
Read article →
04
PolicyHot

Justice Department backs OpenAI on fair use

The Justice Department urged a New York federal judge to accept OpenAI's fair-use argument. OpenAI argues that training models on copyrighted news articles is fair use. In a September 1 statement of interest, the department addressed the publishers' consolidated case. The department said a ruling against OpenAI would raise licensing costs and favor the largest technology companies. It also said such a ruling would weaken the United States against foreign competitors. The filing supports OpenAI's legal theory. It does not decide the case.

Pressure point The New York Times said the administration is siding with large AI companies at creators' expense. The Times argued that model developers should pay for protected work. Judge Sidney Stein still has to apply the four-part fair-use test to the facts before him.

Watch Watch whether Judge Sidney Stein treats the government's innovation and national-security arguments as part of fair use. A ruling on model training would reach beyond this lawsuit. Publishers and AI companies are still negotiating licenses while the legal baseline remains unsettled.

Bloomberg Law ↗Reuters via Ledger-Enquirer ↗
Read article →
05
SocietyWarm

New York City bars younger students from generative AI

New York City barred public school students below high school from using generative AI during the school day. High-school students can use only tools their schools provide. The city also capped device-based instruction at one hour a day for kindergarten through eighth grade. It extended its smartphone ban across elementary and middle schools. Officials said the rules are meant to protect independent thinking and human connection.

Pressure point The policy draws a clear age line, but a huge school system still has to enforce it. Teachers need useful non-AI software, enforcement rules, and a way to separate banned generation from ordinary search, accessibility tools, and tutoring tools.

Watch Watch the implementation guidance and the high-school tool list. New York still has to show whether it can teach students how these systems work while sharply limiting their use in class.

City of New York ↗The Verge ↗
Read article →
06
PowerWarm

Anthropic and Commerce restart talks after Pentagon dispute

Commerce Secretary Howard Lutnick said the Trump administration is in active talks with Anthropic. He said the administration wants the company back on the government's right side after a public dispute over military uses of its models. Lutnick said Anthropic has worked closely with officials on cyber tools. He contrasted Anthropic with OpenAI, which he said gave the government what it asked for. The comments suggest a thaw, not a settlement.

Pressure point Lutnick's remarks and the reports do not show what Anthropic has agreed to change. The core dispute involved limits on autonomous weapons and domestic surveillance. That makes the missing contract language more important than the warmer tone.

Watch Watch for a revised federal contract, changes to Anthropic's usage rules, and any formal removal of restrictions tied to the Pentagon dispute. A deal would show how much leverage the government has over model safety policies.

Reuters ↗Axios ↗
Read article →
07
ScienceWarm

AI-designed proteins delivered RNA in animal tests

Researchers built more than 100 synthetic RNA carriers from natural functional domains and protein assemblies designed by generative models. Their best design, STV-C8, transferred RNA more efficiently than the virus-like particles and lipid nanoparticles tested in cell culture. The team tracked delivery in mice. The team also used the carrier to remove a disease-related section of the dystrophin gene in pig muscle cells.

Pressure point This is early laboratory and animal work, not a human treatment. After intravenous dosing, the carrier delivered RNA mainly to mouse lungs. The team still has to improve targeting and study how the carrier moves through the body.

Watch Watch replication, immune-response testing, and control over which tissues receive the cargo. The result gives researchers another way to build delivery vehicles. Clinical value depends on precise targeting and repeatable safety.

Nature ↗Helmholtz Munich ↗
Read article →

The whole AI power map

AI is no longer a tech beat.

HotTea follows where AI moves power, money, labor, security, and state capacity—not only where a new model scores higher.

01

Politics & regulation

Elections, procurement, courts, surveillance, lobbying, and state power.

02

Economics & labor

Productivity, wages, employment, capital spending, concentration, and who captures the gains.

03

War & security

Autonomy, cyber operations, intelligence, targeting, export controls, and escalation risk.

04

AI geopolitics

Chips, energy, alliances, sovereign capability, supply chains, and strategic competition.

05

Markets & companies

Funding, revenue, margins, model economics, enterprise adoption, and infrastructure bets.

06

Science & society

Medicine, education, climate, culture, research, rights, and measurable public outcomes.

The Deep Read

Agent economics matter more than benchmark wins

Muse Spark 1.3 shows why the next model fight is about the cost of finishing work. Cheaper tokens help. A cheaper completed task changes products.

1

Long loops multiply every price difference

An agent reads files, calls tools, checks results and often retries. Small savings compound across each step, especially when the workflow runs all day.

2

Efficiency must preserve completion

Fewer tool calls and tokens matter only when the system still finishes the job. Teams need task-level tests that count correct outcomes, total spend and human rescue time together.

3

Low cost widens the automation surface

When a capable model gets cheaper, it moves into monitoring, reconciliation and background research that were too expensive to run continuously. That expands the market before any new benchmark record does.

The watchlist

Signals that could change the read

Meta publishes the price and broad availability for the 62-point model.
Judge Stein rules on fair use or gives the DOJ filing weight in a substantive order.
A contract or policy document shows whether military-use limits changed.
Audited numbers, new capital or an IPO filing tests the $103 billion contract pitch.

How HotTea works

No optimism quota. No negativity quota. Just the sourced read.

Every reported item links to its source. Company claims remain company claims. High-risk stories require stronger corroboration. Material caveats, conflicts, and unknowns stay in the story. HotTea’s interpretation is visibly separated so readers can disagree without losing the facts.

Edition validated · 7 stories · 13 unique sources

Audit today’s sources →

Tomorrow’s signal, before tomorrow’s noise

Open HotTea. Know what changed.

A new verified edition every morning. If the evidence or release gate fails, the last verified briefing stays live.

Back to today’s top ↑