Horizon Lens — 8 November 2026
What METR could measure inside Anthropic
Benchmark results suggest Anthropic made about two years’ worth of historical model progress in the space of one year, according to The Verge’s November 7 account of an evaluation by METR. Researchers attributed most of the acceleration they saw by October to AI. They examined confidential data, interviewed employees and consulted informal staff polls. Much of the work they describe involved people using unreleased models to assist their research.
Employees using the most capable internal model estimated that they were around twelve times as productive at their own tasks as they would have been without AI. That number came from staff estimates. METR also said Anthropic removed much of the most informative evidence before publication. Cofounder Sam McCandlish cited commercial sensitivity and the risk of helping rivals accelerate.
Analysis
Connecting the two figures requires knowing which tasks sped up and how that affected the next model.
Action
When quoting a speedup, give its baseline, who measured it and which evidence readers can inspect.
OpenAI and Hacktron plan security tests for connected services
In a November 6 announcement, OpenAI said its engineers would work with Hacktron’s security researchers to choose high-priority tests of OpenAI systems. They plan to focus on connections with third-party services. Hacktron will receive wider access to frontier models through Daybreak Red. The relationship grew out of OpenAI’s existing bug bounty programme, through which Hacktron researchers had disclosed vulnerabilities to the company.
Hacktron builds agents for code review, penetration testing and exercises that simulate an adversary. According to OpenAI, its prior research found weaknesses where applications, cloud infrastructure and identity systems interact. OpenAI can identify important connections inside its own systems, while external researchers bring experience following the same action across multiple services. Findings from the new partnership will be a separate measure of its effectiveness.
Analysis
An agent that connects two services gives a security team a concrete question to investigate. Which identity grants each permission, and what can the second service do with the first service’s request?
Action
For a security review, choose one important integration and trace the permissions at each step. Confirm testing permissions with the programme that defines the scope.
Tesla plans to give its Cybertruck factory space to Cybercab
Tesla plans to end Cybertruck production in 2027 and use the factory space for its two-seat Cybercab robotaxi, Elon Musk said, according to TechCrunch’s November 6 report. When the first trucks reached customers in late 2023, Musk forecast sales of around 250,000 a year. TechCrunch puts total sales over the subsequent three years in the tens of thousands.
TechCrunch reports that Tesla had spent the previous two months putting Cybercabs on Austin streets and that the company planned to expand to Dallas later in November. Tesla had already used Model Y vehicles for robotaxi operations in Dallas and other cities. The distinction matters when reading a city announcement. A service using modified SUVs and the arrival of Tesla’s dedicated two-seater represent different milestones.
Analysis
Factory space is one part of the plan that can be reassigned. Actual Cybercab production and operating locations will show how quickly Tesla can put that space to use.
Action
Track the production wind-down, deliveries of the replacement vehicle and cities where the two-seater actually operates.
A child’s first phone can come with very different controls
A November 7 TechCrunch survey shows how manufacturers design phones for children who need to call home. Bark phones limit calls and texts to approved contacts. Their software can scan messages and supported apps for concerns and alert a parent. Parents can then unlock additional features as a child grows. Gabb takes another approach, offering calls, texts and selected apps while leaving out a browser, social media and an app store.
Ooma’s home phone offers approved contacts and scheduled quiet hours, with an online portal for call logs. Tin Can uses home Wi-Fi and lets parents approve contacts in a companion app. Its free plan covers calls between Tin Can users, while a paid plan allows calls to other approved numbers. Who a child needs to reach can therefore change the cost of using the device.
Analysis
Monitoring conversations, restricting apps and limiting contacts each give a family different choices about independence and privacy.
Action
Start with one ordinary call your child needs to make. Check whether it works away from home, who approves the contact and which plan it requires.
A UV necklace shows where a wearable’s measurement ends
Testing The90’s Gem necklace, The Verge’s Sheena Vasani found that its ultraviolet readings responded quickly as she moved between sunshine, shade and indoor spaces. The sensor measures UVA and UVB separately. The app uses those readings, optionally alongside a profile that includes skin type, clothing and sunscreen, to estimate several outcomes. Vasani could observe the readings changing but could not independently verify the accuracy of the sensor.
Two details from the November 7 review affect how a wearer might read a chart. On Vasani’s iPhone, fully closing the app interrupted tracking because the pendant did not store measurements. She also watched the app’s estimate of additional sun exposure for vitamin D swing from hours to under an hour and then exceed six hours within minutes. She found broader patterns more informative than the precise countdowns.
Analysis
A break in the record can mean the app stopped collecting. A movement in a countdown can reflect a change in the estimate. Those differences matter when interpreting the same afternoon outdoors.
Action
Check how a wearable records missing data and which numbers it measures directly before drawing conclusions from the chart.