01Investigation documents more than 1,000 people who crossed US border surveillance zones without being reached and later died, some under AI-tower watch
The United States has spent billions of dollars over 25 years building a “virtual wall” of surveillance towers along its southern border. The system is intended to detect and help apprehend people crossing the border—and potentially save them before heat, exposure or other dangers become fatal. A new MIT Technology Review investigation found more than 1,000 people who passed through tower-monitored areas without being reached or apprehended and ultimately died nearby, including some in areas watched by newer AI-enabled towers.
Over 15 months, reporters assembled what they describe as the first comprehensive map and analysis of deaths near the virtual wall. They examined cases dating to 2015, combining datasets from humanitarian and civil-liberties groups with records obtained from more than a dozen agencies. A case qualified for analysis only when reporters could establish that the person was believed to have been crossing the border, where the remains were found and roughly when the person died.
To locate the surveillance network, the team relied on mapping led by the Electronic Frontier Foundation, then used satellite imagery to verify individual towers and estimate when they were installed. It compared the resulting map of nearly 600 towers with death locations and counted more than 1,050 people who died nearby.
That figure is not a total for all border deaths, nor proof that surveillance failures caused every death. “Died near a tower” does not necessarily mean a person was detected, and the available records cannot establish what happened in every case. But the investigation identified specific breakdowns: towers that were not working, algorithms that failed to recognize people, and alerts that agents did not answer.
The count is also incomplete. The government says about 800 towers currently operate, but it does not publish a comprehensive map or complete installation and operating dates. Texas records were missing, damaged or still unavailable in some counties, while reliance on county sheriffs excluded cases handled by local or state police.
Those gaps remain as the government prepares to spend $1 billion to triple the virtual wall’s size by 2034. Sources told the reporters that US Customs and Border Protection has never conducted a comprehensive audit matching tower activity with nearby deaths. Its death-record system has omitted cases, particularly those not discovered by Border Patrol, while technology-assisted apprehension data has been found inaccurate and unreliable. A separate migrant-safety program collects information on calls and rescues but lacks a plan for measuring whether it increases rescues or reduces deaths.
02Anthropic launches Claude Opus 5.5, says typical workloads cost 40% less than Opus 5
Anthropic has released Claude Opus 5.5, the first model in its new Claude 5.5 family and its new leading model for complex coding and agent tasks. It is the company’s first model release since Anthropic called for slowing the pace of frontier AI development, amid concerns about increasingly capable systems acting beyond their assigned boundaries.
The company says Opus 5.5 lowers both the cost and waiting time of long-running work while adding safeguards for autonomous operation. At default settings, Anthropic’s tests found that typical workloads cost 40% less than with Opus 5 and generated output more than 30% faster. Those results have not been independently reproduced.
Input tokens cost $4 per million and output tokens $20 per million, both 20% below Opus 5. The larger claimed saving does not come from that price cut alone: Anthropic says Opus 5.5 uses fewer tokens per task, while cache reads—which it says account for most agentic and coding costs—fell 60% to $0.20 per million tokens.
Anthropic is also raising five-hour usage limits for Pro, Max, Team and seat-based Enterprise plans, although it did not disclose the new limits. Subscription users receive a rate-limit reset they can save for later use. A Fast mode in Claude Code and the Claude Platform offers up to 2.5 times the speed, priced at $8 per million input tokens and $40 per million output tokens.
For extended coding jobs, Anthropic says an early tester audited and fixed a 200,000-line codebase in under three hours, compared with more than 20 hours and 2.5 times as many tokens for Opus 5. In an internal test translating HAProxy, software that distributes web traffic across servers, from C to Rust, Opus 5.5 finished in 9.5 hours—2.5 hours faster than Fable 5.1—and cost 51% less. Both versions passed nearly all of HAProxy’s regression tests.
Anthropic says Opus 5.5 is less likely to take hard-to-reverse or unauthorized actions and is more resistant than Opus 5 to prompt injection, an attack that gives an AI malicious instructions through its inputs. Frontier Design and METR evaluated the model before release. Anthropic is applying safeguards similar to those used for Fable 5.1: vetted organizations can seek verified access for biological research, while expanded access for verified cybersecurity practitioners is planned in the coming weeks.
03TypeSafe AI Launches Jev: No Text Generation, Only Probabilities for Classification and Decisions
TypeSafe AI has launched Jev, an API that processes text or semi-structured data but, unlike a conventional large language model, does not generate a written response. Instead, it returns numeric results for classification and decision tasks. The company calls Jev its first “System One” model, though “decision model” more plainly describes its role.
Developers assemble a “state” from a string, an array of strings, or name-value pairs representing an article, customer, or other record. They can then submit one or more typed questions about that same state. In TypeSafe’s formulation, the API takes unstructured state as input and produces probabilistic decisions as output.
Jev supports three question types. A yes-or-no, or “Noul,” question returns a number from zero to one representing the model’s confidence that a statement is true. A choice question selects from supplied options while returning both a confidence score and a probability distribution across every option. A score question uses developer-defined numeric levels and descriptions, then returns a floating-point value along that range. These outputs are model scores, not established facts or necessarily calibrated probabilities.
A single state can carry as many questions as fit within the context window, and Jev evaluates them in parallel. TypeSafe charges only for input, at $0.042 per million tokens, with no output fee. That combination could suit high-volume classification workloads such as spam detection, label suggestions, prioritization, ranking, or search reranking, where many judgments must be made about the same document. These are proposed uses rather than demonstrated business results.
The low price also makes repeated evaluations and structured experiments inexpensive, but Jev’s numeric-only output creates a testing problem. If it flags a message as spam or assigns a low ranking, it does not explain which signals affected the decision. That makes hidden bias harder to identify and investigate than with a language model that can at least produce a rationale, even though such explanations may themselves be unreliable.
Jev’s version 1.13 documentation says the model currently struggles with numbers, dates, and adversarial content. Its architecture, training data, independent performance, probability calibration, and production-scale stability have not been disclosed.
OpenAI releases cheaper GPT-6 Sol and Luna models OpenAI launched GPT-6 Sol for complex tasks such as coding and Luna for high-volume clerical work, pricing API access at half the cost of their 5.6-series predecessors. The company says Sol makes about half as many mistakes as its predecessor in an internal factuality evaluation. techcrunch.com
Qualcomm unveils smartphone chips built for local AI agents Qualcomm announced the Snapdragon 8 Elite Gen 6 and higher-end Extreme version, whose sensing hubs can run small models locally for features such as transcription, speaker identification, and personalized automation. Qualcomm says the Extreme chip can run a 30-billion-parameter mixture-of-experts model on-device. techcrunch.com
Snorkel AI raises $350 million at a $3.5 billion valuation Snorkel AI, which supplies training datasets and simulated environments to AI labs and companies, raised a Series E led by Insight Partners and S32. The company says its annualized revenue run rate reached $375 million after shifting toward completed datasets generated through software, models, and subject-matter experts. techcrunch.com
Rabbit launches a hardware-independent desktop agent Rabbit is rolling out OS3, a cloud-based agentic operating system that works across Windows, macOS, and Linux and can coordinate files, apps, devices, and users’ preferred AI models. Founder Jesse Lyu said Rabbit has stopped manufacturing its R1 device, though the company plans another OS3-based hardware product. theverge.com
AstroForge plans to put an AI system in control of a spacecraft Asteroid-mining startup AstroForge developed Solo, an autonomous control stack combining conventional algorithms with transformer models trained on spacecraft sensor data. Solo is scheduled to operate in shadow mode on DeepSpace-2 before controlling a 2027 spacecraft intended to fly without routine direction from Earth. techcrunch.com
GPT-6 adds prompt-cache diagnostics and explicit breakpoints OpenAI introduced prompt-caching improvements for GPT-6, including higher cache-hit rates, new diagnostics, explicit breakpoints, and additional controls. The features are intended to help API developers reduce latency and inference costs. openai.com
OpenAI calls for shared international AI standards OpenAI outlined a proposed framework for coordinated model evaluation, reporting, and governance across countries. The company argues that common global standards could improve the safety of increasingly capable AI systems. openai.com
Nscale files for an IPO with 85% of contracts tied to two customers British AI infrastructure provider Nscale disclosed more than $103 billion in contracts, about 85% of which comes from agreements with Microsoft and Anthropic. Its filing says the Anthropic agreement depends on Nscale securing financing and meeting milestones, while the company reported a $1.02 billion first-half net loss. techcrunch.com
Andreessen Horowitz creates a 50-student Silicon Valley academy Venture firm Andreessen Horowitz is launching a free, unaccredited one-year program in fall 2027 with partners including Anthropic, Google, Meta, Nvidia, OpenAI, and Palantir. Participants will take short classes and work in company placements, while receiving compute credits and a travel and research budget. theverge.com
RoboDawn uses vision-language models to control robots without task-specific training Researchers introduced RoboDawn, an interface through which a vision-language model observes a robot’s environment and issues discrete movement and gripper commands. In reported benchmarks, one demonstration raised success from 53.2% to 73.6% on RoboTwin 2.0 C2R, and the framework also completed block-manipulation tasks on a physical Franka robot. huggingface.co
WorldCrafter adds 3D-aware memory to interactive video generation Researchers presented WorldCrafter, a video world model that compresses past observations into viewpoint-specific tokens to preserve scenes across camera movements and longer sessions. Their experiments report improved consistency and camera control during minute-scale exploration generated from a single image or text prompt. huggingface.co