Deep-sea cable cutting capability & Anthropic Claude Opus 4.7 launch - Tech News (Apr 17, 2026)

First up: undersea cables, and a test that’s making a lot of governments and network operators uncomfortable. China has reportedly demonstrated a deep-sea cable-cutting device at around 3,500 meters—right in the depth range where many major communication lines sit. Officially, it’s described as a tool for subsea engineering work. Unofficially, it’s a reminder that the internet isn’t just “in the cloud”—it’s physical infrastructure laid across the ocean floor. In a world already jumpy about unexplained cable breaks, publicly showcasing a capability like this raises the pressure for better monitoring, redundancy, and faster repair capacity.

Staying in geopolitics, Australia is moving to lift defense spending to three percent of GDP by 2033, its biggest peacetime ramp-up on record. What’s notable for tech watchers is the emphasis on drones and autonomous systems—another sign that modern deterrence is becoming as much about software, sensors, and unmanned platforms as it is about ships and aircraft. It also signals a longer runway of demand for industrial supply chains that can actually build and sustain these systems, not just prototype them.

Now to the frontier AI race. Anthropic has released Claude Opus 4.7 as its most capable broadly available model, while keeping a stronger successor—called Mythos—limited to select enterprise partners for cybersecurity testing. The headline here isn’t just who’s “winning” a benchmark by a sliver. It’s the direction of competition: models are being judged on whether they can operate like reliable agents—planning, checking their work, and sticking the landing over longer tasks. Anthropic also added more knobs to manage cost and speed for deeper reasoning, which tells you what enterprises are asking for: predictability, not magic.

And there’s a second layer to that Anthropic release: security. As models get better at autonomous tool use, the boundary between “helpful assistant” and “dangerous capability” becomes harder to police. Anthropic is leaning into tighter controls for high-risk cyber requests and gating certain access through a verification program. That’s likely a preview of where the whole industry is headed: more capability, but more locked doors—and more arguments over who gets keys.

On the broader AI power map, a Stanford HAI report argues China has nearly closed the U.S. lead in AI performance, at least on prominent chatbot comparison benchmarks. The U.S. still produces more top-tier models, but China is leading on several scale indicators—things like citations, patents, and industrial robot installations. The report also flags a very practical constraint for the U.S.: electricity. If power generation and grid upgrades can’t keep up with new data centers, “compute” stops being an abstract concept and becomes a concrete bottleneck.

Zooming into how AI actually gets used at work: the agent ecosystem is rapidly shifting toward more standardized tooling. Google announced new tools to help “agentic” Android development work outside Android Studio, including a revamped command-line workflow and a way for assistants to pull up-to-date guidance instead of relying on stale training data. And on the platform side, Cloudflare rolled out a batch of agent-focused capabilities—from persistent memory and versioned storage to expanded workflow controls—aimed at making agents easier to run in production without turning operations into chaos. The common theme is maturity: teams want agents that don’t just generate code or text, but fit into real pipelines with guardrails.

One practical reality check came from a pricing-focused analysis that’s getting shared widely: token costs aren’t as comparable as they look. Different models chop up the same text into very different token counts, which means the “cheapest per token” model can become the most expensive depending on whether you’re sending plain language, structured data, or tool definitions. The takeaway is simple and unglamorous: if you’re budgeting for AI, measure on your actual prompts and workflows. The spreadsheet math only works if the inputs are real.

Let’s talk about personalization—and privacy—because Google is pushing Gemini further into people’s private data. The company is rolling out a feature that lets users connect Google Photos to Gemini for personalized image generation, so the chatbot can create new images based on the photos you’ve already stored. Google says this doesn’t directly train its models on your Photos library, but the product direction is clear: assistants that know you visually, not just via text. That’s powerful for creativity, and also a major trust test for how user data is handled, audited, and protected.

In life sciences, OpenAI introduced GPT-Rosalind, a biology-tuned model aimed at common research workflows like synthesizing evidence, proposing hypotheses, and planning experiments. Access is restricted through a trusted program, reflecting the obvious biosecurity concerns, and OpenAI is also pushing a plugin that connects to a wide set of scientific tools. The bigger signal is where AI is heading: away from one general assistant for everything, and toward domain-specific systems that are trained around how professionals actually work—especially in high-stakes fields.

On the diagnostics front, researchers at NTU Singapore reported an AI-assisted biochip that can detect disease-linked microRNAs from a small blood sample in about twenty minutes. The interesting part isn’t just speed—it’s that they’re aiming to avoid the more time-consuming amplification steps common in older methods. If that holds up in broader clinical validation, it could make microRNA testing more realistic for earlier, less invasive screening and routine monitoring, not just specialized labs.

And on the treatment side, two big clinical stories are worth tracking. First, the Netherlands’ DRUP platform trial looked at what happens when doctors use already-approved cancer drugs off-label based on tumor genomics after standard options run out. A meaningful slice of patients benefited, but toxicity was substantial, and only a smaller group saw exceptional, durable results. It reinforces a key point: precision oncology can work, but it works best when off-label use is captured in structured, data-generating programs that can separate hope from evidence. Second, a large trial in The Lancet Psychiatry found magnetic seizure therapy performed about as well as electroconvulsive therapy for severe, treatment-resistant depression—while causing fewer cognitive side effects, particularly around memory. If this continues to replicate and clears the hurdles for broader adoption, it could widen access to an effective treatment that many patients currently avoid due to fear of cognitive impacts.

Now for one of the more sci-fi items that’s very real: Northwestern engineers created printed artificial neurons that generate electrical spikes realistic enough to activate living brain cells in lab tests. The reason this matters is the materials and form factor—soft, flexible devices that better match the body than rigid silicon. Over time, that could translate into more practical neuroprosthetics and brain-machine interfaces, and it also hints at more energy-efficient “brain-like” computing approaches.

In robotics, Physical Intelligence published research claiming its latest model can direct robots to complete tasks they weren’t specifically trained for by recombining learned skills. A headline demo involved using an air fryer with minimal prior exposure, especially when guided with step-by-step spoken coaching. The cautious framing is important: we’re not at the point where you give one vague command and a household robot flawlessly executes a long sequence. But the trend is clear—robots that can be improved with natural language guidance could reduce the need for endless retraining, making deployments more adaptable in messy real environments.

Two stories to close on how the internet itself is changing. First, an analysis argues “clips” are no longer promotion—they’re the product. The claim is that short snippets now drive the real attention, the embedded ads, and even the creation of new media stars, while full episodes become optional background. If that’s true, it helps explain why creators and platforms keep optimizing for highlights: the business model follows the scroll, not the schedule. Second, Europe has a new push for digital sovereignty with Eurosky, which launched as a European social media infrastructure layer rather than a single app. The pitch is a unified identity and personal data storage under EU law, built on the same protocol ecosystem as Bluesky. It’s early, and it still relies on parts of existing infrastructure, but the direction is notable: a growing appetite for alternatives that reduce dependence on U.S.-centric platforms—especially as regulation, moderation, and AI-generated abuse collide.

Finally, a quick note from the software engineering trenches: developers are debating how to keep code review sane in an era of giant automated diffs and AI-generated changes. One argument gaining traction is to split work into “tastefully broken” commits—separating mechanical refactors from real logic changes—so reviewers can actually follow what happened. Pair that with squash merges, and you get a cleaner main branch without forcing every intermediate step to be perfect. It’s a small process tweak, but it speaks to a big shift: as code generation gets cheaper, validation becomes the scarce skill—and the best teams will optimize for clarity and accountability, not just speed.