Claude Gets Mythical and China Goes Underwater

Until now, Anthropic kept its most capable reasoning features locked behind closed doors — available only to select research partners and enterprise clients willing to sign special agreements. With Claude Fable 5, that wall just came down.

Fable 5 is Anthropic's most powerful publicly available model to date, and its headline feature is the inclusion of Mythos capabilities, a suite of advanced reasoning and creative synthesis tools that the company had previously classified as too sensitive for general release. The fact that Anthropic built an entire internal restriction layer around Mythos — and is now dismantling it — tells you something important about how seriously the company takes capability thresholds. It also tells you they've decided the competitive pressure is too great to keep sitting on it.

The timing is not subtle. OpenAI has been aggressively pushing GPT-4o updates and teasing its next flagship, while Google's Gemini lineup continues to expand. Anthropic has generally positioned Claude as the more thoughtful, safety-conscious alternative. Releasing Mythos to the masses suggests they're now willing to trade a little of that cautious brand identity for market share.

So what does Mythos actually do? Based on what Anthropic has shared, it significantly improves the model's ability to handle long, complex, multi-step reasoning tasks — the kind of thing that matters enormously in legal analysis, scientific research, and advanced coding workflows. Early benchmarks reportedly show Fable 5 outperforming its predecessors by meaningful margins on tasks that require holding large amounts of context while drawing nuanced conclusions. That's the hard problem most frontier models still stumble on.

There's a broader story here about how AI companies manage the gap between what their models can do and what they let the public access. Anthropic has been unusually transparent about maintaining internal capability tiers, treating some features as experiments that need more safety evaluation before wide deployment. Whether you find that reassuring or paternalistic probably depends on how much you trust the company doing the gatekeeping.

For enterprise customers, Fable 5 arriving with Mythos unlocked is a genuine upgrade. Complex document analysis, multi-document synthesis, and sophisticated coding assistance should all see measurable improvements. For individual users and smaller developers, it means access to tools that were, until recently, simply not available at any price point on the consumer tier.

The release also raises a question worth watching: if Anthropic felt comfortable enough to release Mythos now, what are they currently holding back that they're not yet comfortable with? Every frontier lab has a version of that list. The gap between what exists in the lab and what gets deployed is where the most interesting — and occasionally alarming — AI developments tend to live.

Cooling systems eat up nearly half the electricity consumed by a conventional data center. China just built one where that figure drops below 10 percent — and it's sitting on the ocean floor.

Located off the coast of Shanghai in the Lin-gang Special Zone, the new facility is a collaboration between private firm HiCloud Technology and state-owned China Communications Construction. The project cost roughly $236 million and represents the world's first underwater data center powered entirely by offshore wind. It sits at a depth of 10 meters, where the surrounding seawater handles most of the cooling work that onshore facilities burn enormous amounts of electricity to replicate artificially.

The efficiency numbers are striking. Traditional data centers typically post a Power Usage Effectiveness score — the industry's standard energy efficiency metric, where 1.0 is perfect — somewhere between 1.4 and 1.6. The Lin-gang facility is designed to hit 1.15 or better. That gap translates directly into less energy wasted, lower operating costs, and a smaller carbon footprint at scale.

This isn't China's first underwater data center. HiCloud launched a commercial seabed facility in Hainan back in 2023. But that earlier project relied on conventional grid power. The Shanghai installation is the first to pair the underwater cooling concept with offshore wind generation, which is what elevates it from an interesting engineering experiment to a potentially replicable infrastructure model.

The context here matters a lot. A recent UN report found that roughly 90 percent of the world's AI-specialized data center capacity is concentrated in just two countries: China and the United States. Both are scrambling to secure enough clean, reliable energy to feed the AI boom without either bankrupting themselves on electricity bills or inviting serious political backlash over emissions. They're just taking very different roads to get there.

The US has largely pivoted toward nuclear — recommissioning old plants, signing deals with next-generation reactor developers, and in some cases reversing earlier commitments to renewable-first energy policy. China, by contrast, is doubling down on renewables, treating offshore wind as both a climate tool and a strategic hedge against dependence on imported fossil fuels. An underwater data center powered by wind turbines is almost a perfect symbol of that approach.

The facility's first phase runs at 24 megawatts of capacity, which is modest by hyperscale standards but meaningful as a proof of concept. If the operational data holds up — and the PUE stays near 1.15 over real-world conditions rather than just theoretical design specs — expect other countries and private operators to start asking whether seabed infrastructure deserves a spot in their long-term data center strategies.

For now, China has built something nobody else has, and they built it fast. Whether the rest of the world treats that as inspiration or a reason to accelerate their own plans is a question the next few years will answer.