Anthropic hides METR data on ASI timelines

The US midterm elections just sent a shockwave through the tech sector, triggering a significant drop in market value for AI and chip companies. Newly elected officials are openly campaigning against advanced AI training and demanding strict restrictions on data center buildouts. The rhetoric is escalating from vague safety concerns to concrete calls for a pause on frontier development.

For builders, this is a direct threat to compute access. If local and federal politicians start strangling data center permits to score populist points, the cost of inference will skyrocket. The market selloff reflects a grim reality: the physical infrastructure required to scale AI is now a political football.

The winners here are decentralized compute networks and jurisdictions outside the US that welcome data centers with open arms. The losers are the hyperscalers who assumed they could build indefinitely without political friction.

OpenAI just took the wraps off GPT-6 Cyber, a model purpose-built for cybersecurity, and the debut is terrifyingly effective. In a single autonomous run that burned $1.2 million in API credits, the model discovered 948 escalation bugs in the Linux kernel. This isn't theoretical vulnerability scanning; these are concrete exploits found without human intervention.

This fundamentally alters the economics of cybersecurity. Paying $1.2 million for nearly a thousand kernel exploits is a bargain that nation-states and enterprise security teams will gladly take. The scale of this automated discovery makes the idea of a pause on AI development look utterly suicidal. If offensive agents possess this capability, defensive teams have no choice but to deploy equivalent models immediately.

OpenAI just handed defenders a massive weapon, but they also demonstrated that the barrier to finding zero-days is now purely a function of compute. The winner is anyone selling API credits; the losers are the legacy security vendors relying on manual audits.

Ricursive Intelligence just released the preview of Ripple-65B, a dense model pretrained entirely from scratch on a massive 12 trillion tokens. Released under the Apache 2.0 license, the company claims this US-built model holds its own against significantly larger competitors in the current market.

For developers, a highly capable 65B model is the sweet spot for local deployment and enterprise fine-tuning. By dropping this under a true open-source license, Ricursive is directly challenging the narrative that only closed-API giants can afford to train frontier-class models. The 12 trillion token diet is particularly notable, showing that data scale is still yielding massive performance gains for smaller parameter counts.

Ricursive is positioning itself as the new darling of the open-weight crowd. The losers are the mid-tier closed models that just lost their pricing leverage.

Cognition has officially launched SWE-2 Flash, a significantly faster and cheaper iteration of its SWE-2 model. Post-trained on the open Inkling-Small foundation, the new model is currently available entirely for free to users of Devin Desktop and the CLI through November 30.

This is a classic commoditization play. By building on an open-source base model to drastically cut inference costs, Cognition can afford to give away SWE-2 Flash to acquire users for its Devin ecosystem. For working developers, it means access to top-tier agentic coding capabilities without the crushing API bills that usually accompany autonomous dev tools.

The clear winner is the individual developer who gets a free month of high-speed coding assistance. The loser is any proprietary dev tool still trying to charge premium subscription fees for slower, closed-model inference.

Tencent has officially entered the heavy-hitter multimodal arena with the release of Hy4. This is the company's first Hy model to feature integrated vision and computer use capabilities, backed by a massive 1 million token context window. While the model is currently accessible via OpenRouter, Tencent has promised to release the actual weights later.

A 1 million token context paired with native computer use makes Hy4 an immediate threat in the agentic workflow space. Developers can now feed entire codebases or lengthy video feeds into a model that can natively interact with desktop environments. The promise of open weights means Tencent is playing the long game to capture developer mindshare globally.

Tencent wins by instantly becoming a credible alternative to the closed western models. The losers are the proprietary multimodal APIs that just saw their competitive moat shrink.