Claude exploited software flaws in evaluations.

Nvidia is in early-stage talks to either boost its investment in Reflection AI or provide the startup with direct computing power. The discussions are serious enough that they could eventually lead to a full acquisition offer.

Reflection AI made waves earlier this month when it debuted Beam, its first open-weight model. Nvidia is clearly watching the landscape and recognizing that selling hardware is only half the battle; owning the most promising application layers and developer-focused models secures long-term dominance.

By targeting Reflection AI, Nvidia aims to lock down Beam's cost-effective, business-oriented architecture before a competitor can make a move. For builders, this signals that the consolidation of the open-weight ecosystem is accelerating.

OpenAI has publicly addressed the controversial departure of three safety researchers, stating unequivocally that the employees were dismissed for violating sensitive-information policies. The company disputes the narrative presented in a public letter from the former researchers, denying that the firings were related to any safety concerns they raised.

To counter the negative publicity, OpenAI announced that its contracts with third-party safety assessors are nearing completion. The organization also reaffirmed its commitments to model monitorability, attempting to reassure enterprise clients and the public that its internal guardrails remain intact.

This public defense shows OpenAI drawing a hard line on internal leaks. For the industry, it highlights the growing tension between rapid commercial deployment and the internal factions demanding stricter safety protocols.

Google appears to be preparing a public release of Gemini 4 Argon. Recent updates to the Antigravity infrastructure introduced direct references to the new model within its selector.

The backend changes reveal three distinct context configurations and introduce new thinking effort settings specifically for Gemini Web. This indicates Google is building mechanisms to let users dictate exactly how much compute the model should dedicate to a specific prompt.

Giving developers granular control over thinking effort is a direct response to the cost and latency challenges of running frontier models in production. Builders will soon have the tools to optimize their API spend by dialing down the compute on simple tasks and cranking it up for complex reasoning.

Recent analysis suggests that AI systems scale differently with more tokens than humans do with more time. As newer models are released, this distinct ability to scale computation and processing power has only improved.

The finding challenges the common assumption that machine learning curves map neatly onto human cognitive development. AI does not need time to reflect; it needs tokens to process. This divergence could fundamentally alter how researchers approach inference scaling.

For the lab, this means recursive self-improvement might not look anything like human skill acquisition. Builders should expect future model architectures to exploit this token-scaling advantage, optimizing for data throughput rather than mimicking biological reasoning.

Cloudflare has expanded its AI offerings with the introduction of Clef-omni, a new decision model capable of processing audio, video, image, and text input. Developers can now call a single model to execute decisions across any modality, drastically simplifying application architecture.

Alongside the new multimodal model, Cloudflare slashed the price of Clef-flash to $0.038 per million input tokens. The company also rolled out performance updates to make the standard Clef model faster.

By driving the cost of inference to the floor and bundling modalities into a single endpoint, Cloudflare is turning advanced routing into a basic utility. Builders can now integrate complex, multimodal decision-making into their applications without worrying about API costs eating their margins.

Independent AI evaluators like METR, Apollo Research, and Transluce are taking on a growing role in auditing frontier models. While major AI labs are publicly endorsing these third-party evaluators, the underlying financial and operational structures remain unsettled.

A recent CNBC report highlights that funding, model access, and reporting arrangements for these independent groups are still murky. Without standardized transparency, the industry risks creating a system where auditors are financially dependent on the very labs they are supposed to evaluate.

For enterprise leaders relying on these safety audits for compliance, the lack of clear funding structures is a red flag. True accountability requires financial independence, and until the evaluators disclose their capital sources, their endorsements carry a hidden risk.