AI development is becoming more specialized while governments tighten rules around capital and technology transfers. OpenAI released a low-cost API for bounded decisions, and Anthropic showed Claude generating formal software proofs at scale. Pathos raised major financing for agent-assisted drug development, while SK Hyper outlined a vast data center plan. New US investment rules and Chinese export approvals put policy beside products, research and infrastructure.
OpenAI launches a fast API for bounded decisions
TechCrunch reported that OpenAI made its Decisions API generally available for applications that need fast choices and confidence ratings from a defined set of answers. Built with GPT-6 Luna, the public version adds voice and image inputs alongside geographically distributed inference capacity. OpenAI said the system reduces decision latency from one to three seconds to roughly 100 to 180 milliseconds in representative workflows.
The pricing is notable: $0.10 per million input tokens, $0.01 for cached input and no output-token charge. A narrow interface can reduce the cost of routing and classification inside larger agent workflows. It does not replace a generative model for open-ended work, and probabilities still need calibration. The value will depend on teams defining stable answers and measuring errors before automated choices affect users.
Anthropic uses Claude to scale formal verification
Anthropic and Theorem used Claude to produce formal proofs for security properties across 123 open-source libraries at the RISC-V assembly level. The work converted existing fuzz tests into specifications, then asked the model to prove that selected programs would exit normally without defined memory-safety failures under a documented set of assumptions. Anthropic said Claude also found long-standing bugs that extensive conventional testing had missed.
Formal proofs can provide stronger guarantees than more tests, but only for the properties and system model they cover. Anthropic notes that the result does not prove the libraries contain no bugs because trusted components and behavior outside the specification remain possible failure points. Publishing proofs makes independent checking possible. If models lower the expert labor required, formal verification may move from a few critical systems into routine development.
Pathos raises $237 million for agent-assisted drug development
Pathos AI closed a $237 million Series E led by Bristol Myers Squibb at a post-money valuation of about $3 billion. The company says its Foundry platform uses thousands of specialized agents to combine clinical, genomic and competitive data for drug-program decisions. Pathos plans to expand the platform, advance four active clinical assets and move the antibody-drug candidate JSKN016 toward registrational studies.
The round shows investors funding AI as an operating layer for clinical development, not only early discovery. Agents may help teams update evidence and abandon weak paths faster, but recommendations remain inside a regulated process requiring biological evidence and patient outcomes. Pathos now has to show that faster portfolio decisions improve trial success. Capital and model scale cannot substitute for controlled studies, safety monitoring or transparent criteria.
SK Hyper outlines a 15-gigawatt AI data center plan
Seoul Economic Daily reported that SK Hyper is seeking roughly $350 billion for the first 5 gigawatts of a proposed 15-gigawatt South Korean data center network. The SK Telecom unit aims to serve domestic demand and attract training and inference workloads from Japan, Taiwan and global hyperscalers. It plans phased operations from 2029 and an expansion to 15 gigawatts by 2035, subject to customer demand.
The first phase is about five times the size of South Korea's current data center market cited by the company, so international customers are central. That scale exposes the project to financing, power and utilization risk. SK Hyper's focus on cost per generated token reflects a shift from acquiring GPUs to operating them efficiently. Firm customer commitments and staged construction will matter more than headline capacity.
US Treasury proposes broader outbound investment controls
The US Treasury proposed rules to implement the Comprehensive Outbound Investment Security Act. The framework would require notices for some US investments in covered foreign companies and prohibit others when specified strategic technologies are involved. Congress named semiconductors, AI, quantum technology, high-performance computing and hypersonic systems, while Treasury added biotechnology to the proposal.
The proposal shifts more national-security screening to investors and their controlled foreign entities. Funds may need deeper diligence on a target's technology, ownership and military or surveillance connections before completing a transaction. Technical thresholds and exceptions will determine whether the rules focus on advanced capabilities or affect wider commercial activity. Treasury is taking public comments before issuing final regulations.
DeepSeek and MiniMax clear China's overseas process
The Star reported that DeepSeek and MiniMax completed Chinese registration procedures for open-source releases and overseas model use. The report cited a social account affiliated with state broadcaster CCTV and followed China's draft export-control proposal for advanced models and agents. It said API access, web apps and software-as-a-service products would not be restricted even when they use covered models underneath.
The approvals test how China may preserve global commercial reach while controlling sensitive technology transfers. Treating downloadable models differently from hosted access could keep Chinese providers available abroad without allowing every capability to move freely. Developers need clarity on fine-tuning, updates and open-source obligations. Alongside US investment controls, the policy shows both countries building gates around different parts of the AI supply chain.
What matters today
Today's signal is specialization under constraint. OpenAI is separating bounded decisions from general generation, and Anthropic is applying frontier models to a precise assurance problem. Pathos and SK Hyper are raising or seeking extraordinary capital for clinical and physical infrastructure. US and Chinese rules are also shaping where money and model technology can travel. Model quality alone will not decide the winners. Verifiable outputs, disciplined investment and cross-border compliance are becoming equally important.