OpenAI announced a two-week pause on reinforcement learning training for its latest models, and has also deferred scaling up certain frontier training runs, including its largest planned RL project. The company stated this is not a halt to research, but rather a temporary measure to strengthen behavioral monitoring, network isolation, safety testing, and research environment protections for high-risk training. Work will resume once safety, alignment, and monitoring capabilities keep pace with model advancements. OpenAI did not disclose the affected models, compute scale, or exact restart date, and noted that future frontier development speed will increasingly depend on confidence in safety measures.

Anthropic disclosed that Claude autonomously designed novel protein binders for 14 out of 15 targets, with results experimentally validated by external partners. The design success rate ranged from 22% to 35%, surpassing the typical 10%–15% baseline cited in the field, demonstrating that general-purpose models can participate in scientific workflows from target understanding to candidate design. The company also released a technical report and opened access to prompts and datasets used in experiments, enabling researchers to verify methods, compare results, and conduct follow-up studies; however, materials did not specify licensing terms or commercial usability.

Dynatrace announced the acquisition of AI observability company Arize for approximately $915 million. The deal focuses on integrating capabilities for monitoring model outputs, prompt lineages, and agent behaviors into Dynatrace’s existing observability platform. Following integration, enterprises will be able to track traditional applications, generative AI models, and agent workflows within a unified operations framework, identifying quality, performance, and operational anomalies to improve auditability and debugging in production environments. Details such as payment structure, expected closing timeline, regulatory conditions, and post-acquisition product pricing, branding, and staffing arrangements were not disclosed.

Anthropic disclosed to investors that its annualized revenue run rate reached $650 million as of July, representing more than a sevenfold increase year-on-year, driven primarily by rapid expansion among enterprise customers. The company is advancing IPO preparations, though no timeline has been set. This figure represents a forward-looking extrapolation based on current revenue trends and does not equate to actual full-year revenue. Materials did not provide net profit, cash flow, customer count, business composition, or audited financials. The data remains part of reported investor communications and awaits confirmation through formal financial disclosures.

OpenAI launched a program to support democratic oversight of government use of AI in national security, committing $5 million over the next year to fund tool pilots, training, usage credits, and expert consultations. The initiative outlines three principles: AI should augment, not replace, human and institutional judgment; government AI applications should be traceable, explainable, and subject to review by authorized oversight bodies; and oversight institutions themselves may responsibly use AI to enhance their capabilities. The program will engage government, civil society, and technical experts, measuring progress by improvements in review capacity and public trust.

LangChain introduced Tuned Evaluators for LangSmith, enabling automatic evaluation of agent conversations directly on production tracing data, with Perceived Error as the first metric. The evaluator detects implicit failures such as user corrections, repeated requests, contradictions, unresolved issues, and agent admissions of error—without requiring teams to write custom judge prompts, configure model credentials, or maintain inference infrastructure. Official benchmarks indicate the specialized model outperforms frontier models while reducing evaluation costs by 82%. Results can be used to filter cases, cluster errors, and build regression test sets.

A research team tested eight models on AppWorld and found that more memory does not always improve agent performance: DeepSeek-V3.2 saw a 9.5 percentage point improvement in task completion when using full guides, while weaker gpt-oss-120b gained 16.1 percentage points through selective retrieval. In contrast, GLM-5, which was already near saturation, showed no benefit. For gpt-oss-120b, selective retrieval added only 5% more tokens, whereas full guides increased token usage by about 50%. The study recommends calibrating memory dosage based on a model’s residual failure patterns and using static prefix caching to manage cost.

Krea shared infrastructure insights from training and deploying its Krea 2 diffusion model, which involved thousands of GPUs and treated frequent failures as observability and recovery challenges. The team monitored GPU temperature, Tensor Core utilization, InfiniBand wait times and errors, and NVLink signals to detect intermittent anomalies invisible to standard utilization dashboards. The system saves checkpoints every 20 to 30 minutes, writing approximately 1TB of data in under 30 seconds. Kubernetes and Kueue are used for scheduling, enabling workload migration between training and online inference.

Anthropic expanded Claude's workplace execution capabilities: Claude can now send emails and perform threaded replies in Gmail, and manage Google Drive files, with all actions subject to user approval workflows. The company also rolled out Claude Cowork to both mobile and web platforms, making it available across all paid subscription plans so users can access collaboration features without being restricted to desktop environments. This update integrates email, file management, and cross-device access into Claude’s workflow, though details on usage quotas per tier, regional restrictions, enterprise admin policies, or free-tier availability were not disclosed.