OpenAI announced zero data retention for eligible API customers using frontier models: prompts and responses will not be stored after request processing unless customers explicitly opt in for training purposes. The concurrently previewed Private Safety Processing enables detection of cross-interaction abuse patterns on customer-controlled or encrypted storage, returning only narrowed risk signals to OpenAI—raw content remains inaccessible to employees. Customers retain full authority to investigate and decide whether to disclose information; the system is scheduled to begin rollout in September 2026, accompanied by a technical whitepaper. Detection of child sexual abuse material remains an exception due to legal obligations.

Cerebras launched its next-generation AI computer, the CS-4, on August 18, claiming inference speeds up to 30 times faster than GPU systems and support for models exceeding 10 trillion parameters, delivering over 1,000 tokens per second. The company also introduced a modular Nexus architecture and decoupled inference approach, aiming to reorganize compute resources according to workloads. The CS-4 is planned for broad availability in Q3; however, performance claims are from the vendor, with no independent testing, pricing details, or side-by-side benchmarks against NVIDIA systems provided—real-world deployment performance remains to be validated.

Etched completed a $700 million funding round within approximately 26 days, increasing its valuation from around $10.3 billion to $21 billion—an over 100% rise. Jane Street led the round and became the first customer for Etched's chips, aligning capital investment with early commercial adoption. According to reports, the company advanced its chip from testing to actual operation in just 44 days, demonstrating rapid hardware validation and delivery pace. This deal positions Etched among high-valued AI chip startups, though materials do not disclose chip performance metrics, production scale, financing terms, or future delivery timelines.

OpenRouter officially announced it will join Stripe, with both parties planning to integrate multi-model routing capabilities with financial infrastructure including payments and risk controls, supporting growing model invocation and transaction volumes. OpenRouter stated that its API, product roadmap, and model-agnostic neutral gateway positioning will remain unchanged, continuing to serve developers using diverse models. This announcement confirms earlier acquisition reports at the company level, but materials do not disclose transaction amount, closing date, regulatory conditions, or integration plans. Long-term implications for product independence and commercial terms require further observation.

Zhipu's GLM-5.3 API is now live, maintaining pricing at $1.4 per million input tokens and $4.4 per million output tokens. Materials indicate enhanced coding and long-horizon agent capabilities. Co-founder Tang Jie stated that post-training alone improved coding ability by 50%, and further discussed MoE mechanisms and multidimensional Scaling Laws. Model weights remain closed-source; another report links latency to its CyberGym score of 84.5%, reflecting strong cybersecurity performance. The API allows developers to begin calling immediately, though weight release timing has not been announced.

Google launched an AI learning suite for students: eligible U.S. college students can access Google AI Pro free for one year, while students in over 140 countries gain access to AI Plus, with some plans offering quadrupled usage quotas and 5TB of storage. Gemini added new features including a student hub, study notes, diagnostic quizzes, interactive charts, and 3D simulations. Gemini Live now supports voice-initiated, multi-step Deep Research via background execution. Search has integrated custom quizzes, Lens step-by-step explanations, and learning document generation, with related features extending to Google Classroom.

NVIDIA released a method for building federated multimodal AI workflows using FLARE, focusing on collaborative training of vision-language models without data leaving local institutions. The approach recommends freezing large multimodal backbones and exchanging only lightweight adapters such as LoRA, reducing communication volume per round from gigabytes to megabytes. When larger updates are necessary, tensor streaming can bypass message size limits, and disk offloading during aggregation prevents server memory from scaling linearly with client count. NVIDIA also recommends comprehensive evaluation across model quality, communication cost, runtime, memory usage, and failure rates.

Anthropic rolled out centralized security and operational enhancements for Claude Managed Agents. Web search and scraping tools now include 「allowed_domains」 and 「blocked_domains」 parameters, enabling developers to restrict websites accessible to agents. Results from Self-Hosted Sandboxes can now be written into memory for persistent use across subsequent tasks. The Console session viewer has been restructured for multi-agent workflows, introducing per-agent mini-maps, grouped streaming logs, and cost inspection by thread and session—facilitating tracking of concurrent execution paths, resource consumption, and anomalous behavior.

Cursor updated its cloud agents to enable continuity across external events and maintain long-running goals until task completion. The new 「/goal」 command allows users to set persistent objectives, reducing interruptions caused by single-turn context limits or mid-process state loss in complex engineering tasks. Interaction control has also been refined: when issuing redirection commands, users can choose to wait until the next tool call before intervening, rather than interrupting ongoing actions mid-execution. Official materials have not yet disclosed package restrictions, default timeouts, goal persistence duration, or full rollout scope for these features.

smolvm 1.8.3 was tested for hardware-isolated execution of untrusted Python and JavaScript code. Testing showed cold starts taking approximately 0.6 to 1.5 seconds, with hot execution stabilizing around 50ms, while CPU, memory, runtime, and storage quotas were effectively enforced. Offline images, network disabled, read-only input mounts, writable output directories, and non-privileged mode all functioned as intended—suitable for intermittent data transformation tasks. Since the Claude Code web environment does not provide KVM, testers used GitHub Actions with KVM support to run full tests.