Google DeepMind has released Gemini 3.8 Flash and its Cyber variant, maintaining the same input and output pricing as 3.7 Flash at $0.75 and $3.75 per million tokens, respectively. The general-purpose version achieves 73.7% on DeepSWE 1.1; the Cyber version reaches 47.2% pass@1 on CWE-Bench, and in Chrome testing generates correct patches 2.6 times more frequently than large commercial models. The general-purpose model is now available via API and Gemini applications, while the Cyber version is accessible only through Fairwind to government, critical infrastructure, and trusted maintainers.

World Labs has launched an early access version of Atlas, a multimodal spatial intelligence model trained from scratch that accepts text, images, video, and 3D data to build consistent spatial memory and enable world generation, scene reconstruction, and simulation. The model can generate coherent 1440p videos up to one minute long and reconstruct 3D scenes, targeting use cases in VFX, design, and robotics workflows. The release notes indicate that current toolchains are not yet fully compatible with its output formats, and enterprises should evaluate compatibility with asset formats, post-processing steps, and production pipelines before integration.

The European Commission has designated ChatGPT as a 「Very Large Online Search Engine」 under the Digital Services Act (DSA), while naming Reddit and Roblox as 「Very Large Online Platforms」. These services must comply with stricter obligations on systemic risk assessment, mitigation, and transparency within four months. This designation means ChatGPT will no longer be assessed solely as a generative AI product in the EU, but falls under the DSA framework for large-scale information retrieval services; future compliance focus will include risk governance, reporting disclosure, regulatory scrutiny, and platform accountability.

Anthropic has added background computer operation to Claude Cowork and Claude Code desktop apps: Claude can now click, type, and launch applications within authorized apps while users continue other tasks. This feature is currently in beta and available only to Pro and Max plan subscribers on macOS; eligible users are automatically enrolled, or can enable it via settings. Execution shifts from foreground screen takeover to background parallel processing, though scope remains constrained by app permissions and security settings, making it suitable for cross-app, multi-step desktop workflows.

Anthropic has open-sourced blueprints and reference implementations for Claude Commerce Agents, covering retail, travel, telecom, and entertainment, along with a production-ready architecture guide. It recommends prioritizing a 「single model plus Skills」 approach for most commerce conversations to avoid state loss, increased Tokens, and second-long delays caused by sub-agent handoffs; prompt cache hit rates in e-commerce scenarios can reach 90% to 99%. The guide also mandates that financial and business state changes be executed by humans or policies, security rules reside in the harness layer, and evaluation rely on state snapshots rather than fixed dialogue paths.

Cursor now allows Cloud Agents to run on user-managed infrastructure, using machine pools that auto-scale with demand and enabling access to internal enterprise services or dedicated hardware. Agent loops and state management remain on the Cursor platform, while compute execution is offloaded to customer environments. Initial sandbox and deployment providers include AWS Lambda, Cloudflare, Daytona, E2B, Modal, Namespace, Vercel, and Coder. This hybrid architecture primarily targets enterprise development workflows requiring control over data, network connectivity, and compute types.

The U.S. Department of Justice submitted a statement in The New York Times v. OpenAI copyright case, supporting OpenAI’s argument that model training constitutes fair use. The DOJ stated that copying copyrighted text for LLM training is highly transformative and does not substitute for the original works, warning that broad liability could undermine U.S. AI competitiveness and national security. The statement addresses only whether training itself qualifies as fair use, not how training data was obtained, nor does it rule on whether specific model outputs reproduce protected content; final determinations remain subject to judicial review on a case-by-case basis.

DeepSeek has open-sourced an evaluation harness alongside V4 Pro 0813, fully logging every tool call, system prompt, and sub-agent dispatch, enabling external developers to reproduce model performance instead of relying on closed testing environments. The evaluation stack can be forked and adapted into custom benchmarks, compatible with different agent architectures, tool sets, and scoring criteria. Its core value lies in publishing both model outputs and execution processes, providing inspectable experimental traces for independent verification, error diagnosis, and cross-system comparison, reducing reproducibility issues in agent evaluation due to environmental opacity.

A pre-registered, contamination-controlled audit compared 150 pairs of community fine-tuned models against their base models; across 143–144 valid pairs, fine-tuned models performed worse on entirely new tasks: GSM8K dropped by 6.9 percentage points, MMLU by 6.0, and IFEval by 6.5. The study used deterministic scoring instead of LLM judges and predefined exclusion criteria. The authors note that baseline memorization signals have been confirmed only for GSM8K via label permutation tests, while others require further validation—thus the conclusion primarily indicates generalization degradation.

Developers have shared a full implementation of using an agent to optimize the SGLang Diffusion kernel, covering Qwen-Image, FLUX.2, Wan, and SANA Video, with approximately 95% of code written directly or indirectly by the agent. Multiple fusion and quantization optimizations led to a 33.9% end-to-end speedup for Qwen-Image; Wan achieved 17.1% faster VAE decoding, 10.7% pipeline acceleration, and peak GPU memory reduced from 49.6GB to 46.1GB. The case also shows micro-benchmark improvements do not always translate to full-model gains, numerical errors can accumulate across diffusion steps, and human judgment with model-level validation remains essential.