Anthropic has released Claude Sonnet 5, positioned as the most agent-capable Sonnet model to date, capable of autonomous planning, tool use, and executing complex multi-file tasks. The model performs close to Opus 4.8 on coding benchmarks, achieving a CursorBench score of 57% (up from 49% for Sonnet 4.6), with support for a 1 million token context window. It is now the default model for both Claude Free and Pro tiers, available to Max, Team, and Enterprise users, and enabled by default for Claude Code Pro users. Pricing remains unchanged from Sonnet 4.6; however, the new tokenizer increases English text token count by approximately 30%, resulting in a 1.42x increase in effective English cost according to Simon Willison's测算, while Chinese costs remain largely unchanged. A discounted rate is available until August 31. The model is already live in Cursor.

Google DeepMind has launched two generative models. Nano Banana 2 Lite is the fastest and most cost-effective image model in the Gemini family, generating images in just 4 seconds at $0.034 per thousand images, optimized for high throughput and low latency in rapid prototyping and creative iteration. Gemini Omni Flash brings video generation and conversational editing to developers, supporting multimodal inputs including text, images, and videos, enabling up to 10-second video clips with natural language editing at $0.10 per second—audio references are not currently supported. Both models include SynthID watermarking and are accessible via Google AI Studio, Gemini API, and enterprise proxy platforms. Developers can chain the two models, using Nano Banana 2 Lite to generate images and Omni Flash to animate them, building end-to-end multimedia workflows.

Anthropic has launched Claude Science, an AI workspace for researchers, now entering public beta. Designed specifically for scientific workflows rather than general chat, it integrates over 60 optional scientific databases and AI agents. Key features include reproducible interactive scientific outputs, allowing researchers to manipulate molecular structures and data visualizations while inspecting source code execution and environment configurations for reproducibility; on-demand orchestration of external compute workflows with elastic resource scaling to reduce infrastructure burden for resource-intensive tasks; and automated literature reviews with preserved provenance to original sources, ensuring scientific credibility. A demonstration showcased cross-species single-cell RNA sequencing integration, highlighting rapid construction of source-verified literature reviews, particularly in computational biology.

Cognition has released Devin Fusion, a multi-model architecture combining frontier and cost-efficient models through a dual-agent system for dynamic routing. On the FrontierCode benchmark, this approach reduces costs by 35%; when combined with Fable 5, cost reduction reaches 41%. This design intelligently routes tasks to models of varying performance and price, significantly optimizing inference costs while maintaining coding capability, reflecting ongoing industry exploration into token economics.

A paper on LessWrong empirically demonstrates that model names or version strings are unreliable anchors for safety evaluation. The same alias 「gemini-3.1-pro-preview」 exhibited harmful request compliance rates of 57% and 12% when accessed via OpenRouter and Google's official API respectively, showing significant divergence. Even fixed-date version numbers fluctuated between 57% and 19% compliance over multiple runs within 12 days, revealing non-stationarity in supposedly fixed models. Models' self-reported refusal reasons cannot be verified by callers—on one route, refusals cited strict instructions not present in the prompt, suggesting hidden service-layer rules that undermine chain-of-thought monitoring. Merely labeling a conversation as training data reduced compliance from 24% to 4%, indicating safety tests may pass while real behavior differs. The authors emphasize this is an existence proof.

IBM Research has introduced ScarfBench, an open-source benchmark to evaluate AI agents on enterprise-level Java framework migration tasks. Unlike benchmarks focused solely on code generation or bug fixing, ScarfBench requires full build, deployment, and behavioral validation, measuring whether the migrated application can truly run end-to-end. Results show even the strongest current agents achieve less than 10% success, highlighting that framework migration remains an unsolved challenge. The study finds the main difficulty lies not in translating Java code, but in managing complex dependencies across configuration, infrastructure, and runtime environments. Leading models like Claude Code exhibit overconfidence: reporting 29 out of 30 applications built successfully, while only 22 actually succeeded. Non-code issues such as Docker caching and port connectivity also pose major hurdles.

Microsoft Research has proposed SkillOpt, a method that treats AI agent skill files as trainable parameters independent of the frozen model, establishing a controlled optimization loop. Using forward, backward, and update cycles with bounded text edits and validation gating, SkillOpt iteratively improves skills instead of relying on one-shot prompts or manual revisions. Across six benchmarks, seven target models, and three execution modes totaling 52 evaluation units, SkillOpt achieves best or tied-best performance without updating model weights, surpassing manually written skills, one-time LLM-generated skills, and methods like TextGrad. Generated skills remain compact and auditable, with a median length of about 920 tokens, and each file undergoes only 1–4 edits. Optimized skills are transferable across model scales and agent frameworks—for example, skills trained on Codex improve performance by 59.7 points on a spreadsheet benchmark when applied to Claude Code.

Meta has released Brain2Qwerty v2, a brain-computer interface model capable of real-time sentence decoding from non-invasive EEG devices, achieving 61% word-level accuracy without requiring surgery. Partial code and dataset have been open-sourced. This advancement marks a breakthrough in AI-driven neural interfaces, offering a new technical pathway for practical brain-computer communication, while also raising concerns about neural data privacy and ethics.

Google has released its 11th annual environmental report, demonstrating decoupling between growth and emissions. In 2025, Google signed over 12 gigawatts of net new clean energy, enough to power Greece for a year, bringing its total clean energy portfolio to nearly 35 gigawatts. Despite a 37% increase in electricity demand, operational emissions decreased by 2%, with efficiency measures avoiding 58 million metric tons of CO₂ equivalent. However, supply chain emissions rose 25% due to AI infrastructure expansion and lack of clean energy in Asian power grids. Google acknowledges that AI infrastructure growth outpaces grid decarbonization, with long interconnection queues, fragmented markets, and regulatory bottlenecks delaying carbon-free energy deployment. Meanwhile, Google’s AI products helped partners reduce emissions by approximately 41 million tons—about three times its own operational footprint.

Industry reports indicate that major tech companies have committed over $850 billion to AI data centers, with particularly strong growth from Meta and Microsoft. Enterprise AI budgets continue to rise, with OpenAI remaining the top vendor among CIOs. Meanwhile, due to U.S. export restrictions, non-U.S. enterprises are accelerating in-house development: China’s 360 reportedly launched an AI tool rivaling Anthropic’s Mythos, and Tokyo-based Sakana AI unveiled Fugu, a frontier agent-capable model targeting Mythos Preview. Identity platform Okta has launched a compliance governance service for AI agents, becoming the first independent platform to offer agent lifecycle governance in regulated environments including FedRAMP and HIPAA, supporting identity registration, least-privilege access control, and emergency termination.