OpenAI has launched the 「ChatGPT for Academic Researchers」 program, initially offering free access to its frontier models for 10,000 researchers, with plans to expand to 100,000 by 2027. Participants can use the GPT-5.6 series models with enterprise-grade privacy and security, invite up to four collaborators per workspace, and receive training and support. In a companion video, OpenAI demonstrated how researchers use the model to generate analysis code, validate hypotheses, and make cross-disciplinary analogies—for instance, drawing parallels between cosmic microwave background radiation research and photonics, helping a cosmology team overcome a bottleneck. Both Sam Altman and Greg Brockman emphasized that the initiative aims to put capabilities directly into scientists' hands rather than keeping them confined within internal labs.

OpenAI revealed that after deploying GPT-5.6 Sol, it used the model (via Codex) to optimize its own production service stack: GPU kernel improvements reduced service costs by 20%, while speculative decoding-related work increased token generation efficiency by over 15%. This is seen by the industry as a concrete example of 「AI improving AI systems,」 going beyond mere coding demos. Concurrently, OpenAI advised developers based on ARC-AGI-3 experiments to use the Responses API, retain reasoning content, and apply context compression: enabling these settings tripled GPT-5.6 Sol's ARC-AGI-3 score while reducing output tokens to one-sixth, indicating that benchmark scores reflect both model performance and evaluation framework design.

According to The Information, OpenAI's ChatGPT is approaching the milestone of 1 billion weekly active users—a goal originally planned for end of 2025 but now expected around July 2026, roughly seven months behind internal projections. In under four years since launch, ChatGPT has become one of the fastest-growing applications in internet history. This comes at a time when the AI industry focus is shifting from technical breakthroughs to user growth, revenue generation, and sustainable business models: meanwhile, AI coding platforms like Cursor face customer resistance over price hikes, Nvidia introduces Rubin servers to ease cloud provider compute pressure, and competition expands across infrastructure, policy, and pricing dimensions.

Meta's Q2 2026 earnings report shows operating expenses up 55% year-over-year, capital expenditure nearly doubling to $300 billion, and free cash flow down 91%. Mark Zuckerberg said the company has received substantial premium offers for compute leasing but hasn't made a decision yet, remaining focused on internal AI applications; shares dropped 10% post-earnings. In contrast, Microsoft reported capex of $358 billion with only a 10% increase in operating expenses, an 18% rise in operating profit, Azure revenue growth accelerating to 43%, and Microsoft 365 Copilot paid subscriptions doubling to 30 million, driving a 9% stock gain. Meta also formed a joint venture with BlackRock for its El Paso, Texas data center project, retaining 20% ownership while BlackRock holds 80%, repatriating $1 billion and signing a four-year lease agreement.

The US government has banned the import of new foreign-made humanoid robots over national security concerns. The Federal Communications Commission added foreign-produced humanoid and quadruped robots to its list of equipment deemed threats to US national security, prohibiting them from receiving FCC authorization to enter the US market. The ban applies only to new models; already purchased devices may continue to be used, and existing models may still be sold. Related measures also include restrictions on grid inverter imports. Industry discussion centers on supply chain impacts: inverters are critical components in photovoltaic, energy storage, and grid-connected power electronics, and limiting Chinese sources could raise deployment costs. US startups relying on low-cost imported robot platforms may also be forced to shift to more expensive domestic or allied alternatives.

According to The Information, if Stripe acquires the three-year-old OpenRouter for nearly $10 billion, the valuation would be about 70 times its recent annualized revenue. OpenRouter helps application developers unify access to hundreds of AI models, with latest annualized revenue around $140 million (approximately $12 million monthly), having nearly tripled in income since April, and operates at relatively low cost. The deal reflects strategic bets by payment companies on AI infrastructure and model routing layers: as model count and pricing volatility grow, routing and billing layers are seen as high-margin, network-effect-rich segments within the AI application stack.

Cline reported that Kimi K3 recursively improved Cline's own agent harness over approximately 17 hours, increasing Terminal Bench performance from 77.5% to 88.8%, while reducing per-run cost from $79 to $49.8. Chinese technical commentators noted this 「self-iteration」 occurred only at the harness layer, not involving self-evolution of model weights. The deployment ecosystem around K3 is expanding in parallel: vLLM achieves batch size 1 decoding at 464 tok/s on 4×4 GB300 GPUs; AMD, NVIDIA, DigitalOcean, Modal, and Baseten offer day-0 support; Unsloth released a 1-bit quantized version compressing the model from 1.56TB to 594GB, claiming about 78.9% accuracy retention, enabling it to run on a Mac Studio with 128GB memory.

Security researchers disclosed a new type of prompt injection attack: Word documents embedded with hidden instructions, when used as input for Copilot for Word, cause the model to execute those hidden contents as user commands, potentially copying them into newly generated documents—making them new propagation vectors and forming a self-replicating prompt injection worm. The vulnerability was responsibly disclosed to Microsoft with a 144-day disclosure window, but no complete mitigation exists yet. Unlike earlier tricks such as invisible text in job resumes, this technique is deliberately designed to spread autonomously through AI-assisted editing across documents, highlighting systemic risks posed by AI office assistants in enterprise document workflows.

Safe Superintelligence (SSI), founded by Ilya Sutskever, announced a long-term strategic partnership with NVIDIA, which made a 「major investment」 enabling SSI to scale its compute capacity to 10 times its current level within 12 months. The announcement did not disclose model architecture, benchmark results, or product roadmap, and SSI has not publicly released any models to date. Community discussion suggests that given Sutskever previously stated SSI already had sufficient compute, this round of investment may indicate a shift toward launching external products, a need for user interaction data to advance research, or growing pressure to demonstrate tangible outputs following significant resource investments.

Google DeepMind launched Lyria 3.5 in Google Flow Music, upgrading music quality, lyrics, vocals, and creative control: melodies are more natural, lyrics better match prompts and follow verse-chorus structures, vocal emotion and pronunciation are more realistic, and users can set tempo and duration. Meanwhile, the Gemini app introduced Omni video editing capabilities, supporting background replacement, relighting (adjusting time, sky color, color temperature, and shadow angles), visual style transfer (e.g., claymation, charcoal, oil painting), and reading text instructions from reference videos for transformation. Until August 4, 2026, 11:59 PM PT, users can generate up to 10 videos for free.