The global artificial intelligence landscape moves with major announcements and challenges in July 2026, highlighting the continued delay of Google’s Gemini 3.5 Pro. In the same period, China implements strict new laws for AI companion services, forcing giants like ByteDance and Alibaba to disable features for millions of users. At the same time, Tesla expands its Robotaxi service to Miami, operating without safety monitors and sparking debates about vehicle autonomy. Meanwhile, the White House is preparing voluntary standards for AI models in an effort to guide rapid technological advancement.
Delays in Gemini 3.5 Pro and growing competitiveness in the sector
Gemini 3.5 Pro, Google’s artificial intelligence model, enters its second week of July 2026, still in the preview phase for companies, with no confirmed date for general launch. The model has already missed two delivery dates, initially promised for June and with a deadline of June 30 for general availability. The rationale for the latest postponement points to the need to optimize token efficiency in complex tasks and improve performance in long-term reasoning, based on feedback from early testers.
This scenario occurs amid an environment of intense competition. Recently, Anthropic released Claude Sonnet 5 and restored Claude Fable 5, while OpenAI previewed GPT-5.6 for government partners. Google’s urgency in delivering a model with clearly differentiated performance, especially in context retrieval and complex reasoning, is vital to justify the postponements and reverse the negative perception that accompanied the company’s month of June. Initial tests revealed that the model consumed more tokens than expected, making it more expensive to operate on a large scale, and showed gaps in coding performance and multi-step reasoning tasks.
Chinese legislation impacts AI agents: Doubao and Qwen adapt
China is preparing to implement on July 15 its “Provisional Measures for the Administration of Interactive Anthropomorphic AI Services,” legislation that requires AI services with human personality to incorporate anti-addiction systems, mandatory usage notifications, and instant exit mechanisms. This new regulation directly impacts platforms such as Doubao, from ByteDance, and Qwen, from Alibaba, which are deactivating their humanoid and user-created agent functionalities.
Doubao, the most used AI app in China with 345 million monthly active users, will allow view-only access to settings and chat history until October 15, when the data will be permanently inaccessible. Users must export their important content before the deadline. Qwen’s situation is even more critical as Alibaba has not announced any migration path for users with existing agent setups, raising the risk of immediate and permanent data loss. The decision by both companies to disable the features rather than adapt them reflects the fundamental incompatibility between the law’s anti-addiction rules and the functioning of persistent memory agents.
Tesla’s aggressive expansion with Robotaxi and the risks of unsupervised autonomy
Tesla has expanded its Robotaxi service to Miami, Florida, marking the fifth US city to receive the technology, operating without safety monitors on board the vehicle. This is the first time the company has launched autonomous operation without any period of prior supervision, following Florida state regulations. Tesla’s goal is to expand to a dozen US states by the end of 2026, with CEO Elon Musk arguing that Full Self-Driving (FSD) technology has surpassed the safety threshold for commercial deployment.
In contrast, Waymo, the leading robotaxi operator in the US, maintains safety monitors in new markets and adopts a more cautious expansion cadence based on detailed mapping and supervised deployment phases. Tesla’s approach, while promising for rapid cycles of improvement due to its vast fleet, raises concerns from regulators and consumer advocates, who warn of the absence of “human error” should the autonomous system encounter unforeseen situations. This Tesla deployment model is considered one of the most aggressive in the history of autonomous artificial intelligence without human supervision in public consumption contexts.
New benchmarks and the future of artificial intelligence in specialized areas
The development of more sophisticated benchmarks is revealing both the advancement and limitations of AI models. Meta launched SWE-Together, a multi-shift coding agent assessment tool with 109 tasks. This benchmark, which simulates real software engineering sessions, evaluates models’ ability to maintain context, adapt to feedback, and course correct over multiple interactions. Anthropic’s Claude Opus 4.8 demonstrated the least need for human intervention, with 63% correct completion on the first try.
On the other hand, OpenAI introduced GeneBench-Pro, a 129-problem computational biology benchmark covering genomic analysis, protein folding, and experimental design. The most significant finding is that GPT-5.6 Sol Pro, the most advanced version of OpenAI’s models, only achieved 31.5% accuracy. Claude Opus 4.8 achieved 16%. These results underscore the challenge for artificial intelligence in mastering specialized doctoral-level scientific knowledge, which is not yet fully captured by the training pipelines of current frontier models.
The global AI landscape: standards, Chinese models and strategic infrastructure
The artificial intelligence landscape is not just about new models and regulations, but also about the evolution of global standards and the infrastructure that supports them. The White House is expected to announce a voluntary standards framework for frontier AI models this week, establishing criteria for 30-day government review before public release. This movement, which includes Google’s participation in the negotiations, aims to create a precedent for coordination between the government and technology companies.
Meanwhile, Chinese AI models are taking a significant share of traffic on platforms like OpenRouter, accounting for about 45% of the total, compared to less than 2% a year ago. The success is not just due to performance benchmarks, but a combination of competitive pricing, 1 million token context windows and free access. This “cost arbitrage” is transforming the landscape for developers and enterprises. However, the use of these models requires the assessment of three important gaps:
- Content restrictions:Chinese models often refuse politically sensitive topics such as Taiwan, Tiananmen Square and Xinjiang.
- Data Jurisdiction:API calls can route data through servers in China, which could violate data residency requirements for regulated industries.
- Rigidity of the tool-call scheme:There are cases of non-standard JSON in tool calls, requiring additional handling in agent frameworks.
At the corporate level, Alibaba restructured its entire AI operation into what it calls the “Alibaba Token Hub”, consolidating five units under the direct supervision of CEO Eddie Wu. The restructuring reflects strategic recognition that Alibaba’s value creation in AI derives from tokenization at scale, aiming to compete for the Chinese domestic market and international developer adoption.
Building data centers in the US is also seen as a supply chain policy, going beyond a mere investment in AI capacity. Massive investments from companies like Microsoft and Alphabet are being framed as demand anchors that can structure domestic manufacturing chains for semiconductors, power systems, cooling infrastructure and fiber optic networks. This approach links industrial policy directly to the development of AI infrastructure, seeking to direct spending towards US-made equipment.
