China Securities Co., Ltd.: The release of GPT-5 and the open-source of Huawei CANN are expected to drive the development of AI applications.

China Securities Co., Ltd. released a research report stating that GPT-5 has been officially released, providing better soil for the growth of AI applications in long-text memory, reducing hallucination rates, and optimizing reasoning efficiency. Subsequent updates can be expected from Google, Anthropic, and domestic models, as well as continued implementation of AI application tokenization under base model optimization. Huawei has fully open-sourced the CANN ecosystem, providing significantly improved development efficiency, and is expected to compete with CUDA acceleration. 1) Favorable for software companies with data, customers, and scenarios, AI products are expected to drive company ARPU increase and project unit price rise; 2) Increasing demand for model privatization, benefiting all-in-one machines, hyper-converged systems, and B-side service outsourcing companies; 3) The market trading volume continues to maintain high levels, with internet financial targets expected to benefit. Main points from China Securities Co., Ltd.: GPT-5 has been officially released, refreshing multiple abilities SOTA. In the early morning of August 8, 2025, OpenAI officially released GPT-5, including three versions: GPT-5, GPT-5-Mini, and GPT-5-Nano. In terms of performance, GPT-5 refreshes multiple benchmark tests SOTA, especially excelling in mathematics, programming, visual understanding, and health sectors, and achieving the top ranking in all dimensions on the large model competition LMArena. Mathematics: OpenAI believes that Benchmark scores using tools should not be compared with models without tool access capabilities. GPT-5 scored 94.6%/100% on AIME 2025 no tools/Python, reflecting its effective use of tools to a certain extent. Additionally, GPT-5 performed excellently on expert-level problems, achieving a score of 100% on HMMT (Massachusetts Institute of Technology Mathematics Tournament); 89.4% on GPQA Diamond (Doctoral level scientific issues); and 42.0% on HLE (Interdisciplinary expert issues). Programming: GPT-5 scored 74.9% on SWE-bench Verified (With thinking), slightly surpassing Claude Opus 4.1's 74.5%, setting a new SOTA. In addition, GPT-5 scored 1479 points in WebDev on all capability scores of LMArena, significantly ahead of the second-ranked Gemini-2.5-Pro (1403 points). Multimodal: GPT-5 scored 84.2% on visual understanding in MMMU (With thinking) (higher than Gemini 2.5 pro's 81.7%, but both scored equally in LMArena, at 1253 points and ranked first), with varying degrees of improvement over previous models on MMMU-Pro, VideoMMMU, and other benchmarks. Scenarios: In education, GPT-5 can generate hundreds of lines of code in a few minutes, generating interactive content to explain complex concepts; create a French learning webpage in a few minutes to help users practice pronunciation, with more natural speech tones than previous generations; healthcare is a key application scenario for GPT-5, as OpenAI terms it as the best health model to date. In tasks of economic significance spanning over 40 professions including law, logistics, sales, and engineering, GPT-5 has reached expert equivalent or higher levels in 47.1% of scenarios. GPT-5 has a context window of 400K, and can decide the applicable model through a real-time router. Apart from scoring high in initial benchmark tests, GPT-5 is a unified model with a real-time router that can quickly determine which model to use based on user conversation type, complexity, tool requirements, and explicit intent (similar to self-determined fast/slow thinking in a mixed reasoning model). In addition, GPT-5's context window has been increased to 400K tokens (approximately 300,000 words, doubling o3), and the "Memory" function will add access to Gmail and Google Calendar (to be launched next week, initially open to professional users, followed by Plus/Team/Enterprise users), helping users plan schedules according to their preferences. Hallucination rates significantly reduced, with noticeable cost optimization. In anonymous testing of GPT-5's production environment flow in ChatGPT, the fact error rate of GPT-5 is approximately 45% lower than o4; when reasoning function is enabled, the error rate is about 80% lower than o3. In LongFact and FActScore tests of open fact accuracy benchmarks, GPT-5 thinking reduces the hallucination rate by about six times compared to o3; furthermore, GPT-5 further reduces the issue of overconfidence in models, for example, deceiving rates in multimodal problems with non-existent images decreased from 86.7% in o3 to 9%. In terms of pricing and cost, GPT-5 is available for all users, with higher usage for Plus users, and Pro users can use the thinking-enabled GPT-5-Pro. The cost for GPT-5 is $1.25 per million tokens input ($0.125 for cache hit), and $10 for output, on par with Gemini 2.5Pro; Mini and Nano versions are 1/5 and 1/25 of GPT-5 respectively. Furthermore, GPT-5 also shows improved reasoning efficiency, with better performance in some tasks than o3, albeit with a 50-80% reduction in output tokens. Huawei's CANN fully open-sourced, accelerating ecosystem improvement. On August 5th, at the Huawei Ascend Computing Industry Development Summit, Huawei's rotating chairman Xu Zhijun announced that Huawei's Ascend hardware enables CANN to be fully open-sourced and open, with the Mind series application enabling suite and toolchain fully open-sourced, supporting users in independent deep mining and custom development, accelerating the pace of innovation for developers, making Ascend more user-friendly and usable. Specifically, CANN (Neural Network Heterogeneous Computing Architecture) is a software ecosystem composed of various skill stacks and operator acceleration libraries, connecting upper-layer AI training frameworks (such as PyTorch, TensorFlow, MindSpore, etc.) with the underlying Ascend chips, allowing developers to call the underlying computing power without worrying about chip details, thereby accelerating the construction of the Ascend ecosystem. Differentiated openness for developers at various levels significantly improves development efficiency. Ascend CANN has launched a differentiated open strategy for developers at different technical levels. For example, at the algorithm innovation level, developers can conduct "magic-style" innovation based on business scenarios; at the system optimization level, CANN has opened up over 190 underlying interfaces, enabling the release of atomic-level capabilities of hardware. At the compilation ecosystem level, by opening the AscendNPU IR interface of the Bisheng compiler, developers can directly use Python syntax to write high-performance operators after accessing Triton, completely abstracting the underlying hardware differences. In terms of efficiency, CANN improves efficiency through optimization of computing, memory, and communication. On the computing front, Ascend CANN has launched the super operator MLAPO for MoE large model training scenarios, compressing the MLA preprocessing time from the industry average of 109ms to 45ms, achieving over 20% end-to-end performance improvement in financial risk models and advertising recommendation systems, breaking the computational bottleneck of the Transformer architecture; in terms of memory, multiple address mapping technology can automatically splice and utilize memory fragments, solving the "memory fragmentation" problem in dynamic shape scenarios, improving memory utilization rates for business scenarios such as JD.com product search and Tencent short video recommendations by over 20%, breaking the industry limit of single-card concurrent request volume. In the field of distributed training, the NPUDirect communication algorithm reduces the 3 pairs of synchronous operations required for traditional RDMA communication to 1 atomic operation, reducing the cross-machine communication delay of Iflytek Co., Ltd. Voice's large model by 90%, paving the way for the feasibility of large-scale distributed training with thousands of cards. According to official Ascend data, over the past 6 years, over 600,000 developers have learned about CANN, used CANN, and infused CANN with innovative vitality. With the continuous advancement of the Ascend CANN open ecosystem and active participation from developers and enterprises in the industry, the CANN ecosystem is rapidly growing. Currently, there are over 6000 operator certification developers for CANN. With further open-sourcing of CANN, the Ascend CANN ecosystem is expected to further improve, catching up with CUDA acceleration and providing a better foundation for AI applications based on domestic computing power. In summary, GPT-5 has been officially released, providing a better foundation for the growth of AI applications in long-text memory, reducing hallucination rates, and optimizing reasoning efficiency. Subsequent updates can be expected from Google, Anthropic, and domestic models, as well as continued implementation of AI application tokenization under base model optimization. Huawei has fully open-sourced the CANN ecosystem, providing significantly improved development efficiency, expected to compete with CUDA acceleration, and similarly providing a better foundation for domestic AI application development. The overseas launch of GPT-5 reduces hallucination rates, improves reasoning performance, and reduces costs. 1) Favorable for software companies with data, customers, and scenarios, AI products are expected to drive company ARPU increase and project unit price rise; 2) Increasing demand for model privatization, benefiting all-in-one machines, hyper-converged systems, and B-side service outsourcing companies. Risk factors: (1) Downside risk of macroeconomic conditions; (2) Bad debt risk of accounts receivable; (3) Increased industry competition; (4) Impact of changes in the international environment.