AA智能指数60分并列开源第一，成本更低
Key Highlights
Zhipu (GLM) has officially launched the API for GLM-5.3, a new flagship that performs strongly in complex coding, defensive cybersecurity, and long-horizon tasks. What draws even more attention is that it scores 60 on the Artificial Analysis Intelligence Index (commonly called the AA Index in the industry), placing it in the same tier as undisclosed internal flagship models from vendors such as Anthropic and OpenAI, and tying with Kimi K3 as the best-performing open-weight model available today. The result is a rare moment where a Chinese open model sits shoulder to shoulder with the world's most capable closed systems.
What Happened
The biggest selling point of GLM-5.3 is that it lowers the barrier to frontier intelligence further through a smaller parameter size and a lower call cost. According to official disclosures, the model's per-task call cost is the lowest among mainstream flagship models, and its API pricing matches the previous-generation entry-level GLM-5.2. This means developers can obtain near-"next-generation" capability at a price close to the "previous generation." The model weights are scheduled to be open-sourced next Friday, at which point the community can deploy and fine-tune locally, a point of particular importance for domestic teams that require autonomous and controllable capabilities and cannot rely on external APIs for sensitive workloads.
The Technical Details
Looking at its capability profile, GLM-5.3 has built three areas into strengths: first, complex coding that can handle multi-file, cross-repository engineering-level tasks; second, defensive cybersecurity that can assist code auditing and vulnerability hunting; and third, long-horizon tasks that remain more stable in scenarios requiring multi-step reasoning and long-context memory. This combination clearly targets enterprise developers rather than mere chat scenarios, and echoes the recent overall orientation of domestic models toward "practicality and deployment." The emphasis on defensive security also reflects growing enterprise demand for models that can harden, rather than merely generate, code.
Compared with Competitors
At the 60-point tier of the AA Index, GLM-5.3 can already go toe-to-toe with closed-source internal flagships, a position open models rarely reached in the past. Compared with Kimi K3, the two tie for first among open models, but GLM-5.3 is more aggressive on cost control; compared with closed-source APIs that often charge sky-high prices, it directly grabs the small-and-medium enterprise market with an "affordable flagship" pricing strategy. Simply put, Zhipu wants to prove one thing: being open-source does not necessarily mean being in the "second tier." The cost leadership is especially relevant in price-sensitive markets across Asia and the Global South.
Industry Impact and Outlook
GLM-5.3's open-source and affordable strategy will further lower the threshold for enterprises to use frontier models, especially benefiting domestic financial, government, and manufacturing customers that need private deployment with data remaining on-premises. For the developer ecosystem, open-sourcing the weights means a wave of fine-tuning, plugins, and agent applications built around GLM-5.3 will emerge rapidly. As domestic large models accelerate collectively, this combination of "flagship capability plus open-source inclusiveness" is rewriting the landscape in which closed-source models long monopolized the high-end market, and giving more small and medium teams a chance to innovate from the same starting line, ultimately broadening who gets to build with state-of-the-art AI.
The launch also arrives at a pivotal moment for the global open-model landscape. For much of the past two years, the perception persisted that only well-funded, closed labs could reach the top tier of the AA Index; GLM-5.3's result challenges that assumption directly and may encourage other open efforts to aim higher. Its emphasis on defensive cybersecurity is notable given rising enterprise anxiety about AI-assisted attacks, positioning the model as a tool that helps organizations harden their own code rather than merely generate it. The planned weight release will let researchers scrutinize and build upon the model transparently, a dynamic that often accelerates ecosystem improvement faster than closed development. For China's AI industry specifically, GLM-5.3 underscores a strategy of competing on cost and openness rather than raw scale, which resonates in markets sensitive to both price and data sovereignty. If the open-weight release lands as promised, expect a rapid flowering of fine-tunes, domain adapters, and agent frameworks that treat GLM-5.3 as a default local backbone. The longer-term question is whether cost leadership alone can sustain loyalty once competitors match capabilities, but for now, Zhipu has staked out a clear and defensible position at the intersection of performance and accessibility.
The open release also has a signaling effect beyond raw capability. By publishing weights, Zhipu invites global scrutiny and contribution, a move that historically accelerates ecosystem maturation more than closed development alone. It positions GLM-5.3 not just as a product but as a platform others can build upon with confidence in longevity. For an industry wary of vendor lock-in, that openness may prove as decisive as benchmark scores when teams choose a long-term foundation model to standardize on.