Z. ai’s GLM-5.3 has emerged as the more striking of the two launches because of its ability to identify software vulnerabilities. The company said the model scored 84.5 per cent on CyberGym, a benchmark that tests whether AI systems can inspect code, locate security weaknesses and verify them. Anthropic’s restricted-access Mythos 5 scored 83.8 per cent in the comparison, although Z. ai’s figures have not been independently verified.
The results do not mean GLM-5.3 has matched Mythos 5 across offensive cybersecurity tasks. On ExploitBench, which measures whether identified vulnerabilities can be turned into working attacks, GLM-5.3 scored 54.4 per cent against 78 per cent for Mythos 5.
The gap was also visible in timed attack-development evaluations. Z. ai said its model completed 105 tasks in two hours and 130 over six hours, compared with 181 and 247 respectively for Mythos 5. The figures suggest GLM-5.3 is particularly strong at vulnerability discovery but remains significantly weaker when required to develop exploits from those weaknesses.
That distinction matters as AI laboratories wrestle with the dual-use character of advanced coding systems. Models capable of spotting flaws can help developers and security teams repair vulnerable software, but the same capabilities can lower the technical barrier for attackers seeking exploitable weaknesses.
Z. ai has therefore delayed the full public release of GLM-5.3 for roughly two weeks while it carries out additional security assessments. Particularly sensitive cybersecurity functions are expected to be restricted to verified users under a trusted-access programme, while selected launch partners receive earlier access.
The approach marks a shift in China’s open-model ecosystem, where rapid releases and broadly downloadable weights have become important competitive tools. Z. ai has also announced an Open Source Shield initiative designed to audit selected open-source projects, provide defensive access to its models and integrate code-security features into its ZCode development product.
GLM-5.3 uses the same underlying base model as GLM-5.2, with much of the improvement coming through additional post-training. Z. ai says its coding performance has improved by about 50 per cent on its internal benchmark and that extended reinforcement learning produced capabilities extending beyond conventional software development into cybersecurity work.
DeepSeek, meanwhile, has formally launched V4 Pro, an upgraded version of the large open-weight model family first previewed in April. The August 13 release places greater emphasis on autonomous agents, coding, tool use and complex reasoning.
Independent testing gives the reasoning version of V4 Pro a score of 53 on the Artificial Analysis Intelligence Index, which combines evaluations across coding, scientific reasoning, long-context tasks, tool use and agentic workloads. The result represents a notable improvement over earlier DeepSeek V4 configurations and places the model among stronger open-weight systems.
V4 Pro is built on a mixture-of-experts architecture containing about 1.6 trillion total parameters while activating roughly 49 billion for each token. It supports a context window of one million tokens, allowing developers to process unusually large codebases, documents and multi-stage workflows within a single session.
The improved capability carries a substantially higher price. V4 Pro costs $1.32 per million input tokens and $3.96 per million output tokens at standard peak rates. DeepSeek is introducing off-peak pricing at half those rates, creating an incentive for developers running flexible workloads to shift processing outside periods of heavier demand.
Those prices make Pro several times more expensive than V4 Flash, DeepSeek’s smaller and faster model. Flash had previously produced unexpectedly strong benchmark results against an earlier Pro preview, raising questions about whether customers would pay more for the flagship version. The August upgrade has restored a clearer performance advantage for Pro, although Flash remains faster and considerably cheaper.
DeepSeek’s position has become harder to defend as China’s model market crowds with increasingly capable systems from Z. ai, Moonshot AI, Alibaba, MiniMax, ByteDance and other developers. The company that shook the global AI industry with R1 in early 2025 is consequently expanding beyond low-cost reasoning models towards agents, long-context applications, commercial APIs and specialised infrastructure.
Follow Arabian Post
Select Arabian Post as your preferred source on Google and MSN News for trusted business news and Arab politics and updates.
