{
  "id": 882103,
  "title": "GLM-5.3 didn’t change the base model — where did its coding gains come from?",
  "url": "https://urgent.news/2026/08/14/glm-5-3-didnt-change-the-base-model-where-did-its-coding-gains-come",
  "topic": "ai",
  "section": "AI",
  "published": "2026-08-14T15:20:24.000Z",
  "source": {
    "name": "The New Stack",
    "slug": "the-new-stack",
    "url": "https://thenewstack.io/glm-5-3-post-training-coding/"
  },
  "original_language": "en",
  "account": "Z.ai unveiled GLM-5.3 on Friday, a coding and agent model derived from the same base model as GLM-5.2. This model is accessible through Z.ai's GLM Coding Plan with Claude Code, Claude, OpenCode, and Codex. However, direct API access is anticipated to launch soon, pending additional safety testing and two weeks of model hardening. Post-training played a pivotal role, as Z.ai subjected GLM-5.3 to tenfold more extended, long-horizon task environments, along with broader developer tool and engineering workflow access. Certain training tasks mirrored the entire software lifecycle, from bug identification and fix drafting to code writing, testing, and deployment. According to Z.ai, individual tasks equated to the workload of a senior engineer over several days. By focusing compute on the actual environments the model functions in, GLM-5.3 emerges as a compelling post-training compute scaling case study. Z.ai stresses the importance of this approach, asserting it enhances the model's practical applicability. While Z.ai claims GLM-5.3 offers a 50% performance boost over GLM-5.2 on its internal Code Bench, the community should exercise caution until the released weights undergo community testing. On public evaluations, GLM-5.3 demonstrates remarkable gains in agentic coding, achieving a score of 28.3 on Terminal-Bench 3.0 (up from 4.6), 66.9 on DeepSWE v1.1 (from 46.2), and 28.5 on Agents' Last Exam (from 23.8). This DeepSWE score places GLM-5.3 in the vicinity of Google's Gemini 3.7 Flash (65%). However, disparities in testing harnesses necessitate cautious head-to-head comparisons. GLM-5.3 boasts a 1-million-token context window and a 128,000-token completion limit, enabling it to manage large codebases. Developers can configure the larger context window on Claude Code via the glm-5.3[1m] model tag and corresponding 1M-token compaction window. Reasoning effort is adjustable, ranging from low to max (default is max). Although recommended for non-trivial engineering tasks, max reasoning incurs noticeable latency and token overhead. Teams must weigh the downstream accuracy gains against the increased compute costs. Z.ai claims the model strikes a balance between latency, token efficiency, and correctness; however, official per-token API pricing has not been disclosed. While user-friendly for developers eager to leverage the latest model, this feature creates a blind spot for teams attempting to compare results against previous versions using the same plan. This reflects ongoing market recalibration around agent economics, influenced by OpenAI's price reductions and Microsoft's implementation of token caps to curb runaway autonomous agents. Pricing long-horizon agentic compute remains a fluid market dynamic. Z.ai's Coding Plan charges higher multipliers for input, cached-input, and output tokens compared to GLM-4.7. However, these costs are mitigated by a 50% discount during off-peak hours. A noteworthy migration consideration is Z.ai's automatic redirection of calls to GLM-5.2 or GLM-5.1 to GLM-5.3. While convenient for developers seeking the latest model, it poses challenges for teams conducting clean A/B comparisons against previous versions on the same plan. Developers must verify the actual model ID returned by Z.ai's agent to ensure accurate comparisons. Z.ai highlights cybersecurity as an emerging strength for GLM-5.3, noting a 84.5% score on CyberGym, an improvement from 77.2% in GLM-5.2. This score narrowly surpassed Mythos 5 (83.8%) and GPT-5.6 Sol (83.6%). However, GLM-5.3's performance in exploiting found vulnerabilities remains a concern, with an ExploitBench score of 54.4%, significantly higher than Mythos 5 (78%) and GPT-5.6 Sol (76.5%). This suggests that while GLM-5.3 excels at identifying and reviewing vulnerabilities, its ability to complete deeper offensive security tasks lags behind. It is essential to approach these high scores with skepticism, as crafting an exploit, proving its real-world reach in production, or patching the flaw without triggering downstream regressions represents a distinct challenge altogether. The actual weights for GLM-5.3 will become available in two weeks, allowing developers to test if the benchmark improvements translate to local deployments, different inference stacks, and non-Z.ai repositories. This wait mirrors the pattern seen with other frontier AI labs, such as Moonshot's Kimi K3, where there is a notable gap between announcing open weights and making them accessible for actual use.",
  "summary": "Z.ai released GLM-5.3 on Friday, a coding and agent model built from the same base model as GLM-5.2. Developers can The post GLM-5.3 didn’t change the base model — where did its coding gains come from? appeared first on The New Stack .",
  "key_points": [
    "GLM-5.3 built upon GLM-5.2 base model, not changing original structure",
    "Post-training extended tasks with developer tools, akin to senior engineer workload",
    "GLM-5.3 shows 50% performance boost on internal Code Bench, excels in agentic coding"
  ],
  "editors_take": null,
  "illustration": null,
  "coverage": {
    "outlets": 1,
    "also_reported_by": []
  },
  "ai_generated": true,
  "disclaimer": "Summaries, key points and the editor’s take are written by software from other outlets’ reporting and may contain errors — always check the linked original."
}