Z.ai released GLM-5.3 on Friday, a coding and agent model that jumped from 46.2 to 66.9 on DeepSWE v1.1, a benchmark measuring software engineering tasks. The company built the model from the same base architecture as its predecessor, GLM-5.2, but dramatically expanded post-training to expose the model to tenfold more long-horizon task environments while widening its access to developer tools and engineering workflows. Developers can already use GLM-5.3 through Z.ai's GLM Coding Plan with Claude Code, Cline, OpenCode, and Codex, though direct API access remains listed as "coming soon" because the company plans to release model weights after two weeks of hardening and safety testing.
On public evaluations, the model posted dramatic gains in agentic coding. GLM-5.3 surged to 28.3 on Terminal-Bench 3.0, up from 4.6 in the prior version, and climbed from 23.8 to 28.5 on Agents' Last Exam. The 66.9 DeepSWE score lands alongside Google's Gemini 3.7 Flash at 65%, though variations in testing harnesses mean head-to-head comparisons carry caveats. Z.ai claims a 50% performance boost over GLM-5.2 on its internal Code Bench, though self-reported vendor numbers warrant skepticism until the community tests the released weights. On cybersecurity benchmarks, GLM-5.3 scored 84.5% on CyberGym, up from 77.2%, narrowly exceeding Mythos 5 at 83.8% and GPT-5.6 Sol at 83.6%. Its ExploitBench score more than doubled from 24.4% to 54.4%, but still trailed Mythos 5 at 78% and GPT-5.6 Sol at 76.5%. The model operates with a 1-million-token context window and a 128,000-token completion ceiling, and developers on Claude Code can configure the larger window via the glm-5.3[1m] model tag. Reasoning effort is configurable across low, high, and max levels, with max enabled by default.
According to Z.ai, some training tasks simulated the full software lifecycle—from identifying bugs and drafting fixes to writing code, running tests, and shipping results—with single tasks matching the workload of a senior engineer over several days. The company concentrated compute on the specific environments the model actually works in, making GLM-5.3 a case study for post-training compute scaling rather than inflating parameter counts. Z.ai acknowledges the model is currently stronger at the earlier stages of the exploitation chain, including reviewing source code and verifying vulnerabilities, than at completing deeper offensive security tasks. The company's documentation notes that Coding Plan calls to GLM-5.2 or GLM-5.1 are automatically redirected to GLM-5.3, creating a blind spot for teams trying to run clean A/B comparisons against previous versions on the same plan.
The more consequential release is scheduled for two weeks from now, when Z.ai will make the weights available for developers to test whether the benchmark improvements carry over to local deployments, different inference stacks, and repositories the company didn't select. On Z.ai's Coding Plan, usage is metered via credits, with GLM-5.3 carrying higher baseline multipliers for input, cached-input, and output tokens compared to GLM-4.7, offset by a 50% off-peak discount, though official per-token API pricing has not yet been published. Although max reasoning is recommended for non-trivial engineering tasks, it introduces noticeable latency and token overhead, and teams must evaluate whether the downstream accuracy justifies the added compute cost. This absence of transparent pricing reflects ongoing market recalibration around agent economics, between OpenAI's aggressive price drops and Microsoft introducing token caps to mitigate runaway autonomous agents. The wait for weights has become a recurring pattern in the open-model space—Moonshot released Kimi K3's weights under similar conditions, and the gap between "open weights announced" and "weights you can actually run" is now familiar territory. As frontier labs continue building models specifically for cybersecurity work and delaying releases over offensive capability concerns, the pattern Z.ai's results follow suggests the industry is learning to distinguish between finding vulnerabilities and weaponizing them. The real test arrives when independent developers can measure whether post-training intensity translates to reliability on codebases Z.ai never trained against.