GLM-5.3review, pricing and limits

GLM-5.3 succeeds its predecessor as Z.ai's flagship coding model, keeping the same 128,000-token maximum output while requiring one of 3 reasoning-effort levels on every call. It suits agentic coding and security teams who can work through the Coding Plan until standalone API access and open weights ship in late August 2026.

GLM-5.3 is Z.ai's August 2026 post-training refresh of its predecessor, a text-only large language model built for coding agents and cybersecurity vulnerability discovery. It keeps a 1-million-token context window while resetting coding and cyber benchmarks entirely through post-training, with no new pretraining run.

Provider: Z.ai · Family: GLM-5

More about Z.ai on HokAI

Context window: 1,000,000 tokens · Max output: 128,000

Input modalities: text, tool-calls · Output: text, tool-calls

About GLM-5.3

GLM-5.3 is a large language model released by Z.ai (formerly Zhipu AI) on August 14, 2026, as a post-training refresh of GLM-5.2 rather than a new pretraining run. It runs on the same Mixture-of-Experts checkpoint as its predecessor: roughly 744 billion total parameters with about 40 billion active per forward pass. Z.ai shipped it under the tagline "Built to Code. Ready for Cyber Defense," positioning it as the coding- and security-focused member of the GLM-5 lineage rather than a general capability upgrade. Every benchmark gain Z.ai reported came from scaled post-training on a wider set of task environments, not architecture changes. Terminal-Bench 3.0 rose from 4.6 to 28.3, Terminal-Bench 2.1 climbed from 81.0 to 88.2, DeepSWE v1.1 improved from 46.2 to 66.9, SWE-Marathon v1.1 jumped from 19.4 to 42.5, FrontierSWE reached 78.1, and Agents' Last Exam (CLI) moved from 23.8 to 28.5. On Z.ai's private Code Bench at high reasoning effort, GLM-5.3 scored 31.4% using roughly 50,000 tokens, ahead of Claude Opus 4.8's 29.5% result at roughly 120,000 tokens for the same task set. GPT-5.6 Sol still leads on Terminal-Bench 3.0 at 34.6, and Claude Fable 5 tops Code Bench at 39.5%, so the model trails the closed frontier on raw coding capability even as it leads the open-weights track. The model keeps its predecessor's specification: a 1-million-token context window and a 128,000-token maximum output. The one confirmed API-level change is that thinking can no longer be disabled. The prior model accepted a disabled-thinking flag for fast, non-reasoning calls, and this release requires an explicit low, high, or max reasoning-effort setting on every call, defaulting to max. GLM-5.3 is text-only on both input and output. A Z.ai developer poll had returned a unanimous request for vision support in this release, but multimodal input stayed isolated to the separate GLM-5V-Turbo model. Function calling, structured JSON output, streaming, context caching, and Model Context Protocol tool connections are all supported, which matters most for the agentic coding and terminal-automation use cases the model targets. This release has no standalone per-token API price published yet. The only confirmed paid access routes are a tiered monthly Coding Plan subscription and a companion coding client built on top of it; exact subscription pricing and the predecessor's per-token API reference rate are detailed in the pricing FAQ. Deployment is narrow at launch. Z.ai's own docs mark the standalone API as "coming soon" with no date attached, and third-party gateways that carried the predecessor on day one, including OpenRouter, Fireworks, AWS Bedrock, and Google Vertex AI, had not added this release as of the most recent check. A matching Hugging Face repository exists but returns a gated response rather than public weights. The most notable disclosure at launch was safety-related rather than a feature. Z.ai said the model's cyber-exploitation capability "grew faster than it anticipated" during training: GLM-5.3 has already identified 2,436 vulnerabilities across 269 open-source projects, 1,097 rated critical or high severity, with 53 disclosed under CVEs at launch and 2,383 held under embargo on Z.ai's public disclosure ledger at cvd.z.ai. CyberGym accuracy rose from 77.2% to 84.5% and ExploitBench climbed from 24.4% to 54.4%. Unlike GLM-5.2, whose MIT-licensed weights landed on Hugging Face the same week as its announcement, this release's open-weights drop is being held back roughly two weeks, pending a dedicated safety evaluation and hardening pass. GLM-5.3 fits agentic coding teams and researchers chasing long-horizon, terminal-based coding and vulnerability discovery who can work through the Coding Plan while the standalone API and weights are pending. It is a weaker fit for teams needing vision or audio input today, a live per-token API or self-hostable weights right now, or the single best coding score regardless of openness. No system card or independent training-data-cutoff disclosure accompanied the launch, and Z.ai has not published a formal responsible-scaling-policy equivalent to Anthropic's RSP for the GLM-5 line. GLM-5.3 arrived 59 days after GLM-5.2's June 13, 2026 release. Analyst reporting, including a JPMorgan note echoed by Reuters and CGTN, points to a larger successor tentatively called GLM-5.5 at over 1 trillion parameters, though Z.ai had not confirmed a release date for it as of this launch.

Pricing

Z.ai has not published a standalone per-token API price for GLM-5.3. The only confirmed paid access route is the GLM Coding Plan subscription (Lite $18/month, Pro roughly $72 to $80/month, Max roughly $160 to $168/month) or the ZCode client built on top of it. For reference, the predecessor's direct API charged $1.40 per 1M input tokens and $4.40 per 1M output tokens, with $0.26 per 1M cached input.

Key Features

• 744B MoE Architecture: The same Mixture-of-Experts checkpoint as its predecessor: about 744B total parameters, an estimated 40B active per token, with every capability gain coming from this release's post-training pass rather than new pretraining.
• Post-Trained Coding Jump: Its toughest terminal-coding benchmark rose from 4.6 to 28.3, and a separate agentic-coding benchmark climbed from 46.2 to 66.9, purely from extended post-training on the same base checkpoint as its predecessor.
• Cybersecurity and Vulnerability Discovery: CyberGym vulnerability detection climbed to 84.5% (from 77.2%) and ExploitBench reached 54.4% (from 24.4%), reflecting this release's security-agent focus.
• Mandatory Reasoning-Effort Levels: Every call now runs one of 3 reasoning-effort levels (low, high, max); the old option to disable thinking entirely has been removed.
• MCP and Structured Output: Supports Model Context Protocol tool connections, function calling, JSON structured output, streaming, and context caching, matching the integration surface professional coding agents expect.

Pros

• Coding capability improved sharply through post-training alone, roughly 6x higher on its toughest terminal-coding benchmark than the prior release, with no new pretraining run required.
• Matches or beats at least one closed rival on token efficiency: the model's 31.4% score on Z.ai's Code Bench needed only about 50,000 tokens, well under the roughly 120,000 Claude Opus 4.8 used for its 29.5%.
• Keeps the predecessor's context window and maximum output unchanged, so existing long-context coding workflows carry over without adjustment.
• Z.ai has a track record of shipping this line under an MIT license: the predecessor's weights went up on Hugging Face within days of its own announcement, which would make this release unusually permissive for a frontier-class coding model once weights land.

Cons

• Standalone per-token API pricing has not been published yet; access requires a paid Coding Plan subscription instead of pay-as-you-go billing.
• Open weights are withheld pending a safety review, arriving roughly two weeks behind the predecessor's same-week Hugging Face release, so self-hosting or fine-tuning isn't possible yet.
• Text-only input and output despite a unanimous developer request for vision support, and no system card was published at launch.
• Trails closed frontier rivals on raw coding scores: one competitor leads its toughest terminal benchmark at 34.6, and another tops Z.ai's Code Bench at 39.5%, both ahead of this release.

Benchmarks

• cybergym: 84.5
• frontierswe: 78.1
• deepswe v1 1: 66.9
• exploitbench: 54.4
• swe marathon v1 1: 42.5
• terminal bench 2 1: 88.2
• terminal bench 3 0: 28.3
• agents last exam cli: 28.5
• code bench internal high effort: 31.4

Visit GLM-5.3 Official Page