Z.ai releases GLM-5.3 API with 1M-token context

• GLM-5.3 is now available through a hosted API and partner gateways, with a 1M-token context window, according to Z.ai's API announcement and OpenRouter's launch notice.
• The release keeps GLM-5.2's base model and attributes the capability jump entirely to post-training, as OpenRouter's launch notice describes.
• Artificial Analysis scored GLM-5.3 Max at 60, up seven points from GLM-5.2 and level with Kimi K3, in Artificial Analysis's evaluation.
• Z.ai reported large vendor-run jumps in terminal coding, SWE, and cyber tasks, including the Terminal-Bench 3.0 result that the API feature rundown put at 4.6% to 28.3%.
• Public weights remain on a two-week safety and hardening schedule in Z.ai's release note, while the hosted model is already live through Z.ai's API announcement.

Calls that still send thinking.type: "disabled" will fail after a migration to GLM-5.3, according to Z.ai's API migration notes. The same release note describes judge agents testing whether synthetic RL environments are solvable, while Artificial Analysis's model page records 170 million output tokens across its Intelligence Index run.

• Access: Z.ai's first-party API and partner gateways are live, as Z.ai's API announcement says.
• Limits: the text-only model has a 1M-token context window and a 128K output ceiling in Z.ai's release documentation.
• Pricing: input is $1.40 per million tokens, output is $4.40, and cached input is $0.26, the unchanged rate cited by Z.ai's API announcement and listed by OpenRouter.
• Reasoning: low, high, and max are available, max is the default, and thinking is always enabled in Z.ai's migration notes.
• Compatibility: OpenAI- and Anthropic-compatible APIs are part of the API feature rundown.
• Weights: Z.ai says Hugging Face availability follows its two-week safety evaluation and hardening period in the official release.

First-party

• Terminal-Bench 3.0: 4.6% → 28.3%, +23.7 points, in Z.ai's evaluation table.
• DeepSWE v1.1: 46.2% → 66.9%, +20.7 points, according to OpenRouter's launch notice.
• ExploitBench: 24.4% → 54.4%, +30.0 points, in Z.ai's evaluation table.

Third-party evaluators

• Artificial Analysis Intelligence Index: 53 → 60, +7 points, per Artificial Analysis's evaluation.
• GDPval-AA v2 Elo: 1,524 → 1,770, +246 Elo, in Artificial Analysis's results.

Customer-reported

• Voxel-world runtime at high versus max reasoning effort: 12 minutes → 23 minutes, +11 minutes, in cedric_chee's test.

Artificial Analysis put the cost of an Intelligence Index task at $0.44 → $0.68, +$0.24, and attributed part of that increase to roughly 20% more output tokens in its cost analysis. Z.ai's private Code Bench reports the reverse token direction on its own tasks, 96K → 75K output tokens at Max effort, in its release note.

• CritPt physics reasoning: 20.9% → 19.1%, -1.8 points, on teortaxesTex's chart.
• AA-Omniscience accuracy: 24% → 34%, +10 points, while the hallucination rate moved 26% → 30%, +4 points, according to Artificial Analysis's breakdown.
• ExploitBench: 24.4% → 54.4%, +30.0 points, yet GLM-5.3's 54.4% remained below Fable 5's 78.0% and GPT-5.6 Sol's 76.5% in Z.ai's comparison table.

CyberGym measures white-box identification and fault validation, while ExploitBench reaches deeper exploitation reasoning. Z.ai ran CyberGym as single-run Pass@1 over 1,507 tasks in a Claude Code harness with unlimited task time; an independent benchmark breakdown notes that the underlying CyberGym paper reported 759 raw crashes for nine confirmed zero-days after manual triage.

GLM-5.3 retains GLM-5.2's 753B-parameter, 40B-active MoE base, as Artificial Analysis's release analysis reports. Z.ai says every gain came from scaling post-training across three existing components:

1. IndexShare, for efficient long-context processing.
2. SAO with compaction, for long-horizon RL.
3. slime, its large-scale asynchronous RL framework.

Z.ai's training description says research agents turn real-work patterns into executable environments, judge agents check that tasks are solvable, and synthesized verifiers are tested against oracle, no-op, and unsolved-state cases. Solver trajectories then expose reward shortcuts for the environment pipeline to close.

On the systems side, slime puts Megatron training, SGLang rollout, and the data buffer on one dataflow. Z.ai says numerical train-rollout alignment held average log-probability difference around 1e-7 and its resource and scheduling changes raised end-to-end RL throughput more than 2.3x for long-horizon coding workloads; ZhihuFrontier's slime breakdown traces the alignment work through FP32 residuals, BF16 rounding boundaries, FP8 GEMMs, and MoE reduction order.

A ZCode voxel-world run from cedric_chee used 114.5K of the 1M-token context and averaged a 94.6% cache-hit rate during a 23-minute Max-effort session. The same report put usage near three million tokens.

A separate Godzilla Three.js build from ai_for_success priced GLM-5.3 at $0.37 against $0.44 for GPT-5.6 Sol, while reporting substantially higher token use and stronger creature modeling from Sol in that run.

Long-chain behavior also drew mixed observations. haider1's hands-on report described the model testing assumptions aggressively and correcting itself after Chinese words appeared in a chain of thought; HCSolakoglu's initial take preferred GLM-5.2 and called 5.3 benchmark-focused.

• Z.ai Coding Plan and ZCode: all Coding Plan subscribers received GLM-5.3, with token-type-based points billing and 50% off-peak point use in Z.ai's rollout notes.
• OpenRouter: OpenRouter's model listing exposes z-ai/glm-5.3 with low, high, and max reasoning effort.
• Vercel AI Gateway: Vercel's AI Gateway announcement says the model is live there and claims more than a 2x GLM-5.2 gain on DeepsecBench.