GLM-5.3 Just Launched: Specs, Benchmarks, API & How to Use It

Right now

Trending on Kingy

Keep reading with the stories getting the most attention now.

1. 01 Blog · Aug 4

   Can MiniMax H3 Generate Uncensored Video? What Local Deployment Actually Changes

   Read story →
2. 02 Blog · Aug 13

   GLM-5.3 vs Kimi K3 vs DeepSeek V4 Pro: Who Actually Has the Open-Weight Coding Crown?

   Read story →
3. 03 Blog · Aug 12

   Grok 4.6 vs GPT-5.6 Sol vs Claude Fable 5: Which AI Is Best for Coding and Agents?

   Read story →

Z.ai has launched GLM-5.3, and the most important fact is the least glamorous one: this is not a new base model.

The company says GLM-5.3 uses the same base as GLM-5.2. The gains come from another month of post-training—more executable environments, more varied long-horizon tasks and more reinforcement-learning compute. That makes 5.3 a useful test of how far a frontier lab can move a model after pretraining has ended.

The launch numbers are substantial. Z.ai reports Terminal-Bench 3.0 rising from 4.6 to 28.3, DeepSWE v1.1 from 46.2 to 66.9 and SWE-Marathon v1.1 from 19.4 to 42.5. Its private Z.ai Code Bench shows a 50.2% relative improvement at High effort. But the evidence needs disciplined labels: these are Z.ai-run evaluations, not independent GLM-5.3 reproductions, and the private benchmark cannot currently be audited outside the company.

The practical release is also incomplete. The glm-5.3 API and GLM Coding Plan access are live, with new low, high and max reasoning levels. The exact weights, model card, license and a workable local-serving recipe are not. Z.ai says weights will follow two weeks after launch, after safety evaluation and hardening.

Z.ai’s launch post is dated August 14, 2026 in China. This article was source-checked on August 13 in Pacific time, when the same release was already live. Kingy.ai did not perform an inference test because no verified, no-cost exact-model trial was available and no paid credentials were used.

GLM-5.3 verdict

GLM-5.3 looks like a serious coding and agent upgrade, especially on tests where GLM-5.2 had room to grow. The strongest evidence is the consistency of the direction: every reported 5.2-to-5.3 comparison in Z.ai’s launch table improves.

The restraint is equally important. “Public benchmark” means the task set or evaluation project is public; it does not mean this GLM-5.3 run was independently executed. The independent DeepSWE leaderboard had not added GLM-5.3 at publication time. Terminal-Bench 3.0 is public, but Z.ai used its own Claude Code configuration, three rollouts per task and generous limits. Z.ai Code Bench remains private. None of that invalidates the results. It limits what they prove.

For developers, the launch-day recommendation is straightforward:

• Try GLM-5.3 through the API or Coding Plan if coding-agent work is the target.
• Use max effort for difficult repository tasks; use low when latency and token use matter.
• Migrate any direct API request that disables thinking before changing the model ID.
• Do not budget from GLM-5.2 token prices or plan a local deployment until Z.ai publishes the missing 5.3 artifacts.

GLM-5.3 specifications and launch status

Calling GLM-5.3 “open source” on launch day is premature. Z.ai intends to release weights, but a promised future checkpoint is not a downloadable checkpoint. The precise description today is API-available, open-weight release promised.

What changed if the base model did not?

Z.ai says it kept scaling the post-training system built for GLM-5.2. That system combines long-context architecture work, the SAO reinforcement-learning approach and the open-source slime training framework with executable task environments.

The notable change is the environment mix. Instead of short coding exercises, Z.ai describes production-shaped tasks that can require days of expert work: diagnose a training-stack bottleneck, inspect code and internal documentation, run experiments, implement an optimization and prove an end-to-end speedup without breaking correctness. Research agents synthesize tasks, judge agents check whether they are solvable and verifiers are tested against oracle, no-op and unsolved states.

This is a post-training story, not an architecture story. It also means outsiders cannot yet separate the contribution of better task generation, more RL compute, altered system prompts, the agent harness and the model update itself. The missing model card and reproducible checkpoint matter for that reason.

The “50% improvement” claim, recalculated

Z.ai’s headline says GLM-5.3 improves 50% over GLM-5.2 on its in-house Z.ai Code Bench. The chart provides enough detail to check how that headline was produced.

At High effort, GLM-5.3 scores 31.4% while GLM-5.2 scores 20.9%. That is:

(31.4 - 20.9) / 20.9 × 100 = 50.2%

The absolute gain is 10.5 percentage points. At Max effort, 5.3 reaches 34.5% versus 23.4% for 5.2: an 11.1-point gain or 47.4% relative. Z.ai also shows average output falling from roughly 96,000 to 75,000 tokens per task at Max, a reduction of about 21.9%.

That is an encouraging quality-efficiency shift. It remains a private, vendor-designed benchmark evaluated in Claude Code 2.1.207. Z.ai says the benchmark covers realistic local development environments and uses both completion rate and checklist accuracy; it has not released the tasks, graders or run artifacts.

Claude Fable 5 still leads this private chart at Max effort with 39.5%, compared with GLM-5.3 at 34.5%. At High effort, however, 5.3’s 31.4% exceeds Claude Opus 4.8’s 29.5% while using less than half the displayed output tokens.

Every reported GLM-5.3 benchmark delta

The table below recalculates every GLM-5.2-to-5.3 change in Z.ai’s public launch table. A percentage-point gain is not the same as a relative percentage gain, and results with different units should not be compared as one scale.

The biggest relative numbers need the most context. Terminal-Bench 3.0 improves more than sixfold, but the starting score was only 4.6. ExploitGym counts completed tasks rather than percentage points, and its time budgets are normalized using different per-model token speeds. GDPval is an Elo rating. These cannot be collapsed into one average without inventing a unit.

Public, private, vendor-run and independent: what the evidence actually is

The independent DeepSWE leaderboard is the cleanest warning against casual table merging. It runs all models on mini-swe-agent for consistency and, at publication time, listed GLM-5.2 at 44% ±2—not the 46.2 value in Z.ai’s launch table. Z.ai’s 5.3 footnote uses mini-swe-agent too, but specifies temperature 0.95, six-hour timeouts, 400K context and its own run. The benchmark name is the same; the exact evaluation record is not.

Where other models still lead

GLM-5.3 does not win Z.ai’s entire comparison table.

Those rows are not a universal ranking. Harnesses differ across models, Fable’s label explicitly allows fallback behavior and missing cells are not losses. They do show why the responsible launch verdict is “major upgrade” rather than “best model at everything.”

For deeper comparisons, see Kingy.ai’s Kimi K3 benchmark analysis, GPT-5.6 Sol guide, Fable 5 comparison and DeepSeek V4 Pro model profile.

Footnotes and harness differences that change the meaning

Z.ai’s footnotes are unusually detailed, and several materially qualify the table:

• Terminal-Bench 3.0: Claude Code 2.1.207, max effort, 400K context, 128K maximum output, average of three rollouts, up to 600 agent turns and a ten-hour timeout. Tool Search is disabled.
• DeepSWE: mini-swe-agent, temperature 0.95, top-p 1.0, six-hour timeout and 400K context.
• Agents’ Last Exam: Claude Code, max effort, 1M context, 64K maximum output and up to eight hours on task-specific limits.
• FrontierSWE: evaluated by Proximal with 1M context, max effort and 128K maximum output. The table reports a dominance score rather than a simple pass rate.
• PostTrainBench: a weighted average of three runs. Unscored runs fall back to the official zero-shot base-model baseline, and Z.ai replaced some original anti-API checks with LLM inspection.
• SWE-Marathon: Z.ai modified anti-cheat checks for one task and repaired broken NVIDIA package installation for two others.
• ExploitGym: two- and six-hour budgets are rescaled using per-model throughput from Artificial Analysis—115 tokens per second for GLM-5.3, 40 for Kimi K3 and 47 for Qwen3.8-Max—plus non-API overhead.
• CyberGym: single-run Pass@1 over 1,507 tasks with unlimited per-task timeout.
• HLE with tools: up to 163,840 generation tokens and 300K context, with GPT-5.6 Luna Medium as judge.

These disclosures make the table more useful, not less. They also make it impossible to treat every cell as a controlled, identical race.

API changes: thinking is mandatory

GLM-5.3 supports three reasoning levels:

max is the default. On the direct API, thinking.type must be enabled. Z.ai’s migration warning is explicit: if an application currently sends thinking.type: "disabled", change it to enabled and set reasoning_effort to low before switching the model ID. Otherwise the request fails.

There is a subtle integration exception. Z.ai’s Coding Plan adapter guide says Claude Code and Codex inputs such as false, disabled, none or off are converted to low; they do not turn thinking off. That compatibility layer should not be mistaken for direct API behavior.

Dry-run GLM-5.3 API configuration

The following configuration is based on Z.ai’s live API conventions and 5.3 launch parameters. It was reviewed as a dry run only; it was not sent.

curl -X POST "https://api.z.ai/api/paas/v4/chat/completions" 
  -H "Authorization: Bearer $ZAI_API_KEY" 
  -H "Content-Type: application/json" 
  -d '{
    "model": "glm-5.3",
    "messages": [
      {
        "role": "user",
        "content": "Inspect this repository change, identify the highest-risk defect, and propose a minimal verified fix."
      }
    ],
    "thinking": { "type": "enabled" },
    "reasoning_effort": "max",
    "stream": true
  }'

An OpenAI Python SDK equivalent is:

import os
from openai import OpenAI

client = OpenAI(
    api_key=os.environ["ZAI_API_KEY"],
    base_url="https://api.z.ai/api/paas/v4/",
)

stream = client.chat.completions.create(
    model="glm-5.3",
    messages=[
        {
            "role": "user",
            "content": "Inspect this repository change, identify the highest-risk defect, and propose a minimal verified fix.",
        }
    ],
    reasoning_effort="max",
    extra_body={"thinking": {"type": "enabled"}},
    stream=True,
)

for chunk in stream:
    delta = chunk.choices[0].delta
    if delta.content:
        print(delta.content, end="")

Use the general API endpoint for application calls. The GLM Coding Plan uses a separate endpoint and is contractually limited to supported coding tools; it is not a discounted general-purpose API.

Claude Code, OpenCode, Codex and ZCode setup

Claude Code

For an existing Coding Plan installation, Z.ai’s current settings example maps Sonnet and Opus slots to the 1M route:

{
  "env": {
    "CLAUDE_CODE_AUTO_COMPACT_WINDOW": "1000000",
    "ANTHROPIC_DEFAULT_HAIKU_MODEL": "glm-4.7",
    "ANTHROPIC_DEFAULT_SONNET_MODEL": "glm-5.3[1m]",
    "ANTHROPIC_DEFAULT_OPUS_MODEL": "glm-5.3[1m]"
  }
}

Claude Code uses the Anthropic-compatible endpoint https://api.z.ai/api/anthropic. Run /status after restarting and verify that the active model is glm-5.3 or glm-5.3[1m]. The /effort command changes reasoning intensity.

OpenCode

Run opencode auth login, select Z.AI Coding Plan, enter the plan API key and use /models to choose glm-5.3. OpenAI-compatible coding tools use https://api.z.ai/api/coding/paas/v4.

Codex

Z.ai’s integration guide lists https://api.z.ai/api/v1 for Coding Plan use with Codex and maps Codex’s reasoning.effort values to low, high or max. Confirm that the tool supports custom providers before changing an existing configuration.

ZCode

ZCode is Z.ai’s own agentic development environment. The launch page advertises a 98%+ cache-hit rate, Goal mode for plan-code-test-verify loops and remote monitoring through WeChat or Feishu. Those are vendor claims about ZCode’s service, not guarantees for Claude Code or OpenCode. Z.ai says a limited 1.5× quota promotion combines with cache savings through August 31.

Access and pricing: what is known and what is missing

GLM-5.3 is available to all GLM Coding Plan tiers. The live plan page listed monthly prices of $18 for Lite, $80 for Pro and $168 for Max when checked. Annual billing displayed effective monthly prices of $12.60, $56 and $117.60. Lite includes 10,000 credits per week; Pro advertises six times Lite usage and Max fourteen times.

The new points system charges separately for input, cached input and output. Z.ai says calls outside weekday peak hours consume 50% of the standard points. Peak hours are 14:00–18:00 UTC+8, Monday through Friday.

What is missing is the price most API buyers need: GLM-5.3’s per-million-token rate. Z.ai’s official API pricing table still lists GLM-5.2 at $1.40 per million input tokens, $0.26 cached input and $4.40 output, but it has no 5.3 row. Do not assume those rates carry over.

The same staleness appears elsewhere: the public model catalog still features GLM-5.2 while the developer integration guide and Coding Plan page have moved to 5.3. That is a documentation lag, not proof that the model is unavailable.

What is not available yet

As of this launch-day check:

• No official GLM-5.3 weights are downloadable.
• No exact-model Hugging Face repository or model card is linked; Z.ai’s launch button says “Coming Soon.”
• No GLM-5.3 license has been published.
• No official local-serving command, quantization matrix or hardware guidance exists for the exact model.
• No standard API per-token price is listed.
• No independent GLM-5.3 result appears on the DeepSWE or Agents’ Last Exam leaderboards checked for this article.
• No verified no-cost exact-model inference route was identified, so Kingy.ai did not perform a smoke test.

Z.ai says the weights should arrive two weeks after launch. Treat that as a roadmap commitment, not current availability. The practical local-deployment question begins only when the checkpoint size, precision, license, tokenizer, inference-engine compatibility and memory requirements are public.

Should you use GLM-5.3 now?

Try it now if you already use the GLM Coding Plan, can absorb launch-day service changes and have a real repository task with a measurable outcome. The strongest test is not “write a snake game.” Give the model a failing build, a bounded migration, a cross-file bug or a performance regression. Hold the agent, tools, time budget and acceptance tests constant against Kimi K3, DeepSeek V4 Pro and your preferred closed model.

Wait if local control, license certainty, predictable API pricing or reproducible weights are requirements. GLM-5.2 remains the documented option until 5.3’s artifacts land.

The launch is still significant. A same-base post-training update moved every published predecessor comparison in the right direction and produced unusually large gains on new terminal, long-horizon and exploitation tasks. The honest conclusion is not that GLM-5.3 owns the frontier. It is that Z.ai has shown how much capability may still be left in a base model after pretraining—and has made the missing reproducibility artifacts the next thing worth watching.

Frequently asked questions

Is GLM-5.3 a new base model?

No. Z.ai says GLM-5.3 uses the same base model as GLM-5.2. The improvements come from expanded post-training, including more environments, more varied tasks and more RL compute.

Is GLM-5.3 open source or open weight?

Not yet in a practical sense. The API and Coding Plan are live, while weights are promised two weeks after launch. The model card and license are still pending. “Open-weight release promised” is the precise launch-day description.

What are GLM-5.3’s reasoning levels?

low, high and max. The default is max, and Z.ai recommends it for coding. Direct API requests cannot disable thinking.

Why does disabling thinking make GLM-5.3 requests fail?

The direct API no longer supports thinking.type: "disabled" for this model. Set thinking to enabled and use reasoning_effort: "low" for the lightest mode. Official Coding Plan adapters may translate disabled-like inputs to low, but that is adapter behavior.

Does GLM-5.3 have a one-million-token context window?

Z.ai documents a 1M Coding Plan route through the glm-5.3[1m] model suffix and a one-million-token Claude Code compaction window. A standalone 5.3 model page with the full standard-API context and output specification was not yet published.

How much does the GLM-5.3 API cost?

Z.ai had not added GLM-5.3 to its per-token pricing table at publication time. Coding Plan subscriptions are listed separately. GLM-5.2’s $1.40 input, $0.26 cached input and $4.40 output rates should not be presented as confirmed 5.3 pricing.

Did Kingy.ai test GLM-5.3?

No inference test was performed. No verified free exact-model trial was available during the launch-day check, and this article did not use paid credentials. The API examples are clearly labelled dry-run configurations.

Sources

• Z.ai: GLM-5.3 launch post
• Z.ai: how to switch Coding Plan tools to GLM-5.3
• Z.ai: GLM Coding Plan
• Z.ai: standard API pricing
• Z.ai: API quick start
• Z.ai: Thinking Mode documentation
• DeepSWE v1.1 leaderboard and methodology
• Terminal-Bench repository and run instructions
• Agents’ Last Exam leaderboard

Source status checked August 13, 2026 Pacific time / August 14 in China. Evaluation records, pricing and release artifacts may change after publication.