What Is GLM 4.6? Specs, Benchmarks and Pricing
Nov 19, 2026

What Is GLM 4.6? Specs, Benchmarks and Pricing

What is GLM 4.6? Z.ai's September 2025 model with 200K context, 30%+ lower token use and MIT weights. Specs, benchmarks, prices and if it's still worth using.

What is GLM 4.6? Specs, benchmarks and pricing of Z.ai's 200K-context model

GLM 4.6 still turns up in model pickers, older tutorials and production configs. It sits next to newer GLM versions, often at the same price as GLM-4.7. So is it a legacy model to migrate away from, or a sensible budget choice?

What is GLM 4.6? It's the large language model Z.ai (Zhipu AI) released in September 2025 as the successor to GLM-4.5. It's best remembered for three changes: it raised the context window from 128K to 200K tokens, improved real-world coding inside agent tools such as Claude Code, and used over 30% fewer tokens than GLM-4.5 on Z.ai's coding tests. Its weights are open under the MIT license.

This article covers GLM 4.6's specs, what it improved, how it compares with GLM-4.7, what it costs, and whether it's still worth using in late 2026. It's based on Z.ai's official GLM-4.6 model card and documentation, the GLM-4.7 model card (which includes GLM-4.6 comparison scores) and Z.ai's pricing page, all reviewed November 19, 2026. Benchmark figures are Z.ai's own published results. GLM5 publishes this article and operates independently of Z.ai.

GLM 4.6 at a Glance

GLM-4.6
Made byZ.ai (Zhipu AI)
ReleasedSeptember 2025 (official repository published September 29, 2025)
PredecessorGLM-4.5 (shares the same inference method)
ParametersAbout 357B in the published checkpoint
Context window200K tokens (up from GLM-4.5's 128K)
Max output128K tokens
ModalitiesText in, text out (vision is a separate GLM-4.6V model)
LicenseMIT
API price (Z.ai)$0.60 input / $0.11 cached input / $2.20 output per 1M tokens

The parameter count comes from the official Hugging Face repository's file metadata. Z.ai's model card says GLM-4.5 and GLM-4.6 use the same inference method. GLM-4.5's technical report describes a 355B-total, 32B-active mixture-of-experts design.

What GLM 4.6 Improved Over GLM-4.5

Z.ai's GLM-4.6 documentation lists five upgrades:

  1. Longer context: 128K → 200K tokens, enough for bigger agent sessions and larger code or document inputs.
  2. Better real-world coding: higher code-benchmark scores and better results inside Claude Code, Cline, Roo Code and Kilo Code, including cleaner front-end pages.
  3. Stronger reasoning: with support for tool use during inference.
  4. More capable agents: better tool use and search-based agent performance.
  5. Refined writing: a more natural style, including in role-play.

The most useful evidence is Z.ai's real-world test. It ran 74 coding tasks inside Claude Code and reported that GLM-4.6 beat Claude Sonnet 4 and other Chinese models in those tests. It also used over 30% fewer tokens on average than GLM-4.5. Z.ai published all the test questions and agent trajectories as the CC-Bench dataset on Hugging Face, so the results can be checked independently. That level of transparency is uncommon for a vendor benchmark.

Why the token saving matters: a coding agent bills you for every token it reads and writes across a long session. A model that finishes the same task with 30% fewer tokens costs roughly 30% less per task at the same price, before any difference in quality.

GLM 4.6 vs GLM 4.7: Benchmarks

Z.ai's GLM-4.7 model card compares the two directly:

Benchmark (Z.ai's published results)GLM-4.6GLM-4.7
SWE-bench Verified68.073.8
SWE-bench Multilingual53.866.7
Terminal Bench 2.024.541.0
LiveCodeBench v682.884.9
AIME 202593.995.7
τ²-Bench (tool use)75.287.4
HLE (with tools)30.442.8

GLM-4.7 is ahead everywhere, by the widest margins on terminal tasks, multilingual coding and tool use. On pure maths and competitive programming (AIME, LiveCodeBench) the gap is small.

Is GLM 4.6 Open Source?

Yes, in the open-weight sense. GLM-4.6's weights are published on Hugging Face under the MIT license, and Z.ai's GLM-4.5 GitHub repository provides the inference code shared by GLM-4.5 and 4.6. Expect a multi-GPU server to run the full model.

Is GLM 4.6 Free?

The weights are free. Z.ai's API charges $0.60 per million input tokens and $2.20 per million output tokens for GLM-4.6. Z.ai's pricing page shows the same rate for GLM-4.7. Z.ai's free API tier covers Flash models, such as GLM-4.7-Flash, not GLM-4.6. See Is GLM free? for every free option across the family.

Should You Still Use GLM 4.6?

Here's the awkward fact: on Z.ai's pricing page, GLM-4.6 and GLM-4.7 cost exactly the same, and GLM-4.7 scores higher on every benchmark in Z.ai's comparison. For a new project, there's little reason to choose GLM-4.6 over GLM-4.7.

SituationRecommendation
New projectUse GLM-4.7 (same price, better) or a GLM-5.x model
Existing app validated on GLM-4.6Keep it running, but test GLM-4.7 on a sample, since the swap is usually just a model-ID change
Prompts tuned heavily for GLM-4.6Re-test before switching, because output style can change between versions
Need image inputUse GLM-4.6V, the separate vision model

Rule of thumb: when two versions cost the same, migrate to the newer one as soon as it passes your test set. Staying on an older version should be a deliberate choice, not something you forgot to change.

For more on the successor, read what GLM 4.7 is. For the full lineage, see what GLM is.

Try GLM 4.6 Against Its Successor

GLM5's API serves glm-4.6 and glm-4.7 (plus the GLM-5.x models) through one OpenAI-compatible endpoint, so you can run the same prompt on both and compare. Get the model IDs from the GLM5 API docs, or learn more on the GLM 4.6 model page.

Frequently Asked Questions

Who made GLM 4.6?

Z.ai, the international brand of Zhipu AI.

When was GLM 4.6 released?

September 2025. The official Hugging Face repository was published on September 29, 2025.

What is GLM 4.6's context window?

200K tokens, with up to 128K output tokens. That was an increase from GLM-4.5's 128K context.

Is GLM 4.6 multimodal?

No. GLM-4.6 is text-only. Z.ai offers image understanding through the separate GLM-4.6V model.

Is GLM 4.6 better than GLM 4.5?

Yes, according to Z.ai. It has a longer context, better coding and agent results, and over 30% lower token use in Z.ai's Claude Code tests.

Is GLM 4.6 or GLM 4.7 better?

GLM-4.7, on every benchmark in Z.ai's comparison, and at the same API price.

The Bottom Line

GLM 4.6 was the release that gave GLM a 200K-token context and made it noticeably more efficient for agentic coding, with unusually transparent real-world tests behind it. In late 2026 it's been overtaken. GLM-4.7 costs the same and performs better, and the GLM-5 generation goes much further. Keep GLM 4.6 if it's working in production, but test the upgrade.

The quickest test is a direct comparison: run one real task on glm-4.6 and glm-4.7 through the GLM5 API and see which one your workload prefers.

Sources

Reviewed November 19, 2026. Benchmarks are Z.ai's published figures. Prices and availability can change, so confirm them on Z.ai's official pages.

Start Using GLM 5 Today

Try GLM 5 free — reasoning, coding, agents, and image generation in one platform.