GLM-6: Release Date Predictions and What to Expect vs GLM-5.3

A2Agent Team · 2026-10-22T00:00:00Z

GLM-5.3 came out in August, Zhipu (Z.ai) has shipped a new flagship roughly every two months this year, and its founder has started talking about GLM-6 by name. The next big GLM release looks close.

As of October 22, 2026, GLM-6 has not been released or formally announced. There's no model card, no API model ID, no price and no weights. At least one site claims GLM-6 already launched in June 2026, but that's wrong. June's release was GLM-5.2.

Everything below about GLM-6's timing and performance is our prediction, built from what Zhipu has said and how it has released models so far. We'll update the post when Zhipu announces the model.

What Zhipu has said about GLM-6

On September 2, in coverage of Zhipu's first-half 2026 results, founder and chief scientist Tang Jie described the direction for GLM-6.0 as self-evolution: a model that learns when to stop, corrects its own mistakes, and works autonomously for longer, on tasks well beyond coding.

On September 14, Zhipu's financing disclosure said about 60% of its latest proceeds would go to next-generation GLM foundation models, a "Fully Self Training" system, and the compute to train and serve them. The plan has models helping to build their own training pipeline:

  • generating, filtering and checking their own training data, with humans sampling the results
  • building training environments, tasks and verifiers for agents
  • helping optimize the training and inference stack itself

The same disclosure lists stronger native multimodal modeling, more efficient computation without runaway inference costs, and reinforcement learning for long-horizon tasks as goals.

At the GLM-5.3-Flash launch in August, Zhipu also said the lessons from building it "are already shaping our next frontier model." It didn't name that model.

When will GLM-6 be released?

In 2026, Zhipu has shipped a new flagship every 54 to 70 days:

Model Released Gap since previous
GLM-4.7 December 22, 2025
GLM-5 February 12, 2026 52 days
GLM-5.1 April 7, 2026 54 days
GLM-5.2 June 16, 2026 70 days
GLM-5.3 August 14, 2026 59 days

At that pace, the next flagship was due between October 7 and October 23. Today is day 69 since GLM-5.3, so something new from Zhipu could land any day.

That next release isn't necessarily GLM-6, though. GLM-5.3 was built on the same base model as GLM-5.2, with all of its gains coming from about a month of extra post-training. A ".x" release can be another round of post-training on an existing base. GLM-6, going by Tang Jie's description and the self-training plan, sounds like a new base model trained a new way, and that takes longer than a post-training cycle. The money earmarked for it was only disclosed in mid-September.

At the last major version jump, GLM-5 arrived about seven months after GLM-4.5, following two point releases, and launched in mid-February, shortly before the Lunar New Year holiday.

Our estimate:

Window Likelihood Why
Late October to December 2026 Low for GLM-6, high for a GLM-5.x update The two-month cadence says something is due now, but the new base model and self-training system were only funded in September
January to March 2027 Most likely Leaves time for a new base model, and matches the timing of GLM-5's launch before the Lunar New Year (February 6, 2027)
After March 2027 Possible If the self-training system takes longer to build or Zhipu ships GLM-5.5 first

If GLM-6 follows GLM-5.3's rollout, expect it to appear first in the GLM Coding Plan, then in the API a few days later, with open weights following a safety review. GLM-5.3's weights arrived 11 days after launch, under a custom license instead of MIT, and a smaller Flash version followed the next day.

Where GLM-5.3 stands today

GLM-5.3 has a 1M-token context window and up to 128K output tokens, and it takes text only. These scores come from MiniMax's comparison table published with M3.1 Flash on October 19, which puts GLM-5.3 next to the two Flash models released this month:

Benchmark GLM-5.3 DeepSeek V4.1 Flash 1017 MiniMax M3.1 Flash Claude Opus 5
CyberGym 84.5 n/a 82.2 n/a
BankerToolBench 74.5 n/a 76.2 76.2
HLE (no tools) 42.3 42.6 43.4 55.6
Terminal-Bench 4.0 37.9 42.3 38.1 51.8
DeepSWE v1.1 66.9 77.0 73.7 74.0
NL2Repo-Bench 58.0 68.2 61.2 75.3
LOCA-Bench 256K 50.3 n/a 60.3 53.7
Automation-Bench v1.0.6 48.8 60.5 62.2 50.3

GLM-5.3's standout is security. Its CyberGym score is the highest in the table, which is why Zhipu held its weights back for a safety review. On agent and coding benchmarks, the October Flash models from DeepSeek and MiniMax have pulled ahead, and they cost a fraction of GLM-5.3's $1.40 / $4.40 per million tokens on Zhipu's API.

GLM-6 vs GLM-5.3: what to expect

The table below is our estimate. It's based on what Zhipu says it's building, how far a single post-training round moved GLM-5.3, and where competing models already score.

Area GLM-5.3 today Our guess for GLM-6 Reasoning
Long-horizon agent work (Automation-Bench) 48.8 58 to 65 Self-evolution and long-horizon RL target exactly this, and two Flash models already score above 60
Agentic coding (DeepSWE v1.1) 66.9 75 to 80 One post-training round took GLM-5.2's 46.2 to 66.9; a new base plus RL should at least match the 73.7 to 77.0 of October's Flash models
Terminal work (Terminal-Bench 4.0) 37.9 45 to 52 Longer autonomous runs and better self-correction help most on multi-step terminal tasks
General reasoning (HLE, no tools) 42.3 45 to 50 A new base model typically lifts general knowledge more than post-training does
Security (CyberGym) 84.5 85 or higher Already a strength; expect another staged release with controlled access
Input types Text only Text and images likely "Stronger native multimodal modeling" is a stated goal
Price $1.40 / $4.40 per 1M (GLM-5.2 API rate) Similar or higher for the flagship, with a cheaper Flash version later Zhipu has kept flagship prices high and added Flash variants for lower-cost use

If GLM-6 really knows when to stop and fixes its own mistakes, it should waste fewer tokens on long tasks. That would make it cheaper per finished task even if the token price goes up, which is the same trick DeepSeek and MiniMax pulled this month.

Should you wait for GLM-6?

If you're choosing a model today, there's no reason to wait. GLM-5.3 is a strong pick for security work such as finding vulnerabilities, and for agent loops and everyday coding, DeepSeek V4.1 Flash 1017 and MiniMax M3.1 Flash give higher scores on those benchmarks at a much lower price. All of them take the same OpenAI-format request, so switching to GLM-6 later is a change to the model name.

Be careful with any API or website offering "GLM-6" today. Zhipu hasn't published a model ID, so anything labeled GLM-6 right now is either renamed from another model or doesn't exist.

GLM on A2Agent

You can call GLM-5.3, GLM-5.3 Flash and the rest of the GLM family on A2Agent today, along with DeepSeek, Qwen, Kimi and MiniMax, using one API key. We'll add GLM-6 as soon as Zhipu releases it and update this post with the real release date, specs and benchmarks.

Sources