Z.ai never shipped GLM 5.5. For about ten weeks the internet treated that sentence as a miss. It is not one. The model Zhipu actually named in its place is a far bigger swing than any 0.2 version bump could have been, and the first receipts are already on the table.
The paper trail is a single line. A CGTN report on June 30, published around the GLM-5.2 launch, said Zhipu's next model, GLM 5.5, was expected in August. That was the whole basis for the date: no spec sheet, no parameter count, no price. Zhipu, which markets itself as Z.ai, never confirmed the name, and its developer docs have never carried a GLM 5.5 model ID.
August arrived and the calendar was half right. A flagship did land in mid-August, and a second model followed on August 26. Neither carried the 5.5 badge. The two-month rhythm that made the August guess look smart held perfectly, and that is the first sign nothing went wrong: GLM-5 in February, GLM-5.1 in April, GLM-5.2 in June, GLM 5.3 in August. Only the number moved.
What August Actually Delivered
GLM 5.3 is the current flagship, and it is serious engineering. It is a 744-billion-parameter mixture-of-experts model that activates roughly 40 billion parameters per token, with the 1-million-token context window carried over from GLM-5.2. Z.ai also claims a 50 percent gain over GLM-5.2 on its own Code Bench, plus a step change in vulnerability discovery work.
Then came the part that annoyed the open-weight crowd. Every GLM flagship through 5.2 shipped under MIT. GLM 5.3 did not. Its weights landed about two weeks after launch under a custom GLM 5.3 License, a first for the family and a clear signal that Z.ai now wants to protect the commercial edge of its best model.
Pricing stayed aggressive anyway: $1.40 per million input tokens and $4.40 per million output. Twelve days later, GLM 5.3 Flash arrived as the lighter sibling, 320 billion total parameters with 18 billion active, MIT licensed, with native image input and $0.15 in, $0.50 out. For individual developers the entry point is the GLM Coding Plan at $18, $80 or $168 a month, still the cheapest serious route to a frontier-adjacent coding model.
The Bet Z.ai Named Instead
On August 31, at Zhipu's first-half results call, founder Tang Jie skipped the rest of the 5.x line entirely and named the next-generation base model GLM-6.0. Its theme is what Zhipu calls full self-training, a step toward recursive self-improvement in which the model helps build and train its successors.
Tang was unusually specific about where the difficulty sits. The hard problem is not generating training data, he said. It is self-judgment: knowing when to stop training, and how to correct its own mistakes without a human sitting in the loop as referee.
The money followed within two weeks. In mid-September Zhipu raised roughly HK$39.3 billion net through a Hong Kong share placing and convertible bonds, and earmarked about 60 percent of it for the next-generation GLM base model, the self-training system and the compute behind them. That is not a research line item bolted onto a product roadmap. It is the roadmap.
Then on September 17, Tang reported a first result that deserves more attention than it received. A GLM 5.3 agent helped build and tune a production inference system running on more than 100,000 domestically produced chips, and roughly tripled end-to-end throughput in under two weeks.
Zhipu also said plainly that it has not achieved recursive self-improvement. It is the difference between an agent that tunes a serving stack and a system that rewrites its own training recipe, and the company is not pretending otherwise. The interesting part is that the serving stack is exactly where the feedback loop starts: an agent that can measure a system, change it, measure again and keep the change that wins.
The ambition has a deadline. Zhipu has committed to an open-weight model at Fable and Mythos level, the frontier tier, before the end of 2026. The community nicknamed the target Open Fable, and it reframes GLM 5.5 as a stepping stone that was never the destination. In June, Elon Musk suggested GLM could reach Mythos-level benchmarks by year-end, and Tang replied that it would not take that long.
Here is the honest counterweight: enthusiasm without numbers is marketing. On the Artificial Analysis Intelligence Index v5, read on November 7, GLM 5.3 scored 31.7 at maximum reasoning effort, while Kimi K3.1 Pro sat at 46.5. The open-weight lead that GLM-5.2 grabbed in June has moved. Shipping a self-training loop is not the same as shipping a model that beats the closed frontier, and nobody at Z.ai has claimed the two are equivalent.
Which makes the next few weeks worth watching: the cadence that made August look predictable points at a window that has already opened. The signals to track are boring and reliable:
- Z.ai release notes: a new dated entry is always the first public artifact.
- The zai-org Hugging Face organization: a weights repository appears before most benchmarks do.
- API model IDs and pricing pages: anything with 5.5 or 6.0 in the string is real, everything else is a placeholder.
If a model called GLM 5.5 appears now, it will be an in-between release, not the August flagship the reports predicted. The flagship is GLM-6.0, and it is aimed at a much harder target than closing four points on Terminal-Bench. Zhipu is trying to build a model that helps train the model after it, with 100,000 chips of its own silicon and a war chest sized for the attempt.
That is why the missing version number matters so little. GLM 5.5 was a label attached to a rumour. GLM-6.0 is a thesis about how the next generation of models gets built, and Z.ai has already spent real money, real chips and real engineering hours on it. The direction of travel is no longer ambiguous.
Comments