ModelBrief · GLM-5.5 · Topic Page
Last updated Nov 20, 2026
China · LLM · Unreleased · Tracking

GLM-5.5 Tracker

GLM-5.5 has never been announced and never shipped: the name appears in none of Zhipu (Z.ai)'s release notes, docs, pricing page, Hugging Face organization zai-org, or OpenRouter's model list. But the rumors have never stopped — from a June JPMorgan research note to a GitHub PR attribution on November 17. This page collects every GLM-5.5 rumor and claim online — release date, parameters, grey-test signals, official denials — and labels each one confirmed, unverified or debunked, in a dated timeline. Latest signal: on November 17, 2026, a merged GitHub PR carried an agent attribution reading "GLM-5.5 session".

Not announcedAnnouncement status (Z.ai has never published a GLM-5.5 announcement)
UnreleasedRelease status (no model card, weights, API or price)
GLM-5.3Current flagship (announced Aug 14, 2026; 744B/40B)
GLM-6.0Next generation named (named by Tang Jie on Aug 31, 2026; no date)
NoneWeights and API (no GLM-5.5 repo on HF zai-org)
Not announcedPricing (Z.ai has announced no price)

01Status overview

In one sentence: GLM-5.5 exists only in rumors. Every item below was checked by this site on November 18, 2026.

ItemStatusNotes
Official announcementNoNo GLM-5.5 in Z.ai release notes, blog or official accounts
ReleaseNoNo release date, no launch event
Model cardNoneNo such model in z.ai docs
Open weightsNoneNewest HF zai-org repos are the GLM-5.3 family, Aug 25, 2026
APINoneNo z-ai/glm-5.5 on OpenRouter; no such model on Z.ai's pricing page
PricingNot announcedZ.ai has announced no price
Benchmark scoresNoneNo GLM-5.5 benchmark exists
Grey testUnverifiedOnly signal is the Nov 17 GitHub PR attribution; meaning unclear

02Confirmed vs rumored

Separating fact from rumor. Three grades: confirmed (official or verifiable source), unverified (sourced but unconfirmed by Z.ai), debunked (contradicted by facts or an official denial).

ClaimStatusSource and date
GLM-5.3 is the current flagship: 744B total / 40B active, 1M contextConfirmedZ.ai official, Aug 14, 2026
The next generation is named GLM-6.0, themed "full self-training", no dateConfirmedH1 results call, Aug 31, 2026
Zhipu committed to an open-weight model at Fable/Mythos level before the end of 2026ConfirmedZhipu public statement
GLM-5.3-FlashX speed tier launched, up to 200 tokens/sConfirmedSep 18, 2026
JPMorgan note: Zhipu is its top China AI pick; a November GLM-5.5/GLM-6 launch is the next catalystConfirmed (the note exists; its content is an analyst forecast)JPMorgan, Nov 4, 2026
Zhipu denied the leaked 32B/122B/256B parameters, calling them fabricatedOfficial response (as reported by media)Nov 9, 2026
"1T+ parameters"UnverifiedJuly 20 post, no source
"3T+ parameters"UnverifiedFrom July 29, no source
"Well into post-training"UnverifiedLumina X post, Oct 31, 2026, screenshot of unknown origin
Tsinghua lecture: six data engines (10 billion web pages a day, and more)Unverified@bebuu_ai X post, Nov 4, 2026, no recording or event page
"GitHub PR attribution means a grey test"UnverifiedAttribution string is real; meaning unclear, Nov 17, 2026
"Launch window is late October/early November 2026"Not deliveredCometAPI, Oct 25, 2026
"November 19, 2026 launch"Not deliveredapihop.ai, predicted date Nov 19, 2026
"August 2026 launch"DebunkedAugust actually brought GLM-5.3
"Skipping 5.3 to ship 5.5 directly"DebunkedGLM-5.3 shipped on schedule, Aug 14
"September launch, beating Mythos 5.1"DebunkedSeptember passed with no launch; Anthropic lists no Mythos 5.1
"32B/122B/256B checkpoint tiers"DebunkedDenied by Zhipu on Nov 9, called fabricated
"The opencode.ai page is a leak"DebunkedAuto-generated platform page, marked "model exists: no"
"Training data from hundreds of millions of repos"DebunkedThe GLM-5 paper reports about 10 million issue-PR pairs

03Rumor timeline

Newest first. Status labels match section 02.

  • 2026-11-19

    This site checked: apihop.ai's predicted "November 19 launch" did not happen. As of 20:06 that day, Z.ai had no announcement, HF zai-org had no new repo, and docs.z.ai had no update. Reseller TokenRa claimed the same day that the launch date was "confirmed", while admitting the model is not callable and there is no model card, price or official link — not credible.

  • 2026-11-18

    This site checked: still no GLM-5.5 repo on HF zai-org. The same day, WeChat account RXD智能 published "Breaking! Zhipu GLM-5.5 suspected grey test!", reading the GitHub PR attribution as a grey-test signal, in "suspected" and "very likely" wording.

  • 2026-11-17

    An agent attribution reading "GLM-5.5 session, 2026-11-17" appeared in merged GitHub PR shd101wyy/Yo#1991. The string is real; its meaning is unverified.

  • 2026-11-13

    CellCog argued that on the GLM release cadence (51–83 days per release, median 65), the next version should be called GLM-5.4, not 5.5, and that the "Mythos 5.1" comparison does not exist.

  • 2026-11-09

    Zhipu officially denied the leaked parameters: the circulating "32B/122B/256B checkpoint tiers" were called fabricated, with parameter count and release date still unannounced (official response as reported by media). The same day, llmtr checked the rumors item by item: the GLM-5 paper reports about 10 million issue-PR pairs, not "hundreds of millions of repos".

  • 2026-11-07

    Traictory checked the Tsinghua lecture claim: no recording, no event page, no attendee account, no Z.ai confirmation or denial.

  • 2026-11-06

    X account ai_funss relayed the six-engine figures (165,000 views); kompozy covered the post-training leak and the Tsinghua lecture.

  • 2026-11-04

    JPMorgan note: Zhipu is its top China AI pick, with a November GLM-5.5/GLM-6 launch as the next catalyst. The same day, X account @bebuu_ai posted that Tang Jie had disclosed a six-engine data pipeline and 32B/122B/256B tiers in a Tsinghua lecture (over 200,000 views).

  • 2026-11-01

    kie.ai relayed Lumina's 32B/122B/256B claim; OmniaKey noted 79 days since the GLM-5.3 blog — the longest gap in the GLM line.

  • 2026-10-31

    X account Lumina posted that GLM 5.5 was well into post-training, with three internal models of 32B, 122B and 256B parameters to merge into the next-generation base; the attached screenshot gave no source.

  • 2026-10-28

    An auto-generated opencode.ai glm-5-5-flash page was misread as a leak; its metadata reads "model exists: no" — later debunked.

  • 2026-10-25

    CometAPI said the launch window had moved from August to "late October/early November 2026"; the window has passed with no launch.

  • 2026-10-21

    X account @NFT_Chen posted a leak claim (cited by OmniaKey).

  • 2026-09-18

    GLM-5.3-FlashX speed tier launched: up to 200 tokens/s at about 2.5x Flash pricing, a service-side speedup.

  • 2026-09-17

    Tang Jie reported that a GLM-5.3 agent helped build Flash's production inference system (100,000+ domestic chips, about 3x throughput); Zhipu said recursive self-improvement has not been reached.

  • 2026-08-31

    H1 results call: Tang Jie named the next-generation base model GLM-6.0, themed "full self-training", with no date; chairman Liu Debing said the next-generation base was already in progress.

  • 2026-08-28

    GLM-5.3 weights went public under a custom GLM 5.3 License (not MIT).

  • 2026-08-26

    GLM-5.3-Flash released: 320B/18B, the first natively multimodal GLM-5, MIT license.

  • 2026-08-14

    GLM-5.3 announced: the model that actually took the August slot, 744B/40B MoE, 1M context; API live August 18.

  • 2026-07-29

    The "3T+ parameters" claim appeared, with no source.

  • 2026-07-26

    The "Mythos-level" claim appeared, with no source.

  • 2026-07-20

    A post claimed Zhipu would skip 5.3 and ship GLM-5.5 in August with 1T+ parameters; both predictions were debunked.

  • 2026-07-03

    Fello AI started tracking the model with a GLM 5.5 preview page (later rewritten as a "why it never shipped" record).

  • 2026-06-30

    CGTN's report on the GLM-5.2 launch said the next model, GLM 5.5, was expected in August.

  • Late June 2026

    A JPMorgan research note (reported by Reuters) predicted Zhipu would release trillion-parameter GLM-5.5 in August — the start of the whole rumor chain.

04What Zhipu actually said

Setting every rumor aside, here is everything Zhipu has said in public about its next model.

Name of the next generation
GLM-6.0 (named by Tang Jie at the H1 results call, Aug 31, 2026; theme "full self-training"; no date)
Progress
The next-generation base model is already in progress (chairman Liu Debing, Aug 31, 2026)
Year-end goal
An open-weight model at Fable/Mythos level before the end of 2026 (nicknamed "Open Fable" by the community)
Self-training progress
A GLM-5.3 agent helped build Flash's production inference system (Sep 17, 2026; 100,000+ domestic chips, about 3x throughput); Zhipu says recursive self-improvement has not been reached
Response to leaked parameters
The 32B/122B/256B figures were called fabricated; parameter count and release date remain unannounced (Nov 9, 2026; official response as reported by media)
On the name "GLM-5.5"
Zhipu has never used it in any public setting (checked Nov 19, 2026)

Note: JPMorgan's November 4 note predicting a November GLM-5.5/GLM-6 launch is an analyst forecast, not a Zhipu statement; the note named Zhipu its top China AI pick on model demand and falling costs, which has nothing to do with whether GLM-5.5 exists.

05Leak details

The five leads worth examining in detail: what each one actually says, and which link in its evidence chain is missing.

The GitHub PR attribution (Nov 17, 2026 — latest signal, unverified)

In merged GitHub PR shd101wyy/Yo#1991, an agent auto-wrote the attribution "Authored by a coordinator subagent (GLM-5.5 session, 2026-11-17)". The project requires agents to label their model and harness in PRs, but neither the project nor its prompts ever mention 5.5. The WeChat account RXD智能 reported it on November 18, 2026 as a suspected grey test, using "suspected" and "very likely" wording, ending with "personal opinion, for reference only" and stating plainly that "the publicly available models are still the GLM-5.3 family".

Three readings are all possible: an internal or grey-test deployment, a model misreporting its own name, or a harness placeholder string. They cannot be distinguished yet. Watch whether later PR attributions repeat GLM-5.5.

The Tsinghua lecture and six data engines (Nov 4, 2026 — unverified)

On November 4, 2026 at 18:48, X account @bebuu_ai posted that Zhipu co-founder and chief scientist Tang Jie had disclosed the next GLM's pretraining pipeline in a Tsinghua University lecture: six in-house data engines replacing public datasets — 10 billion web pages a day, 2 billion research PDFs, 50 million books, 500 million repositories converted into executable environments — plus 32B/122B/256B tiers. The post drew over 200,000 views.

Traictory's November 7 check found no lecture recording, no event page, no attendee account, and no Z.ai confirmation or denial. The 32B/122B/256B figures had already appeared on November 1 in a kie.ai write-up relaying Lumina, described then as internal checkpoints rather than products. On November 6, X account ai_funss relayed the numbers (165,000 views), spreading them further; kompozy covered the leak the same day.

"Well into post-training" (Oct 31, 2026 — unverified)

X account Lumina posted on October 31 that GLM 5.5 was well into post-training, with Zhipu training three internal models of 32B, 122B and 256B parameters to merge into the next-generation base. The attached screenshot gave no source, so its origin cannot be confirmed.

The JPMorgan and CGTN "August launch" (June 2026 — debunked)

A late-June JPMorgan research note (reported by Reuters) predicted Zhipu would release GLM-5.5 in August; CGTN's June 30 report on the GLM-5.2 launch said the next model, GLM 5.5, was expected in August. That was the start of the whole rumor chain. August actually brought GLM-5.3 (August 14) and GLM-5.3-Flash (August 26). The Baidu Baike "trillion-parameter" entry cites no JPMorgan document.

Fake leak: the opencode.ai auto-generated page (Oct 28, 2026 — debunked)

The page opencode.ai/data/zhipu/glm-5-5-flash was misread as a GLM-5.5-Flash leak; its metadata reads "page: auto-generated" and "model exists: no" — a platform-generated placeholder, not a Zhipu product.

06What to use now

GLM-5.5 cannot be used. But the GLM-5.3 family can — and GLM-5.3 is the model that actually took the August slot.

GLM-5.3 (current flagship)
744B total / 40B active MoE, 1M context, 128K output, text input; API $1.40/$4.40 per 1M tokens; weights public since August 28, 2026 under a custom GLM 5.3 License (not MIT)
GLM-5.3 benchmarks
AA Intelligence Index v5 reading 31.7 (felloai, November 7, 2026, maximum-reasoning setting); CyberGym 84.5 vs Mythos 5's 83.8 (Z.ai self-reported)
GLM-5.3-Flash
320B/18B, the first natively multimodal GLM-5, MIT license; $0.15/$0.50; the FlashX speed tier launched September 18, up to 200 tokens/s at about 2.5x Flash pricing, a service-side speedup
GLM Coding Plan
$18/$80/$168 per month
Where to get it
Z.ai open platform, Hugging Face zai-org, OpenRouter (z-ai/glm-5.3)
Related topic
The model in our Nex-N2.6 topic page is a computer-use agent post-trained on GLM-5.3-Flash

07FAQ

When is GLM-5.5 coming out?

Z.ai has given no date. The earliest "August" came from a JPMorgan note and a CGTN report, both debunked; CometAPI's "late October/early November 2026" window has passed with no launch; apihop.ai's November 19 date also passed with no launch; JPMorgan's November 4 note expects a November GLM-5.5/GLM-6 launch, but that is an analyst forecast, not a Z.ai statement.

Is GLM-5.5 released? Can I use it now?

No. No model card, weights, API or price; no z-ai/glm-5.5 on OpenRouter; no GLM-5.5 repository on Hugging Face zai-org (checked November 19, 2026). The publicly available models are still the GLM-5.3 family.

How many parameters does GLM-5.5 have? Is the trillion-parameter claim real?

Z.ai has published no size. "1T+" comes from a sourceless July 20 post, "3T+" from July 29, "2.5–3.5T" from CometAPI and an X post; the Baidu Baike "trillion-parameter" entry cites no JPMorgan document. All unverified.

Will GLM-5.5 be open source?

Unknown. The GLM-5.3 flagship uses a custom license (not MIT) while Flash kept MIT; Zhipu has committed to an open-weight model at Fable/Mythos level before the end of 2026, but that is a separate goal, not a GLM-5.5 promise.

How much will GLM-5.5 cost?

Not announced. For reference: GLM-5.3 API is $1.40/$4.40 per 1M tokens, Flash $0.15/$0.50.

Is the GLM-5.5 grey test real? What does the GitHub PR attribution mean?

Unverified. The only signal is an agent attribution reading "GLM-5.5 session" in merged GitHub PR shd101wyy/Yo#1991 on November 17 — the string is real, but it could be an internal test, a grey test, a model misreporting its name, or a harness placeholder, and the readings cannot be distinguished yet. The WeChat article that reported it used "suspected" and "very likely" wording itself.

Are the 32B/122B/256B checkpoint sizes real?

Z.ai denied them on November 9 (official response as reported by media), calling them fabricated. The figures first appeared in Lumina's October 31 X post (screenshot of unknown origin) and were repeated in @bebuu_ai's November 4 lecture post.

Is the next model GLM-5.5 or GLM-6.0?

The only next-generation name Zhipu has used in public is GLM-6.0 (August 31, no date). CellCog argued on November 13 that the release cadence points to GLM-5.4 next; GLM-5.5 could still appear as an in-between release, but Zhipu has never used the name.

08Sources

Every fact on this page comes from the public sources below (Hugging Face model cards and official pages read on November 18, 2026). Entries without links are cited by name and date; this site read them during verification but does not fabricate links.

Official pages and model cards

Original rumor sources

Tracker and fact-check pages

Chinese coverage

More indexes: All Zhipu models · All Language Models · All models