A post is doing the rounds this weekend: “GLM-5.5 Leak: Mythos Level.” Reportedly beats Mythos 5.1 and GPT-6 Astra, reportedly a September release, rumored 3T+ parameters, expected to stay open-weight, 1M context carried over. Every clause has “reportedly” in it, and the headline does not.
We traced it. The chain starts on July 20, six weeks before the version you are seeing, and the first two things it predicted have already not happened. This page is the dated record: what Z.ai has actually said and shipped, where each rumored number entered the story, and what Z.ai’s own release rhythm suggests instead.
On this page · 8 sectionsOpen
- GLM-5.5 is not announced, dated, sized, priced or licensed as of September 6, 2026. There is no GLM-5.5 model card on Z.ai’s Hugging Face organization, no glm-5.5 identifier in Z.ai’s API docs or on OpenRouter, and no Z.ai post that uses the name.
- The leak is older than it looks. The earliest version we can trace is a July 20 post claiming Z.ai would skip GLM-5.3 and ship GLM-5.5 in August with more than 1T parameters. Z.ai then shipped GLM-5.3 on August 14 and GLM-5.3-Flash on August 26, and August ended without a 5.5. Both dated predictions failed.
- The numbers grew as the rumor traveled: 1T+ on July 20, ‘Mythos-level’ by July 26, 3T+ by July 29, and ‘beats GPT-6 Astra’ with a September window in the version circulating this week. No primary source appeared at any step.
- The ‘Mythos 5.1’ in the current version matches nothing Anthropic lists. Anthropic’s model with that name is Claude Mythos Preview; the 5.1 that shipped on September 1 is Claude Fable 5.1. The one real fact nearby is Z.ai’s own claim that GLM-5.3 beats Mythos 5 on the CyberGym vulnerability benchmark, 84.5 to 83.8.
- A version skip has no precedent in the GLM line. Since GLM-4.5 in July 2025, Z.ai has shipped 4.6, 4.7, 5, 5.1, 5.2 and 5.3, every step 51 to 83 days apart (median 65). On that cadence the next point release lands between October 8 and November 9, and it would be called GLM-5.4.
- This is a living tracker. When Z.ai names a model, a zai-org repository appears, or a glm-5.x identifier lands on the API, the confirmed facts go here the same day, next to a scorecard of how each claim above held up.
§ 01What we actually know
| Claim | Status |
|---|---|
| GLM-5.5 exists as an announced or planned model | No Z.ai statement uses the name; nothing on z.ai, docs.z.ai, the zai-org Hugging Face organization or Z.ai’s X account |
| Z.ai would skip GLM-5.3 for GLM-5.5 | Failed: GLM-5.3 announced August 14, API August 18, open weights August 28 |
| August 2026 release | Failed: August closed with GLM-5.3 and GLM-5.3-Flash as the newest models |
| September 2026 release | Rumor only; no venue, no teaser, no repository activity |
| 3T+ parameters | Rumor only; started as “1T+” on July 20, became “3T+” on July 29 with no source for either |
| Beats Mythos 5.1 and GPT-6 Astra | No GLM-5.5 benchmark exists; “Mythos 5.1” matches no Anthropic model name |
| Open-weight | Unknown for 5.5; a pattern, since GLM-5.3 and GLM-5.3-Flash are both open, but not a statement |
| 1M-token context | Unknown for 5.5; GLM-5.3 and GLM-5.3-Flash both ship 1M today |
| A glm-5.5 API identifier or OpenRouter route | None; OpenRouter lists z-ai/glm-5.3, z-ai/glm-5.3-flash and the aliases ~z-ai/glm-latest and ~z-ai/glm-flash-latest |
| Z.ai’s current flagship | GLM-5.3, 744B total, 40B active, 1M context, 128K output, open weights since August 28 |
§ 02What Z.ai has said, and has not
Z.ai’s public record for the last month is dense, and GLM-5.5 is not in it.
On August 14, Z.ai announced GLM-5.3 as a post-trained model on its 744B base, focused on coding, agentic work and cybersecurity, available first through its Coding Plan and staged toward API access and open weights while security partners tested it. On August 18 the API went live and the docs.z.ai release notes recorded it: a 50 percent gain over GLM-5.2 on Z.ai Code Bench, state-of-the-art among open models on Terminal Bench 3.0, and the sentence that later got bent into the rumor, that GLM-5.3 “matches Mythos 5” in white-box cybersecurity work. On August 26 it announced GLM-5.3-Flash and confirmed it had been the anonymous Ox Alpha model on OpenRouter. On August 28 it opened the GLM-5.3 weights on Hugging Face.
Four announcements in two weeks, all about the 5.3 generation. What Z.ai has not said: any next model’s name, number, size, date or license. The GLM-5 repository README lists GLM-5.3 (744B-A40B) and GLM-5.3-Flash (320B-A18B) as the newest entries. The Hugging Face organization’s newest repositories were created on August 25 and last touched on September 4, and they are the 5.3 family.
§ 03The rumor chain, dated
| Date | Event | Source type |
|---|---|---|
| Jul 20 | Post on X: Z.ai may skip GLM-5.3, ship GLM-5.5 in August; 1T+ parameters, open weights, 1M context; “not confirmed by Z.ai” | Commentator post |
| Jul 25 | 13-post X thread titled “GLM-5.5 Leak: It Beats Fable 5”; positions 5.5 against Fable 5 and Mythos; no benchmarks or spec | Commentator thread |
| Jul 26 | Trade coverage recounts the thread and notes it “provides no benchmarks, model details, or official confirmation” | Secondary coverage |
| Jul 26 | “Mythos-level” framing appears; size still 1T+ | Commentator post |
| Jul 29 | Size becomes “3T+ parameter open-weight model” | Commentator post |
| Aug 2 | Blog roundup: rumored 3T+ model to “challenge GPT-6”, positioned against Fable 5, Mythos and Kimi K3 | Secondary coverage |
| Aug 14 | Z.ai announces GLM-5.3; the “skip 5.3” prediction fails | Z.ai, official |
| Aug 26 | Z.ai announces GLM-5.3-Flash, reveals it was Ox Alpha | Z.ai, official |
| Aug 28 | GLM-5.3 weights open on Hugging Face | Z.ai, official |
| Aug 31 | August ends with no GLM-5.5; the August prediction fails | Calendar |
| Sep 6 | Current version circulates: beats Mythos 5.1 and GPT-6 Astra, September window, 3T+ | Commentator post |
The size inflated with distance from the source. The July 20 post said more than 1T. Nine days later the same story said 3T+, with nothing new to point at. For scale, Z.ai’s flagship is 744B and its first-ever trillion-class model does not exist; the largest open model anyone has shipped this year is Alibaba’s Qwen3.8-2.4T-A95B. A 3T open-weight release from Z.ai would be a bigger event than any of the leaks describing it, which is exactly why it should have left a trace, and has not.
The rivals inflated too. Fable 5 on July 25 became Mythos on July 26, then “Mythos 5.1 and GPT-6 Astra” this week. Astra launched on September 3; the rumor absorbed it within three days. “Mythos 5.1” appears to be Fable 5.1, which shipped on September 1, fused with Z.ai’s real Mythos 5 comparison. That comparison is worth stating correctly because it is the only benchmark in this whole story that exists: by Z.ai’s own numbers, GLM-5.3 scores 84.5 percent on CyberGym against Mythos 5 at 83.8 and GPT-5.6 Sol at 83.6, and 54.4 percent on ExploitBench against 78.0 and 76.5. Ahead on one, well behind on the other. A rumor built to be quoted does not have room for the second half.
Two predictions have already been graded. The skip did not happen and August passed. When a chain’s first two checkable claims fail and its unfalsifiable claims keep growing, the honest reading is that it was never a leak; it was a guess that found an audience.
§ 04Why “skip to 5.5” has no precedent
Z.ai’s release history is the most useful thing on this page, because it is the only evidence about Z.ai’s next model that actually comes from Z.ai.
| Model | Date | Days since previous |
|---|---|---|
| GLM-4.5 | Jul 25, 2025 | start |
| GLM-4.6 | Sep 30, 2025 | 67 |
| GLM-4.7 | Dec 22, 2025 | 83 |
| GLM-5 | Feb 11, 2026 | 51 |
| GLM-5.1 | Apr 7, 2026 | 55 |
| GLM-5.2 | Jun 16, 2026 | 70 |
| GLM-5.3 | Aug 18, 2026 | 63 |
Seven releases, each one a single step in the version number, spaced 51 to 83 days apart with a median of 65. Dates are OpenRouter’s catalog dates, which have tracked Z.ai’s own announcements within a day; GLM-5.3 was announced to Coding Plan users on August 14 and reached the API on August 18.
Two things follow. First, the naming: Z.ai has never skipped a number in this line. The step after 5.3 has always been 5.4 in this pattern, and a jump to 5.5 would be the first exception. Second, the timing: 65 days after August 18 is October 22, and the full historical range puts the next release between October 8 and November 9. A September launch would be the shortest gap Z.ai has ever run. Neither rule is a law, and Z.ai is free to break both. But a rumor that needs the company to break two habits at once, silently, is a rumor that should show its evidence, and this one has none.
The Flash line is the other honest caveat. GLM-5.3-Flash arrived eight days after GLM-5.3’s API as a differently sized sibling, not a version step, after a two-week anonymous run as Ox Alpha. If Z.ai has something new in September, the precedent says it is likelier to look like that: a stealth checkpoint on OpenRouter or an Arena entry under a codename, weeks before a name.
§ 05What shipped instead, and what it costs
| Model | Shipped | Size | Context and output | Weights | Z.ai API price per 1M tokens |
|---|---|---|---|---|---|
| GLM-5.3 | Announced Aug 14; API Aug 18; weights Aug 28 | 744B total, 40B active | 1M in, 128K out; text only | Open, Hugging Face | $1.40 in, $4.40 out |
| GLM-5.3-Flash | Aug 26 (was Ox Alpha) | 320B total, 18B active | 1M in, 128K out; text, image, video, file | Open, MIT | $0.15 in, $0.50 out at list; $0.075 and $0.25 through Sep 9 |
That is a frontier-adjacent open model at $1.40 and a multimodal one at a tenth of the price, both released inside the window the rumor said would belong to GLM-5.5. Anyone waiting on 5.5 to try GLM has been waiting past the thing they were waiting for. The Flash pricing and the September 9 promo deadline are on our Ox Alpha pricing page; the reveal record is on the GLM-5.3-Flash page; how the flagship compares with its closest open rival is in GLM-5.3 vs Qwen3.8-Max.
We do not route to GLM today. CellCog’s tiers run on Claude Fable 5.1, Claude Opus 5 and Gemini 3.8 Flash, and we moved to each on the day it shipped. GLM matters to us because it is the open-weight family closest to the frontier, which makes the question of running an AI employee on an engine nobody else controls a live one. That decision deserves dated facts, and this page is where we keep them.
§ 06What we are watching for
We track open-weight frontier models as an operator weighing engines, not as spectators.
- A name from Z.ai. A post on z.ai or X, a docs.z.ai release note, or a Coding Plan announcement. Z.ai announced GLM-5.3 to subscribers four days before the API; that surface tends to move first.
- A zai-org repository. The moment a GLM-5.4, GLM-5.5 or unnamed new repository appears on Hugging Face or ModelScope, this stops being a rumor page.
- The rolling aliases. OpenRouter’s ~z-ai/glm-latest and ~z-ai/glm-flash-latest point at 5.3 and 5.3-Flash today. A change in what they resolve to is a same-day tell.
- A stealth checkpoint. Ox Alpha ran anonymously on OpenRouter for two weeks before it became GLM-5.3-Flash. A new codename with a 1M context and Z.ai’s tokenizer fingerprints would be the strongest early signal.
- The size, when it is real. 744B to 3T is a fourfold jump. If Z.ai’s next model card says anything in the trillions, that is news on its own; if it says 750B, the rumor’s central number was invented.
- Our own routing. If a new GLM is real, open and better on our axes, we test it on our infrastructure and this page records the result. If it is not, this page records that too.
§ 07The tracker
This page updates when facts change, not when rumors get louder. As of September 6, 2026: no GLM-5.5 is announced, dated, sized, priced or licensed; the leak’s two dated predictions have failed; Z.ai’s own cadence points at a next release between October 8 and November 9, most likely numbered 5.4. When Z.ai moves, the confirmed facts and a graded scorecard of the claims above land here the same day.
§ 08Update log
This is a living page; when the story moves, the update lands here.
As of September 6, 2026: page opened. No official movement to log.
Q1Has Z.ai announced GLM-5.5?
No. We checked every surface Z.ai publishes to. The docs.z.ai release notes list GLM-5.3 (August 18) and GLM-5.3-Flash (August 26) as the newest entries. The zai-org organization on Hugging Face has GLM-5.3 and GLM-5.3-Flash repositories created August 25 and last updated September 4, and nothing newer. OpenRouter’s catalog carries z-ai/glm-5.3 and z-ai/glm-5.3-flash plus two rolling aliases, ~z-ai/glm-latest and ~z-ai/glm-flash-latest, and no glm-5.5. Z.ai’s X account announced GLM-5.3 on August 14, the API on August 18, GLM-5.3-Flash on August 26 and open weights on August 28. The name GLM-5.5 does not appear in any of it.
Q2What exactly did the leak claim, and what happened?
The July 20 version: Z.ai might skip GLM-5.3, launch GLM-5.5 in August, more than 1T parameters, open weights, a 1M-token context, a focus on long-running coding agents. The poster added that none of it was confirmed by Z.ai. What happened: Z.ai shipped GLM-5.3 on August 14 (so nothing was skipped), shipped GLM-5.3-Flash on August 26, opened the GLM-5.3 weights on August 28, and August ended with no GLM-5.5. The two checkable predictions both failed. The size and the rivals kept growing anyway: 3T+ by July 29, and by September 6 the story reads ‘beats Mythos 5.1 and GPT-6 Astra, September window’.
Q3Is 'Mythos 5.1' a real model?
Not under that name. Anthropic’s public model list as of September 6 includes Claude Fable 5.1, Claude Opus 5 and Claude Sonnet 5, and its documentation refers to a separate research model as Claude Mythos Preview. The rumor appears to have merged two things: Anthropic’s September 1 release of Fable 5.1, and Z.ai’s own August 18 statement that GLM-5.3 matches Mythos 5 on white-box cybersecurity tasks. Specifically, Z.ai reports GLM-5.3 at 84.5 percent on CyberGym, ahead of Mythos 5 at 83.8 and GPT-5.6 Sol at 83.6, and at 54.4 percent on ExploitBench, where Mythos 5 scores 78.0 and GPT-5.6 Sol 76.5. A real, mixed comparison got flattened into a slogan.
Q4Why do we think the next GLM will be 5.4, not 5.5?
Cadence and naming. Using OpenRouter catalog dates, which track Z.ai’s launches within a day: GLM-4.5 on July 25, 2025; GLM-4.6 on September 30; GLM-4.7 on December 22; GLM-5 on February 11, 2026; GLM-5.1 on April 7; GLM-5.2 on June 16; GLM-5.3 on August 18. Seven releases, every one a single step up, never a skip, spaced 51 to 83 days apart with a median of 65. Sixty-five days after August 18 is October 22. Z.ai could break its own pattern, and the page will say so if it does, but a rumor that requires the pattern to break twice (a skipped number and a compressed gap) is carrying a lot of weight with no evidence under it.
Q5Should I wait for GLM-5.5 before building on GLM?
No. GLM-5.3 is here, open-weight, 744B total with 40B active, a 1M-token context and 128K output, at $1.40 per million input tokens and $4.40 output on Z.ai’s API. GLM-5.3-Flash is the cheap sibling: 320B total, 18B active, natively multimodal, MIT weights, $0.15 and $0.50 at list with a 50 percent launch discount through September 9. For our part, CellCog does not route to GLM today; our tiers run on Claude Fable 5.1, Claude Opus 5 and Gemini 3.8 Flash, and we say so plainly. We track GLM because it is the open-weight family closest to the frontier, and the engine under an AI employee is a decision we would rather make on dated facts than on a screenshot.