I Traced 9 GLM-5.5 Claims. August Looks Real; the 1T Spec Sheet Doesn't.

On July 20, GLM-5.5 posts started converging on three claims:

"It launches in August."

"It's a 1T-plus open-weight model."

"Z.ai's founder just confirmed it."

Two of those statements overreach. The remaining one is plausible, but it still isn't a release date.

I spent an afternoon checking Z.ai's live release notes, pricing table, API docs, Hugging Face organization, Reuters reporting, and the Chinese article behind the "epic-plus" quote. What I found is more useful than another speculative spec sheet.

TL;DR

• NO: Z.ai has not announced GLM-5.5. There is no official model page, API ID, price row, release note, or model card.
• YES, BUT: August is credible. Reuters reported it as the expected window, and it fits Z.ai's recent 54-70-day release cadence.
• The "epic-plus" quote is real as a media-reported reply, but Jie Tang did not name GLM-5.5 or give a date.
• The 1T-plus parameter number is an analyst forecast. The more specific 1.6T claim is speculation.
• I'd benchmark GLM-5.2 now and prepare a canary. I would not plan a migration around an endpoint that doesn't exist.

What actually exists today

The official record ends at GLM-5.2.

Z.ai released GLM-5.2 on June 16, 2026. Its official documentation lists a 1M-token context and the model ID glm-5.2. Its official Hugging Face card lists 753B parameters, an MIT license, and local-serving instructions.

I checked five places where a real GLM-5.5 launch should appear:

Surface	Latest model	GLM-5.5?
Z.ai release notes	GLM-5.2	No
Z.ai model docs	GLM-5.2	No
Z.ai pricing	GLM-5.2	No
Z.ai Hugging Face	GLM-5.2	No
TokenMix model catalog	GLM-5.2	No

That doesn't prove Z.ai isn't training a successor. It proves you cannot responsibly publish an exact GLM-5.5 API spec today.

I keep seeing pages that list a 1M context, MIT license, August date, and 1T-plus parameters as if those fields came from one document. They don't. They come from different evidence levels:

Claim	My label	Why
A major GLM upgrade is being teased	Confirmed media report	Two Chinese reports recorded "epic-plus"
The product name will be GLM-5.5	Likely	Reuters and analyst reporting use it
It will arrive in August	Likely	Reuters report plus cadence
It will exceed 1T parameters	Analyst forecast	Not in Z.ai docs
It will use exactly 1.6T parameters	Speculation	No primary source
It will be open weight on day one	Speculation	Predecessor precedent only
It will beat a named model by N points	Made up today	No public GLM-5.5 eval exists

The habit I use is simple: if I can't point to the model card, price page, or working request, I don't call the field confirmed.

Why I still take August seriously

The August window isn't random.

Z.ai's three current GLM-5 releases landed on:

Release	Date	Gap
GLM-5	February 12	-
GLM-5.1	April 7	54 days
GLM-5.2	June 16	70 days

The mean of those two intervals is 62 days.

(54 + 70) / 2 = 62 days

June 16 + 62 days = August 17

I am not predicting an August 17 launch. Two intervals are not a law of nature. But the full month of August sits 46-76 days after GLM-5.2, almost exactly around the recent cadence.

Then there is the reporting. A June 25 Reuters story said the next GLM-5.5 model was expected in August. Separately, JPMorgan-linked reporting forecast August and more than one trillion parameters.

Those are useful signals. They are not the same as Z.ai writing "available August 17" in its docs.

My current wording would be:

August is the leading reported window for a model widely called GLM-5.5.

I would not write:

GLM-5.5 launches in August with 1T parameters.

The first sentence survives scrutiny. The second combines two predictions and presents them as a product announcement.

The "epic-plus" quote is narrower than the headlines

This was the most interesting part of the source trail.

According to QbitAI's Chinese report, someone asked Z.ai founder and chief scientist Jie Tang whether GLM still had a response after major Kimi and Qwen upgrades. Tang replied with two words: "epic-plus."

The same report immediately noted what was missing:

• no model name
• no release date
• no parameter count

I like this clue. A founder chose an unusually strong phrase in public. I don't think it means nothing.

But I can't turn two words into a context window, modality list, API price, benchmark score, or license.

There is also a naming wrinkle. Chinese coverage has mentioned GLM-5.3, GLM-5.5, and even GLM-6 as possibilities. Reuters gives GLM-5.5 the strongest evidence, but only Z.ai can lock the name.

My read:

def label_glm_55_claim(claim):
    official_surfaces = {
        "glm-5.2 release date",
        "glm-5.2 1m context",
        "glm-5.2 api pricing",
        "glm-5.2 mit license",
    }

    if claim in official_surfaces:
        return "CONFIRMED"

    if claim in {
        "glm-5.5 is the likely name",
        "august is the leading window",
    }:
        return "LIKELY, NOT ANNOUNCED"

    if claim in {
        "1t+ parameters",
        "open weights on launch day",
        "1m context",
    }:
        return "FORECAST OR SPECULATION"

    if "benchmark score" in claim or "official price" in claim:
        return "NO PUBLIC DATA"

    return "CHECK THE PRIMARY SOURCE"

This function is deliberately boring. Boring is good when every SEO page wants to be first.

The 1T number matters less than people think

Even if the 1T-plus forecast is right, it doesn't answer the questions developers pay for:

• How many parameters are active per token?
• What is the p95 latency?
• Does 1M context remain useful at the end of a long agent run?
• How often do tool calls succeed?
• What is the cost per accepted patch?
• Are weights available, and under which license?

GLM-5.2's official Hugging Face page shows 753B total parameters. Its launch materials focus on a 1M context, long-horizon coding, and IndexShare, an efficiency technique that reuses the same indexer across four sparse-attention layers.

That architecture story is more important than a round total-parameter headline. A larger model with an inefficient serving path can be slower and more expensive. A model with better sparse routing, speculative decoding, and post-training can improve production outcomes without doubling active compute.

So I would treat "more than 1T" as a capacity hypothesis, not a capability claim.

The current price math

There is no GLM-5.5 price. Anyone giving you one is guessing.

There is a useful current baseline. Z.ai lists GLM-5.2 at:

Token type	Price per 1M
Input	$1.40
Cached input	$0.26
Output	$4.40

For 10M input and 1M output:

10 x $1.40 + 1 x $4.40 = $18.40

At 100M input and 10M output:

100 x $1.40 + 10 x $4.40 = $184/month

At 1B input and 100M output:

1,000 x $1.40 + 100 x $4.40 = $1,840/month

The price of a successor matters, but cache behavior may matter more. With 50M input, 80% cache hits, and 5M output:

Without cache:
50 x $1.40 + 5 x $4.40 = $92.00

With cache:
10 x $1.40 + 40 x $0.26 + 5 x $4.40 = $46.40

That's a $45.60 reduction, or 49.6%, without changing the model.

This is why I won't recommend "wait for GLM-5.5 because it will be cheaper." That claim has no evidence, and production cost is more than a sticker price.

The GLM-5.5 launch checklist

Here is what I need before I call it real:

Artifact	What it settles
Z.ai release note	Date and name
Model documentation	Context, output, modalities
Working API example	Model ID and request fields
Pricing row	Input, cache, output
Model card	Weights, license, architecture
Benchmark methodology	Harness, effort, tools, retries
Migration notes	Contract changes
Region/access list	Whether I can actually deploy it

And here is what I'd benchmark:

def should_migrate(glm_52, glm_55):
    if not glm_55.official_model_id:
        return "wait"

    if not glm_55.price or not glm_55.documented_limits:
        return "wait"

    if glm_55.accepted_tasks <= glm_52.accepted_tasks:
        return "keep GLM-5.2"

    if glm_55.cost_per_accepted_task >= glm_52.cost_per_accepted_task:
        return "route only the workloads where GLM-5.5 wins"

    return "start a 5% canary, not a full migration"

Notice what's missing: total parameter count.

What I'd do this week

If I were running GLM in production:

1. I'd freeze a 100-300-task canary set now.
2. I'd record GLM-5.2 tokens, retries, latency, tool success, and human acceptance.
3. I'd move the model ID into configuration if it is hard-coded.
4. I'd keep a GLM-5.2 or cross-provider fallback.
5. I'd poll the official release note, pricing page, and Hugging Face catalog.
6. I'd ignore every GLM-5.5 benchmark screenshot until the methodology and source are public.

If I were experimenting:

1. I'd use GLM-5.2 today.
2. I'd test where its 1M context actually helps.
3. I'd save the exact prompts and expected outputs.
4. I'd rerun those tasks when a documented successor appears.

If I were writing about the model:

1. I'd call August "reported" or "likely."
2. I'd call 1T-plus an analyst forecast.
3. I'd call exact scores, price, API ID, and license unknown.
4. I'd update the same URL after launch.

The bigger picture

This is becoming a recurring pattern in Chinese frontier-model coverage.

A credible analyst or wire-service report names a window. A founder posts a deliberately exciting clue. Secondary pages combine those signals with predecessor specs. Within days, the web has a detailed table for a product that has no official page.

The result is not always malicious. It is often just citation drift:

1. GLM-5.2 has 1M context.
2. GLM-5.5 is expected after GLM-5.2.
3. Therefore a page writes "GLM-5.5: 1M context."
4. Another page cites the first page as confirmation.

That chain creates the appearance of multiple sources while all of them depend on one extrapolation.

The useful response isn't to ignore rumors. Rumors can help developers prepare. The useful response is to keep the label attached to the claim.

If you want the 15-table version with every source and cost scenario, I published the full cited GLM-5.5 breakdown. I also keep the current GLM-5.2 baseline separate so the future model doesn't inherit benchmark numbers it never earned.

If you want to swap among GLM, OpenAI, Anthropic, Google, and other models through one OpenAI-compatible endpoint, that's roughly what TokenMix does. Disclosure: I work on the research side.

Bottom line

GLM-5.5 is probably more than internet fiction. Reuters used the name and an August expectation; a founder teased an "epic-plus" jump; recent cadence points to the same month.

But no official product exists today. I would prepare a canary, not a migration.

Which piece of evidence would change your plan first: a working API ID, an open model card, or independently reproduced coding results?