查看中文版 →

Deutsche Version ansehen →

Открыть русскую версию →

Ver versión en español →

Log inChoose a plan
Blog · Model tests

Stealth model Gestalt and GLM Coding chat leaks: reading the GLM-5.5 launch signals

Gestalt surged to #1 on the Frontier Design Benchmark in November just as a leaked GLM Coding user-group screenshot showed Zhipu staff saying a new model is coming soon. Comparing both signals against Zhipu's Pony Alpha and Ox Alpha playbook shows why developers expect GLM-5.5—and where the proof still stops short.

In mid-to-late November 2026, two separate threads converged in AI developer circles: an anonymous model called Gestalt climbed to the top of frontend design leaderboards, and a screenshot from Zhipu's official GLM Coding user group began circulating with staff teasing an upcoming model release. Put together, they have fueled speculation that GLM-5.5 is right around the corner.

Bottom line: Zhipu staff in the GLM Coding user chat confirmed that a new model is coming soon with ongoing promotional campaigns, echoing an earlier compensation notice promising extra quota resets for the next release. Whether Gestalt is actually the rumored 3T-parameter GLM-5.5 remains unverified—there is no official confirmation and no public tokenizer or error-code fingerprint yet.

Gestalt's run on the design leaderboard

In November, Contra Labs and BenchFlow launched the Frontier Design Benchmark to test models on real client briefs—complete with budgets and multi-layered constraints—rather than toy landing-page prompts. Professional designers judge the generated websites in blind head-to-head A/B votes.

An unannounced stealth entry named Gestalt debuted at #1 on the board:

  • Elo rating: Scored roughly 1,928 to 1,931 across live leaderboard snapshots, comfortably ahead of public frontier models
  • Head-to-head win rate: Won roughly 78% of direct blind matchups against Claude Fable 5.2 and outscored human reference designs overall
  • Broader buzz: Also drew attention across UI and web-generation blind arenas like DesignArena for layout polish and fewer broken styles

Because its lab, model ID, pricing, and weights are hidden, speculation split across communities: some Western developers suspect an unreleased frontier lab checkpoint or multi-model ensemble, while Chinese developer circles largely point to Zhipu's next flagship, GLM-5.5.

Zhipu's stealth-testing track record

Connecting an anonymous leaderboard topper to Zhipu is not pure guesswork—Zhipu has already used this exact pre-launch playbook twice in 2026.

Zhipu's 2026 stealth model codenames and confirmed releases
Stealth codenameFirst spottedPlatform and behaviorConfirmed model
Pony AlphaFebruary 2026Appeared on OpenRouter with strong coding and agentic resultsGLM-5
Ox AlphaAugust 2026Dropped on OpenRouter and OpenCode with 1M context and near-unlimited free testingGLM-5.3-Flash (claimed August 26)
GestaltNovember 2026~1,928–1,931 Elo on Frontier Design Benchmark, 78% blind win rate vs. Claude Fable 5.2Unconfirmed (rumored GLM-5.5)

In both prior launches, Zhipu pushed preview checkpoints under temporary codenames onto public platforms weeks ahead of schedule, gathered real-world blind feedback, and then published model cards alongside API endpoints.

What the GLM Coding group-chat screenshot shows

Alongside the leaderboard run, a screenshot from Zhipu's GLM Coding 用户交流群 (GLM Coding User Group) added weight to the timeline.

Screenshot from the GLM Coding user group showing Zhipu official operations staff replying that the new model will have ongoing promotional activities and answering 'soon' when asked when the new model launches
Screenshot from the GLM Coding user group showing Zhipu operations staff commenting on upcoming model plans and launch timing.

After a member recalled a message that triggered playful replies in the thread, Zhipu's official operations account (智谱官方运营-筛肉) posted two direct hints:

  • Ongoing campaigns: "The new model will have new-model activities," adding that "this is long-term"
  • Launch timing: When asked directly when the new model is coming out, the operator replied simply: "Soon"

That lines up with Zhipu's earlier GLM Coding Plan notice following hardware-node anomalies on some inference servers. Alongside issuing quota reset cards to affected subscribers, Zhipu explicitly stated that users would receive an additional quota reset opportunity once the new model launches.

Rumored GLM-5.5 specifications

Industry notes and Chinese tech media reports cite several unconfirmed details about GLM-5.5:

  • Parameter scale: While GLM-5 through GLM-5.3 sit around 740B total parameters, reports point to a 2.5T to 3.5T (roughly 3-trillion) MoE architecture for GLM-5.5
  • Architecture direction: Building on GLM-5.3-Flash's hybrid linear/sparse attention and mHC (Manifold-Constrained Hyper-Connections), expectations center on stronger asynchronous RL, long-context handling, and multi-step agentic engineering
  • Target window: Analysts and leaks generally place the launch window in Q4 2026, around late November or December

Where the evidence stops short

Connecting the dots is tempting, but three gaps keep Gestalt = GLM-5.5 in speculation territory for now:

  • No technical fingerprint: Back in August, developers linked Ox Alpha to Zhipu using reproducible clues—matching tokenizer behavior and Zhipu-specific error codes like 1210. Gestalt has mainly surfaced in blind design evaluations without public API endpoints, headers, or error signatures to inspect
  • Design vs. full-stack coding: Topping the Frontier Design Benchmark measures frontend layout, visual hierarchy, and brief execution—not backend architecture, multi-repo refactoring, or tool-calling reliability
  • "Soon" does not specify the model: The group chat confirms a new GLM Coding rollout and promo structure are close, not that the release is necessarily a 3T-parameter GLM-5.5 rather than an intermediate or specialized variant

What to use right now

Until Zhipu publishes an official model card, pricing sheet, and live API endpoint, production workflows still rely on shipped models:

  • Low-cost GLM workflows and long context: Use glm-5.3-flash or glm-5.3-flashx
  • Production frontend design and multi-file refactoring: Stick with Claude Fable 5.2, Claude Sonnet 5.5, or Gemini 4 Neon

CodingPass includes glm-5.3-flash, glm-5.3-flashx, and those frontier models across Go, Plus, and Pro, metered at official API rates. Once Zhipu ships the next model, we can benchmark real task cost and first-pass success rates directly against them.

Boarding

Ship with live models while waiting for the announcement.

Go, Plus and Pro include GLM-5.3-Flash and GLM-5.3-FlashX at official API rates. No surprise charges, cancel anytime.

  • Request ID on every response
  • Full refund if you've used 5% or less
  • No surprise charges
Choose GoOpenCode guide

$10/mo until you cancel.