model-release
GLM-5.3 Release: What Shipped and What Changed for Developers
Jessie
COO
August 3, 2026
Updated on August 27, 2026
12 min read
GLM-5.3 is officially released. Z.ai announced it on August 14, 2026 in the blog post "GLM-5.3: Frontier Coding with Emergent Cyber Capabilities". To be precise about naming: this release is GLM-5.3. The "GLM 5.5" and "GLM-6" labels that circulated during the rumor period refer to possible later models — GLM 5.5 remains unannounced and is tracked separately.
August 26 update: the release has moved beyond its subscription-only launch state. Z.ai's official pricing page now lists GLM-5.3 at $1.40 input, $0.26 cached input, and $4.40 output per 1M tokens. EvoLink's route uses model ID
glm-5.3 behind the same unified API key and exposes both Chat Completions and Anthropic Messages endpoints. The staged state below remains useful as release history; it is no longer the current API answer. For live route pricing and code, use the GLM-5.3 API page.August 27 family update: GLM-5.3 Flash is also live as
glm-5.3-flash, adding image, video, and file input at a much lower route price. Use the GLM-5.3 Flash vs GLM-5.3 comparison to decide whether a workload should start on the volume tier or go directly to the text-only flagship.What Z.ai Actually Announced
GLM-5.3 uses the same base model as GLM-5.2 — Z.ai states plainly that "every gain comes from post-training." That makes this an unusual release: no new pretraining run, no parameter-count reveal, and a set of very large benchmark jumps attributed entirely to post-training scale.
| Item | Confirmed value (August 14, 2026) |
|---|---|
| Official name / API model ID | GLM-5.3 / glm-5.3 (from Z.ai's own API example) |
| Base model | Same as GLM-5.2; all gains from post-training |
| Context window / max output | 1M tokens / 128K tokens |
| Modalities | Text in, text out — no vision, a widely reported pre-release community request that did not ship |
| Thinking | Always on: thinking.type: "disabled" is no longer supported |
| Reasoning control | New reasoning_effort parameter: low / high / max (max recommended for coding) |
| Tooling | Function calling, MCP, streaming, context caching, structured outputs (per official docs) |
| Parameter count | Not published (same-base inference ≠ an official figure) |
| Positioning | "Built to Code. Ready for Cyber Defense." |
Sources: Z.ai announcement, BigModel GLM-5.3 model page.
How Access Changed After Release Day
GLM-5.3 is a useful example of a launch where the announcement, subscription product, API, and weights did not all ship on the same day. The release-day state and current state are different:
| Channel | Release-day state | Current state checked August 26 |
|---|---|---|
| GLM Coding Plan / ZCode | Live; Coding Plan requests to 5.2/5.1 auto-routed to 5.3 | Live; subscription economics remain separate from per-token API billing |
| Z.ai international API | Request example existed, but pricing was absent | Priced: $1.40 input / $0.26 cached / $4.40 output per 1M tokens |
| EvoLink unified API | No verified public route on release day | Route configured as glm-5.3; live price and code are on the model page |
| Open weights | Promised after a safety evaluation; license unstated | Verify the current repository and license before planning self-hosting |
The practical consequence today is the opposite of the release-day warning: pricing is now a confirmed input to the adoption decision. The official rates happen to match GLM-5.2, but equal token rates do not guarantee equal task cost because GLM-5.3 always reasons and bills reasoning inside output tokens.
The Benchmark Claims — and Their Caveats
All numbers below are Z.ai's own reported results (vendor-claimed). No independent replication existed on release day.
| Benchmark | GLM-5.3 | GLM-5.2 | Note |
|---|---|---|---|
| Terminal Bench 3.0 | 28.3 | 4.6 | Claimed #1 among open models; Claude Fable 5 = 33.7 |
| DeepSWE v1.1 | 66.9 | 46.2 | |
| FrontierSWE | 78.1 | 67.5 | |
| SWE-Marathon v1.1 | 42.5 | 19.4 | |
| CyberGym | 84.5 | 77.2 | Z.ai's Chinese docs say 83.5% in body text — two official figures circulate |
| ExploitBench | 54.4 | 24.4 | Z.ai itself notes closed models remain ahead (Mythos 5 = 78) |
| Agents' Last Exam (CLI) | 28.5 | 23.8 | Claimed #1 among open models |
Z.ai also reports a "50% improvement over GLM-5.2" on its internal Code Bench — a private benchmark — and openly acknowledges GLM-5.3 "remains behind Claude Fable 5" at max effort. The day-one community reaction on Hacker News captured both poles: "only just by a hair" behind frontier closed models, and "is post-training magic just overfitting to benchmarks?" Both are quotes worth keeping in mind until third-party evaluations land.
The cybersecurity results came with an unusual artifact: Z.ai simultaneously launched a public Security Disclosure Ledger for vulnerabilities the model finds. That helps explain the deliberately staged release and remains relevant even after API pricing became public.
GLM-5.3, GLM 5.5, GLM-6 — Which Is This?
For two months, three names circulated around Z.ai's roadmap. Where each stands after today's release:
- GLM-5.3 — the label used in developer discussion since July. This is what shipped today.
- GLM 5.5 — a separate analyst/media label (JPMorgan forecast via Reuters), often attached to a rumored >1T-parameter flagship. Still unannounced — if it comes, it would be a later release, possibly the one after GLM-5.3. Our GLM 5.5 release watch continues to track it as its own entity.
- GLM-6 — a speculation track with no official standing.
One rumor detail worth closing out: the widely shared GitHub "glm-5.3 branch" URL was never release evidence (the page showed no commit history when checked, and the repository later vanished from public access). The release happened anyway — through an official announcement, exactly the artifact this tracker said to wait for.
What Changed for Developers: The API Contract
Two API-level changes are confirmed in the announcement and matter before any migration:
- Thinking can no longer be disabled. GLM-5.2 accepted
thinking.type: "disabled"; GLM-5.3 does not. Pipelines that relied on non-thinking calls for latency or cost control have a breaking change to handle. reasoning_effortis the new control surface —low/high/max, withmaxrecommended for coding tasks.
Z.ai's official request example:
{
"model": "glm-5.3",
"thinking": { "type": "enabled" },
"reasoning_effort": "max"
}Inside the Coding Plan, the migration has effectively already happened without user action: GLM-5.2 and GLM-5.1 requests are auto-routed to GLM-5.3. For a full switching decision — what improved, what the auto-routing implies, and when staying on GLM-5.2 is the right call — see the dedicated comparison: GLM-5.3 vs GLM-5.2: what actually changed and is it worth switching?
How to Use GLM-5.3 Today
There are now two different buying paths: a points-based coding subscription and a per-token API. Do not compare their headline units directly.
| Your situation | Realistic option today | What to watch |
|---|---|---|
| You use Claude Code / OpenCode / a coding agent | GLM Coding Plan subscription (from $18/mo, points-based quota; GLM-5.3 bills at the highest point multipliers in the family) | Off-peak hours bill at 50% of points |
| You bill by the token via API | Call glm-5.3 through the EvoLink model page; start at reasoning_effort: "low" for routine work | Cost per accepted task, cache hits, output-token growth |
| You self-host | Wait for the weights (~August 28) and the license terms | The Hugging Face zai-org org; license is not guaranteed to match GLM-5.2's |
| You need vision | GLM-5.3 does not have it; use a multimodal route | Whether the GLM-V line converges with mainline GLM |
EvoLink now exposes GLM-5.3 through one gateway alongside other production model choices. Keep the model ID in configuration, validate streaming and tool behavior on your own payloads, and retain a tested fallback; route availability is not a substitute for workload acceptance testing.
What We Still Don't Know
Several production-relevant facts still require workload-level verification even though pricing is now public:
- Rate limits and concurrency. No RPM/TPM figures published.
- Open-weight license. GLM-5.2 was MIT-style; Kimi and Qwen have recently moved to restricted licenses, and the community is watching whether Z.ai follows. Two weeks of safety review also means terms could differ.
- Parameter count. "Same base as GLM-5.2" invites an inference; Z.ai has published no figure.
- Independent benchmarks. All release-day numbers are vendor-claimed.
Signals That Will Trigger the Next Update
- GLM-5.3 weights appearing under huggingface.co/zai-org, with explicit license terms.
- Published rate-limit or concurrency guidance for production sizing.
- Independent, matched-harness evaluations that reproduce or challenge the vendor benchmark table.
- Material pricing or API-contract changes on Z.ai or EvoLink.
FAQ
Is GLM-5.3 released?
Yes. Z.ai officially announced GLM-5.3 on August 14, 2026. Release-day access was subscription-first; per-token API pricing is now published.
How can I actually use GLM-5.3 today?
Use model ID
glm-5.3 through the EvoLink API route, or use Z.ai's GLM Coding Plan/ZCode subscription products. The first is per-token; the latter use plan economics.When will the GLM-5.3 API be available, and what will it cost?
Z.ai's official rates are $1.40 input, $0.26 cached input, and $4.40 output per 1M tokens. EvoLink's current route price and calculator are on the GLM-5.3 API page.
What is the GLM-5.3 model ID?
Use
glm-5.3. The same public model ID is used in the official example and EvoLink route.What does "same base, all gains from post-training" mean?
Z.ai did not pretrain a new model. GLM-5.3 is GLM-5.2's base network with substantially scaled post-training (fine-tuning, RL, and related stages). That is why there is no new parameter count — and why some developers question whether benchmark gains will generalize.
Is GLM-5.3 the same size as GLM-5.2?
Presumably, since the base is the same — but Z.ai has not published a parameter count for GLM-5.3, so no figure should be treated as official.
Does GLM-5.3 have vision?
No. It is text-only. Vision was a heavily reported community request before launch (TechTimes and explainx both covered the developer-poll thread), and it did not ship; Z.ai's multimodal work remains in the separate GLM-V line.
When will the open weights be released, and under what license?
Z.ai promises the weights about two weeks after launch (around August 28, 2026), after safety evaluation and hardening. The license has not been stated — do not assume GLM-5.2's terms carry over.
Is this the same thing as GLM 5.5?
No — they are separate. GLM 5.5 remains unannounced: during the rumor period, analysts used that label for a possible larger flagship, which may still arrive as a later release. Today's release is GLM-5.3, and the separate GLM 5.5 watch continues.
Are the benchmark numbers trustworthy?
They are Z.ai's own reported results, including comparisons where GLM-5.3 trails closed models — a point in their favor candor-wise. But no independent replication exists yet, and one figure (CyberGym) appears as both 84.5 and 83.5 in different official documents. Treat them as vendor claims pending third-party tests.
What breaks if I move a GLM-5.2 integration to GLM-5.3?
The confirmed breaking change: thinking can no longer be disabled. Latency- or cost-sensitive non-thinking call paths need redesign, and
reasoning_effort becomes your control knob. Full migration checklist in the GLM-5.3 vs GLM-5.2 guide.Should I switch from GLM-5.2 now?
If you are on the Coding Plan, 5.2 requests were already auto-routed. For per-token traffic, run a controlled challenger lane on
glm-5.3, remove thinking.type: "disabled", compare accepted-task cost, and keep GLM-5.2 as the rollback route until your gates pass.Sources and Update Log
- Z.ai — GLM-5.3 announcement
- BigModel — GLM-5.3 model documentation
- Z.ai — official GLM-5.3 pricing
- Z.ai — GLM Coding Plan / devpack documentation
- Hacker News — release-day discussion
- Unite.AI — release coverage
- Hugging Face — zai-org (no GLM-5.3 weights yet)
August 26, 2026 — API activation revision. Z.ai's pricing page now lists GLM-5.3 at $1.40/$0.26 cached/$4.40 per 1M tokens. EvoLink's product route uses
glm-5.3 and exposes live pricing, both supported API protocols, migration checks, and code examples. All present-tense “API unavailable” guidance was retired; the release-day state remains only as dated history.August 14, 2026 — release-day revision. GLM-5.3 was officially announced today; this page flipped from rumor tracking to release tracking. Confirmed: name,
glm-5.3 model ID (per official example), 1M/128K limits, text-only modality, always-on thinking with reasoning_effort, vendor benchmark table, Coding Plan availability with auto-routing from GLM-5.2/5.1, ~2-week open-weight timeline. Still open at that checkpoint: per-token pricing, API open dates, rate limits, license, parameter count, independent evaluations.August 3, 2026 — first publication. Documented the rumor-period evidence: the GitHub commits URL that showed no commit history (archived snapshot; the repository later vanished — correctly treated as a signal, not release evidence), the unresolved GLM-5.3 / GLM 5.5 / GLM-6 naming split, and the recommendation to ship on a verified GLM-5.2 route while treating the successor as an evaluation candidate. That recommendation held at the release-day checkpoint, when the per-token API was the part that had not yet shipped.
Evidence last reviewed August 26, 2026. This page updates when pricing, API behavior, weights, license, or independent evaluations change.
