The GLM-5.3 launch on August 14, 2026 makes an unusual bet: Z.ai says the model reuses GLM-5.2’s base model exactly as it was, and that every benchmark gain in the announcement — including a Terminal-Bench 3.0 score that jumped from 4.6 to 28.3 — comes from post-training alone. No new architecture, no new pre-training run, no bigger parameter count.
That framing matters beyond one release. If scaled post-training — more reinforcement-learning environments, more diverse tasks, more compute pointed at the same frozen base — can move an agentic coding score by a factor of six, the industry’s center of gravity shifts from pre-training runs to post-training pipelines. It is also, for now, unverifiable by design: the weights that would let anyone check the same-base-model claim are a stated two-week promise, not a download link.
This guide covers what actually shipped, what changed versus what stayed identical, an honest read of a benchmark table that is not a sweep, who graded each claim, the security capability Z.ai says outgrew its training intent, the breaking API change, and the new Coding Plan points system — with every absence (no price, no weights, no OpenRouter listing) verified directly rather than assumed.
- 01Same base model — every gain is post-training.Z.ai states GLM-5.3 uses GLM-5.2's identical base model, scaled through the same IndexShare, SAO, and slime training stack with more environments and compute. Until weights ship, the claim is unverifiable by design.
- 02Big jumps, but not a sweep.Terminal-Bench 3.0 goes 4.6 → 28.3 as an open-source high, yet GPT-5.6 Sol posts 34.6 there and 88.8 on Terminal-Bench 2.1, Sol leads DeepSWE at 72.7, and Kimi K3 takes Toolathlon Verified at 76.5.
- 03Only two rows were graded by third parties.Of roughly nineteen benchmark rows, Artificial Analysis ran GDPval-AA v2 (GLM-5.3: 1769, the top value) and Proximal ran FrontierSWE (78.1). Everything else was scored by Z.ai — with Claude Code 2.1.207 named in the footnotes as the harness for the coding and agentic rows.
- 04Weights, price, and listings are all pending.Weights are promised “in two weeks,” the docs.z.ai pricing table has no GLM-5.3 row, a dedicated GLM-5.3 docs page 404s, and OpenRouter's newest z-ai listing is still glm-5.2 — each absence checked directly at launch.
- 05The API contract changed under you.Thinking can no longer be disabled; requests using the old parameter fail. The effort ladder is low / high / max with no medium rung, and the GLM Coding Plan moved to a points system with a narrow weekday peak window.
01 — The LaunchOne sentence carries the whole release. #
Z.ai’s announcement is explicit about what GLM-5.3 is: “It uses the same base model as GLM-5.2 — every gain comes from post-training.” The company kept the training stack it documented for GLM-5.2 — IndexShare for long context, SAO for reinforcement learning on long-horizon tasks, and the open-source async RL framework slime — and scaled it with more environments, more diverse tasks, and more compute. Nothing about the base model’s architecture or pre-training changed, per Z.ai.
The infrastructure side of that scaling carries its own vendor-stated numbers: a claimed improvement of more than 2.3× in end-to-end RL training throughput for long-horizon coding tasks, and training/rollout log-probability divergence controlled at the 1e-7 level, both attributed to slime framework work. Neither is independently checkable — they are internal metrics with no user-facing verification path — but they explain where the “more compute” went.
The launch drew mainstream pickup the same day: Bloomberg covered it under the headline “Z.ai Aims to Catch Anthropic, OpenAI in Coding With New AI Model,” and Unite.AI’s same-day report corroborates the core facts — the same-base-model claim, the two-week weight delay, and the security benchmark figures — with no material contradictions against the primary announcement. If you want the baseline this release is measured against, our breakdown of how GLM-5.2 stacked up against Opus 4.8 covers the predecessor’s launch in detail.
“Scaling post-training is all we did for GLM-5.3.”— Z.ai, GLM-5.3 launch announcement, August 14, 2026
cannot be independently verifieduntil the weights actually ship. Z.ai’s own wording is future-tense: “We will release the weights in two weeks after launch, once safety evaluation and hardening are complete.” At launch, GLM-5.3 is a closed model with an open-weight promise attached.
02 — The Release MatrixWhat changed versus what didn’t. #
The cleanest way to understand a post-training-only release is a before/after matrix. Most rows below read “unchanged” — which is exactly Z.ai’s point. The rows that did move fall into three buckets: scores that post-training shifted, contract changes that will break existing integrations, and things that existed for GLM-5.2 but do not yet exist for GLM-5.3.
| Dimension | GLM-5.2 (baseline) | GLM-5.3 (this launch) | What moved |
|---|---|---|---|
| Identical by design — the launch thesis | |||
| Base model | GLM-5.2 base | Same base, reused as-is | Nothing — vendor-stated, unverifiable until weights ship |
| Training stack | IndexShare · SAO · slime | Same stack, scaled harder | More environments, more diverse tasks, more compute |
| Moved by post-training — all scores as reported by Z.ai | |||
| Terminal-Bench 3.0 | 4.6 | 28.3 | +23.7 — roughly 6× the predecessor; still under GPT-5.6 Sol’s 34.6 |
| FrontierSWE | 67.5 | 78.1 | +10.6 — evaluated by Proximal, not Z.ai; Fable 5 posts 88.2 |
| CyberGym | 77.2% | 84.5% | +7.3 — described by Z.ai as the best result on the benchmark |
| ExploitBench | 24.4% | 54.4% | More than doubled; closed frontier still 22–24 points ahead |
| Z.ai Code Bench (private) | Baseline | +50% reported | In-house benchmark — no external replication is possible |
| Changed at the contract level — breaks existing callers | |||
| Thinking toggle | disabled supported | disabled removed — requests fail | Migration required before swapping the model ID |
| Effort ladder | Prior scheme | low / high / max , default max | No medium rung; Z.ai recommends max for coding |
| Coding Plan quota | Rolling-window multipliers | Points system + peak window | Off-peak hours billed at 50% of standard points |
| Not yet available for GLM-5.3 — each absence verified directly | |||
| Open weights | Public | Promised “in two weeks” | Closed at launch, pending safety evaluation and hardening |
| Per-token API price | $1.40 in / $0.26 cached / $4.40 out | None published | No GLM-5.3 row exists in the docs.z.ai pricing table |
| OpenRouter listing | Listed | Absent | Newest z-ai entry on the public models API is glm-5.2 |
Sources: Z.ai’s GLM-5.3 announcement for every score and contract change; the docs.z.ai pricing page and OpenRouter’s public models API for the availability rows, both checked directly at launch. GLM-5.2 prices are the predecessor’s published per-million-token rates — they are not GLM-5.3’s, because GLM-5.3 has none yet.
03 — BenchmarksBig jumps, honest losses — not a sweep. #
The most striking chart in the release is GLM-5.3 against its own predecessor. Because the base model is unchanged, every delta below is — per Z.ai — attributable to post-training. Terminal-Bench 3.0 roughly six-folds. ExploitBench more than doubles. Even the rows that were already strong moved.
Same base model, new scores · GLM-5.2 vs GLM-5.3
Source: Z.ai GLM-5.3 announcement, August 14, 2026 — vendor-reported; FrontierSWE rows evaluated by ProximalNow the other half of the table — the half most launch coverage skips. Z.ai published rows it loses outright, and there are several. GPT-5.6 Sol leads both Terminal-Bench generations. Opus 4.8 leads NL2Repo by 11.7 points and SWE-Marathon v1.1 by 6.3. Kimi K3 takes Toolathlon Verified. On Agents’ Last Exam CLI the gap to Sol is a single tenth of a point — 28.5 versus 28.6 — which we’d call a tie, not a win for either side.
Contested rows · who leads each benchmark, per Z.ai's own table
Source: Z.ai GLM-5.3 announcement benchmark table, August 14, 2026 — all rows vendor-reported, evaluated in Z.ai's harnessOne efficiency claim deserves its own sentence, hedges included. On Z.ai’s in-house Code Bench at High effort, GLM-5.3 reports 31.4% at roughly 50K output tokens, against Opus 4.8’s 29.5% at roughly 120K — about 2.4× the tokens for a slightly lower score. But the same sentence in the announcement concedes GLM-5.3 remains behind Fable 5’s 39.5% at Max effort, and the whole comparison lives on a private benchmark nobody outside Z.ai can rerun. Treat it as a directional token-efficiency signal, not a standings change.
“Fable 5 (w/ fallback)”— while the prose two paragraphs later attributes the identical numbers (83.8% on CyberGym, 78.0% on ExploitBench) to a model it calls
*“Mythos 5.”*Both names appear in the same official document for the same values. We are reporting the inconsistency, not resolving it — the document itself gives no basis to say which label is authoritative.
04 — The Verification LadderWho actually graded each claim. #
Z.ai’s footnotes are unusually transparent — harness, context length, timeout, and temperature are specified per-benchmark. Read closely, they also reveal the release’s biggest caveat: out of roughly nineteen benchmark rows, exactly two were evaluated by named third parties. Everything else — including both headline claims, the Terminal-Bench 3.0 jump and the security numbers — was scored by Z.ai itself. The footnotes name Claude Code 2.1.207 as the harness for the coding and agentic rows — Terminal-Bench 2.1 and 3.0, DeepSWE, ALE-CLI, SWE-Marathon, PostTrainBench, CyberGym, ExploitGym and ExploitBench — with Z.ai running every competitor’s model itself, mostly at reasoning effort max.
| Benchmark | GLM-5.3 result | Evaluator & harness | Outside replication today? |
|---|---|---|---|
| Independently evaluated — 2 rows of ~19 | |||
| GDPval-AA v2 | 1769 — top of the row (Fable 5: 1743, Sol: 1730, Opus 4.8: 1588) | Artificial Analysis (third party) | Strongest data point in the table — the only non-self-graded win |
| FrontierSWE | 78.1 (GLM-5.2: 67.5 · Fable 5: 88.2) | Proximal — 1M context, max effort, 128K output cap | Third-party run, result reported via Z.ai’s post |
| Self-run by Z.ai — including both headline claims | |||
| Terminal-Bench 2.1 / 3.0 | 88.2 / 28.3 | Z.ai · Claude Code 2.1.207, effort max | Public benchmark — replicable once API access is priced or weights ship |
| DeepSWE · ALE-CLI · SWE-Marathon | 66.9 · 28.5 · 42.5 | Z.ai · Claude Code 2.1.207 | Same caveat — Z.ai ran every competitor’s model itself |
| CyberGym · ExploitGym · ExploitBench | 84.5% · 105/130 tasks · 54.4% | Z.ai · Claude Code 2.1.207; ExploitGym budgets throughput-normalized with Artificial Analysis TPS figures | Methodology partially third-party-sourced; the runs are Z.ai’s |
| Z.ai Code Bench | +50% vs GLM-5.2 | Z.ai — private, in-house benchmark | Not replicable by anyone, by design (contamination avoidance) |
To be fair to Z.ai on two counts: publishing rows you lose is a real credibility signal most vendors skip, and per-benchmark methodology footnotes are more disclosure than the industry norm. But transparency about method is not the same as independent verification, and at launch no independent reproduction of GLM-5.3’s numbers exists — expected for a same-day release, and worth restating plainly. Unite.AI’s report lands on the same open question: whether independent evaluators replicate these numbers will determine how much of this launch is a genuine step for open-weight coding models, and how much is evaluation choice.
05 — Security CapabilityThe capability that outgrew its training intent. #
The most consequential subplot of this release is not a coding score. Z.ai says it deliberately added vulnerability-discovery data and environments to post-training, expecting incremental gains in security reasoning. Its own account of what happened next: “We expected this to make the model better at finding and reasoning about vulnerabilities. What surprised us was how quickly the capability continued to develop as training scaled.” That is Z.ai’s narrative of intent versus outcome — a claim about its own surprise, not an independently checkable fact.
The numbers behind it, all vendor-reported: ExploitBench more than doubled from 24.4% to 54.4%, and ExploitGym task completions rose from 29 within a two-hour budget and 39 within six hours on GLM-5.2 to 105 and 130 on GLM-5.3. In both cases the closed frontier stays well ahead — 78.0% and 76.5% on ExploitBench for the disputed-name frontier column and GPT-5.6 Sol respectively, and 181/247 and 216/293 on ExploitGym. GLM-5.3’s two-hour ExploitGym count is still under half of Sol’s.
self-criticalsentence to find in a launch post, and the honest way to read the security rows.
Z.ai’s response to holding this capability is a public disclosure program: a security ledger at cvd.z.ai that tracks AI-assisted vulnerability findings through a structured embargo process. The ledger’s stats are Z.ai’s own, but the live ledger site and the announcement quote matching figures, and its “recently added” feed lists named CVE IDs — among them a Linux kernel 6lowpan use-after-free (CVE-2026-64452), a WebKit/Safari memory-handling bug (CVE-2026-43663), a FreeBSD ptrace parameter-validation flaw (CVE-2026-45253), a GStreamer heap out-of-bounds write (CVE-2026-59691), a Suricata SMTP/MIME parsing bypass (CVE-2026-57229), and a Joomla stored XSS (CVE-2026-48952) — identifiers the ledger records as assigned CVEs rather than vendor-invented labels.
Vulnerabilities in the ledger
53 publicly disclosed, 2,383 under embargo. Severity split: 107 critical, 990 high, 1,286 medium, 53 low — 1,097 rated critical or high. All figures are Z.ai's own ledger stats, consistent between the announcement and the live site.
Open-source codebases
Findings span the Linux kernel, WebKit, FreeBSD, GStreamer, Suricata, and Joomla, each disclosed through a structured embargo process, with the ledger recording an assigned CVE ID where one exists.
Before discovery
The ledger's oldest flaw dates to 1981, and Z.ai puts the mean latency across findings at 26.6 years — the average bug sat undiscovered for over two and a half decades before AI-assisted analysis surfaced it.
The framing matters: this is a disclosure story, not an offensive one. The model surfaces flaws; a human-run embargo program discloses them responsibly; patches follow through the normal CVE process. Z.ai gives only “safety evaluation and hardening” as the reason for the two-week window; it does not name this capability as the cause, though the timing is the obvious context. That tension — hardening the very artifact you have promised to open — is the part of this launch worth watching over the next fortnight.
06 — Breaking ChangeThinking can no longer be disabled. #
The most practical section of the announcement is three lines of migration guidance. GLM-5.3 drops support for thinking.type: "disabled"
entirely. Per Z.ai’s docs, any application still sending that parameter must change it to enabled
and set reasoning_effort
to low
before updating the model ID to glm-5.3
— otherwise the request fails outright. There is no non-thinking mode anymore; effort low
is the new floor.
The effort ladder itself is unusually short: three rungs, no medium. If you followed our guide to running GLM models inside Claude Code, that setup is exactly the kind of integration this change can silently break — audit the thinking parameter before you swap model IDs.
Effort low
The migration target for anything that previously ran with thinking disabled. Thinking still happens — there is no way to turn it off — but at the smallest budget the ladder allows.
Effort high
The only intermediate step — GLM-5.3 has no medium rung, unlike several other vendors' ladders. The token-efficiency comparison Z.ai published against Opus 4.8 was measured at this level.
Effort max
The default — and Z.ai's explicit recommendation for coding tasks. Most of the benchmark table was run at this level, so published scores assume it.
07 — Pricing & AccessThree absences, each verified directly. #
Launch coverage tends to assert availability facts without checking them. We checked each one directly, and all three come back the same way: GLM-5.3 exists as a service, not yet as a priced, listed, downloadable artifact.
No per-token API price exists. The Text Models table on Z.ai’s docs pricing page lists GLM-5.2 at $1.40 input / $0.26 cached / $4.40 output per million tokens, alongside GLM-5.1, GLM-5, GLM-5-Turbo, and older GLM-4.x models — and contains no GLM-5.3 row as of publication, so treat any per-token number quoted elsewhere as unsourced until a row appears.No dedicated docs page exists. The URL pattern Z.ai used for prior releases (a model page under docs.z.ai/guides/llm/) returns a 404 for glm-5.3 at launch.No OpenRouter listing exists. A direct query of OpenRouter’s public models API at launch returns twelve z-ai and GLM model IDs; the newest is glm-5.2 and its batch variant. We did not check Artificial Analysis’s listing directory either way — its role in this launch is as the evaluator of one benchmark row, not a marketplace listing.
Where does that leave access? The API serves GLM-5.3 now — priced, for the moment, only through subscription quota rather than a published per-token rate. The GLM Coding Plan includes it (next section). And self-hosting is a two-week wait at minimum: if you are sizing hardware for the weight drop, our look at the hardware reality of self-hosting a GLM flagship is the right starting point, and our GLM API access and price comparison maps the provider landscape the 5.3 listing will eventually join.
08 — Coding Plan & ZCodeA points system with a narrow peak window. #
GLM-5.3 rolled out to all existing GLM Coding Plan subscribers at no extra cost — and the plan’s quota mechanics were restructured at the same time. Two separate changes landed together, and they are worth keeping distinct.
First, the new baseline: points. Quota is now calculated as points, computed separately for input tokens, cached input tokens, and output tokens. Peak hours are Monday through Friday, 14:00–18:00 UTC+8 — four hours a day, five days a week. Every other hour, including all of Saturday and Sunday, is off-peak and billed at 50% of standard points. Convert the window before assuming it misses you: 14:00–18:00 UTC+8 is 08:00–12:00 CEST, so the European morning sits inside peak, while in the Americas it falls overnight — US working hours sit entirely off-peak, at half the metered cost of the same work. This replaces the older rolling-window multiplier system — descriptions of that scheme are now out of date.
Second, a time-boxed promo on top. ZCode, Z.ai’s agentic development environment, carries a 1.5× limited-time quota boost through August 31, 2026, which Z.ai says stacks with cache savings for “up to 180%” of standard quota — alongside a claimed 98%+ cache hit rate billing cached context at the lower rate for roughly 30% more effective tokens. The boost is a dated offer, not a permanent rate; the points system is the durable structure. Our GLM Coding Plan value breakdown covers the plan economics this restructure just rewrote, and our guide to ZCode, Z.ai’s agentic dev environment covers the tool the promo applies to.
points system— with its off-peak 50% rate — is the new permanent baseline. The
1.5× ZCode boostis a launch promotion with an August 31, 2026 end date. Model your September costs on the points system alone.
09 — ImplicationsWhat this launch means for your stack. #
Four situations, four different answers. The common thread: what you can act on today is the service and the contract change; what you should wait on is everything the launch only promised.
You already have GLM-5.3
The rollout reached all subscribers at no extra cost. Test it at effort max — the level the published scores assume — and learn the peak window: outside Mon–Fri 14:00–18:00 UTC+8, the same work costs half the points.
Migrate before you swap
If any caller sends thinking.type: disabled, requests against glm-5.3 fail. Switch to enabled with reasoning_effort low first, then change the model ID. And budget carefully — no per-token price exists yet to forecast against.
Nothing to download today
Weights are a stated two-week promise pending safety evaluation and hardening. Until they land, GLM-5.2 remains the newest GLM you can actually run on your own hardware. Size your cluster now; deploy when the promise is kept.
Treat the table as a claim
Two of ~19 rows were third-party-graded; no independent reproduction exists yet. Run your own harness on your own repos before routing work — the same discipline we apply to every launch, this week's included.
The trend worth interpreting: this is the clearest public test yet of post-training as the primary scaling axis. GLM-5.2’s launch story was a new open-weight flagship; GLM-5.3’s is the same artifact plus more RL environments, more diverse tasks and more compute — and the deltas on long-horizon agentic benchmarks are larger than most generation-over-generation jumps. If independent evaluators replicate even most of the reported gains once weights land, the lesson for every lab — and every buyer — is that the frozen base model matters less than the post-training pipeline wrapped around it. It also lands in a crowded week: this week’s other model launch competed on price, while GLM-5.3 competes on method.
Looking forward, the two-week window is the story to watch. If the weights ship on schedule and the community reproduces the Terminal-Bench and FrontierSWE numbers, GLM-5.3 becomes the strongest open-weight coding claim of the quarter and a genuine routing candidate for agentic workloads. If the window slips, or reproductions come in materially lower, the launch reads differently in hindsight. For teams deciding how — or whether — a model like this fits their delivery pipeline, our AI transformation engagements start with exactly this kind of evidence-first evaluation rather than benchmark-table vendor picks.
10 — ConclusionA launch built on a promise and a premise. #
The premise is post-training; the proof arrives in two weeks.
GLM-5.3 is two things at once. As a service, it is real today: live on the API and the Coding Plan, with a six-fold Terminal-Bench 3.0 jump, a genuinely honest benchmark table that publishes its own losses, and a breaking API change you need to handle before you touch the model ID. As an open-weight release, it is entirely prospective — a two-week promise, no published per-token price, no OpenRouter listing, and no independent reproduction of any headline number.
The honest reading of the scores: GLM-5.3 wins the open-source agentic-coding story on several long-horizon evals and takes GDPval-AA v2 outright — the one of the two third-party-graded rows it actually leads — while GPT-5.6 Sol, Fable 5, and Kimi K3 each keep rows of their own, and the closed frontier stays more than 22 points ahead on ExploitBench, the row Z.ai itself calls its fastest-growing gap. Not a sweep; not dismissible either.
The bigger signal is the method. If a frozen base model plus scaled post-training can rebuild a coding ladder this thoroughly, the competitive question for the next year is not who pre-trains the largest base — it is who runs the best post-training pipeline on the base they already have. Two weeks from now, when the weights either land or don’t, we will know how much of that thesis survived contact with independent evaluation.