GLM-5.3

GLM-5.3 is Z.ai's flagship language model, released August 18, 2026. It accepts text only, carries a 1 million token context window and a maximum output length of 128,000 tokens, and always runs with reasoning enabled at one of three effort levels. Z.ai states that GLM-5.3 uses the same base model as GLM-5.2, with every improvement coming from post-training. On version 4.1.1 of the Artificial Analysis Intelligence Index, read August 28, 2026, GLM-5.3 at max effort displays 60 against 53 for GLM-5.2 at its own max effort — on identical list rates of $1.40 per million input tokens, $0.26 per million cached input tokens and $4.40 per million output tokens. One change is not backward compatible: disabling reasoning is no longer supported, and a request that still sends the old parameter will fail.

Our Verdict

GLM-5.3 scores 8.6 out of 10 in our assessment. The headline is unusual enough to state plainly: Z.ai moved its flagship up a generation on an independent index and did not move the invoice. For anyone already budgeting at GLM-5.2 rates for agentic coding, that is close to a free upgrade on the per-token line. Three things keep it below a nine. Reasoning can no longer be turned off, which removes a mode some teams were using deliberately for cheap high-volume calls. The cost measured per completed task went up even though the rate per token did not, because the model emits more output tokens at its default setting. And we could not find published GLM-5.3 weights, so a self-hosted deployment cannot be built on this release the way it can on GLM-5.2.

How we assessed GLM-5.3 without running it

We have not run GLM-5.3 ourselves. We hold no Z.ai API credentials, so nothing on this page is a hands-on measurement by us, and we are not going to write it as one. What this review does is assemble what can be verified from primary sources and say which party is behind each number. Every specification, price and parameter below comes from Z.ai's own developer documentation at docs.z.ai, read August 28, 2026. Every index score, cost-per-task and throughput figure comes from Artificial Analysis, an evaluator with no commercial relationship to Z.ai, read the same day. Where a claim exists only in Z.ai's release notes or marketing, we say so in the same sentence as the claim.

That distinction carries most of the weight on a page like this one. A vendor benchmark and an independent one are not two readings of the same thing, and the difference is not a matter of trust — it is a matter of whether anyone outside the lab can reproduce the run. We keep them in separate sections below for exactly that reason.

Our scores are editorial judgments built on that evidence, not stopwatch readings. They are stated with their reasons in the score breakdown further down, so you can disagree with a number by disagreeing with the argument behind it.

What is GLM-5.3?

GLM-5.3 is the latest flagship in Z.ai's GLM series, announced in the vendor's release notes under an entry dated August 18, 2026 and aimed squarely at complex software engineering and long-horizon agent work. Z.ai describes it as delivering comprehensive advancements in those two areas over GLM-5.2.

The most striking sentence in the documentation is one most vendors leave out. Z.ai writes that GLM-5.3 "uses the same base model as GLM-5.2, with all improvements driven by post-training." No new pre-training run is claimed, no new parameter count, no new architecture — a full version bump attributed entirely to what happened after the base model was frozen. Artificial Analysis independently lists the same total and active parameter counts for GLM-5.3 as it does for GLM-5.2, which is consistent with that statement rather than a confirmation of it.

Practically, this means the specification sheet a team engineered around for GLM-5.2 still describes GLM-5.3. Same context window, same output ceiling, same text-only input. The one thing that changed shape is the reasoning interface, which is the subject of its own section below and the single item that belongs on a sprint board.

Z.ai is the international brand of the lab known in China as Zhipu AI. GLM-5.3 reaches production through three documented protocols: an OpenAI Chat Completion endpoint, an OpenAI Response endpoint, and an Anthropic Messages endpoint, which is what makes the model a drop-in for agent harnesses built against either API shape, Claude Code among them.

Specifications

Two entries deserve a second look. The first is the input modality: GLM-5.3 is text-only, in Z.ai's own words it "currently supports text-only inputs." Vision runs through a separate line in the catalog and, as of late August, through the Flash variant covered further down — which is a different model, not a cheaper tier of this one. The second is the output ceiling. A 1 million token context paired with a 128,000 token output limit is a wide funnel with a normal spout, which matters when planning whole-repository work: the model can read the repository in one pass, but it will not write one back in a single response.

The breaking change: reasoning can no longer be disabled

GLM-5.3 always operates with reasoning enabled. Z.ai's documentation states that disabling reasoning is no longer supported, and that an application still sending thinking.type: "disabled" must change it before switching model IDs or the request will fail. This is the one item on the page that can break a working production system, and it breaks quietly: nothing about a model ID swap announces that a parameter has been retired.

The migration order Z.ai specifies

The sequencing is explicit in the vendor's migration notice, and it is worth following exactly. Change thinking.type to enabled and set reasoning_effort to low before updating the model ID to glm-5.3. Done in that order, the failure mode never reaches production: the first step is harmless against GLM-5.2, and the second step is the actual switch, made only once the first is deployed everywhere.

The request that stops working:

{
  "model": "glm-5.2",
  "thinking": { "type": "disabled" }
}

Step one, still on the old model and already compatible:

{
  "model": "glm-5.2",
  "thinking": { "type": "enabled" },
  "reasoning_effort": "low"
}

Step two, once step one is live everywhere:

{
  "model": "glm-5.3",
  "thinking": { "type": "enabled" },
  "reasoning_effort": "max"
}

Z.ai recommends max for complex tasks such as coding, and max is also what the API applies when the parameter is absent. Be deliberate about that default: an omitted parameter now buys the deepest and most expensive reasoning tier, which is the same configuration the independent cost figures below were measured on.

One more item belongs on the migration checklist. Z.ai's documentation states that anyone who has previously subscribed to a GLM Coding Plan, including an expired subscription, can currently reach the model API only through the OpenAI Chat Completion-compatible protocol. A service written against the Anthropic Messages endpoint may therefore need a transport change on top of the parameter change, which is a larger job than it sounds.

Pricing

Z.ai's pricing page, read August 28, 2026, lists GLM-5.3 at $1.40 per million input tokens, $0.26 per million cached input tokens and $4.40 per million output tokens, with cached input storage marked free for a limited time. The rows for GLM-5.2 and GLM-5.1 carry the same three numbers. A generation of measured capability was added to the same line on the invoice.

The cache discount is the number that decides real bills, and it is steep: $0.26 against $1.40 is an 81 percent reduction on cached input. For any workload that re-sends a large stable prefix — a repository, a specification, a long system prompt — that rate, not the headline input rate, is what the invoice tracks.

Alongside metered access, Z.ai sells the GLM Coding Plan, a points-based subscription. The documentation states that calls made during off-peak hours, which Z.ai defines as including all day on weekends, consume only 50 percent of the standard points. That is a real lever on the subscription side with no equivalent on the per-token side, where the rates are flat around the clock.

Holding a price flat across a capability jump is a decision rather than a default, and it is becoming a pattern worth watching. Anthropic made the same call when Claude Opus 5 launched on the previous flagship's price card, and DeepSeek made it when its cheapest model overtook its own flagship at unchanged rates. In each case the vendor absorbed the improvement instead of repricing it.

Same rate per token, higher cost per task

A flat rate per token does not mean a flat cost per job, and this is the part that surprises budgets. Artificial Analysis measures cost per Intelligence Index task at $0.68 for GLM-5.3 against $0.44 for GLM-5.2, read August 28, 2026 — roughly 55 percent more, on rates identical to the cent. The evaluator also publishes its own totals: $1,238.50 to run GLM-5.3 through the index against $843.44 for GLM-5.2.

The reason is token volume, not price. Reasoning is always on and the default effort is max, so the model thinks at length unless instructed otherwise, and longer reasoning traces carried through multi-turn evaluations grow the input side as well as the output side. The practical translation is short: if your budget is modeled on tokens, nothing changed. If it is modeled on completed tasks, assume the same job costs meaningfully more at default settings, and treat reasoning_effort as the lever that brings it back down.

GLM-5.3-Flash is a different model, released eight days later

GLM-5.3-Flash is not a cheap tier of GLM-5.3. Z.ai's release notes date it August 26, 2026, eight days after GLM-5.3, and describe it as the first native multimodal model in the GLM-5 series. Its documented input modality covers video, image, text and file, where GLM-5.3 is text-only. Z.ai states it has 320 billion total parameters with 18 billion activated, and positions it as delivering stronger intelligence than GLM-5.2 at a much lower cost — not as a downgrade from GLM-5.3.

The prices are on their own schedule. Z.ai's pricing page lists GLM-5.3-Flash at $0.15 per million input tokens, $0.03 per million cached input tokens and $0.50 per million output tokens, and states that the model is available at a 50 percent discount: $0.075, $0.015 and $0.25 respectively. That promotion is dated — it ends at 24:00 on September 9, 2026, in UTC+8, Singapore time. Read from Europe or the Americas, the discount therefore lapses earlier in the local day than a naive reading of the date suggests. Anything budgeted on the promotional rates should be planned against the list rates, which are the numbers that apply afterward.

Artificial Analysis has measured the Flash variant separately: on Intelligence Index version 4.1.1, read August 28, 2026, GLM-5.3-Flash displays 57, below GLM-5.3 at 60 and above GLM-5.2 at 53. That ordering supports Z.ai's positioning. We cover the Flash here because it shares a version number and gets confused with this model constantly; the two are separate products with separate IDs, separate modalities and separate price cards.

What the independent evaluation measured

Artificial Analysis scores GLM-5.3 at max reasoning effort at 60 on version 4.1.1 of its Intelligence Index, read August 28, 2026. That version aggregates nine evaluations, named by the evaluator: GDPval-AA v2, Tau-cubed Banking, Terminal-Bench v2.1, SciCode, Humanity's Last Exam, GPQA Diamond, CritPt, AA-Omniscience and AA-LCR. The same index version, on the same day, displays 53 for GLM-5.2 at max effort. Because both readings come from one version of one index, the seven-point gap is a like-for-like comparison rather than two measurements taken against different rulers.

Two things in that table are easy to misread. Effort levels are separate entries rather than variants of a single number — the evaluator lists Claude Opus 5 at several different scores depending on the effort configuration, so comparing a max-effort score against a medium-effort score describes a settings choice, not a model. And the displayed scores are rounded. GLM-5.3 and Kimi K3 both display 60, but the evaluator ranks Kimi K3 above GLM-5.3, which is what a shared displayed value looks like when the underlying numbers are not equal. GLM-5.3 is inside the leading group on this index; it is not at the front of it, and it is not alone at the top among models from Chinese labs.

Throughput is where we have to be most careful, because it is the figure that moved. On August 28, 2026 Artificial Analysis measured GLM-5.3 at 66.5 output tokens per second and GLM-5.2 at 69.9 — the older model faster, on that day. Our own reading of the same pages on August 22 recorded 93.2 and 92.5, with the ordering reversed. Both readings are honest records of what the evaluator displayed; taken together they say that throughput on this leaderboard is a live measurement that re-runs, not a specification. Treat any speed claim about either model, including ours, as valid only on its stated date.

Z.ai's own benchmarks, and what each is attached to

The strongest numbers Z.ai publishes for GLM-5.3 are its own. That is not a criticism, it is a provenance label: a private benchmark is a legitimate engineering tool and an illegitimate basis for cross-vendor comparison, because nobody outside the lab can run it. Z.ai reports a 50 percent gain over GLM-5.2 on Z.ai Code Bench, an in-house benchmark it describes as private specifically to reduce contamination from public test sets. On public benchmarks it reports Terminal-Bench 3.0 rising from 4.6 to 28.3, DeepSWE v1.1 from 46.2 to 66.9, and Agents' Last Exam from 23.8 to 28.5, all against GLM-5.2.

The cybersecurity claims are the ones circulating most, and they are more layered than their summary. On CyberGym, which starts from white-box source code, Z.ai reports GLM-5.3 at 84.5 percent against GLM-5.2 at 77.2 percent, ahead of the models it names as comparison points. On ExploitBench, which requires deeper reasoning about exploitation, it reports 54.4 percent against 24.4 percent for GLM-5.2 — while placing the closed frontier models it compares against well above, in the high seventies. On ExploitGym, it reports GLM-5.3 completing 105 tasks within two hours against 29 for GLM-5.2, and again places the closed comparison point far ahead. Z.ai draws the honest conclusion itself in the same document: capability is growing fastest exactly where the remaining gap is widest.

That nuance matters, because the release-notes summary — that GLM-5.3 matches a leading closed model in white-box code review and vulnerability discovery — holds for the first of those three benchmarks and not for the other two. A reader who takes the summary without the table will overstate the result by a wide margin.

Separately, Z.ai reports that in collaboration with several security teams the model identified 2,436 vulnerabilities across 269 projects, of which 1,097 were medium-to-high severity, and says it maintains a public disclosure ledger as those findings move through disclosure. That figure comes from real-world testing described by the vendor, not from CyberGym, and the two are frequently merged into one claim that neither supports.

There is a structural reason to keep CyberGym numbers at arm's length regardless of who publishes them. When a headline CyberGym score made the rounds in July, the result turned out to belong to a harness running two models rather than to a single model. CyberGym outcomes are properties of a system — model plus scaffolding plus tool access plus attempt budget — so two honestly reported scores from different labs still cannot be lined up against each other without the harness described. And a comparison against a model behind invitation-only access is, by construction, one almost nobody outside that access list can check.

Are the weights open?

Z.ai describes GLM-5.3 as reaching state-of-the-art performance among open-source models, and the question of whether that describes GLM-5.3's own license or the field it competes against has a clearer answer now than it did a week ago. Searching Hugging Face on August 28, 2026, we found no published GLM-5.3 weights under the zai-org account that released GLM-5, GLM-5.1 and GLM-5.2. We did find GLM-5.3-Flash there, published August 25, 2026, alongside a BF16 variant and a substantial ecosystem of community quantizations that appeared within days.

Artificial Analysis draws the same line independently. It classifies GLM-5.3 as a proprietary model, while classifying both GLM-5.2 and GLM-5.3-Flash as open weights and publishing an openness score for each of those two but not for GLM-5.3.

What we are reporting is the result of a search, not the state of the world. We did not find GLM-5.3 weights in the places where Z.ai has published its previous releases; that is different from asserting none exist, since a release can land on a mirror, behind a gate, or on a platform we did not query. The distinction matters for one specific decision: if a deployment depends on running weights yourself, GLM-5.2 remains the newest GLM flagship we could confirm as downloadable, and GLM-5.3-Flash is now the newest GLM release of any kind with weights we could confirm.

Score breakdown

The overall 8.6 is not the average of those four. It weights the one fact that is both independently measured and financially concrete — a higher index score at an unchanged rate per token — above the operational friction, because the friction is a one-time migration cost while the pricing position persists for as long as the rates hold.

Pros and Cons

The balance sheet, with the provenance of each item attached.

What stands out: a displayed Intelligence Index score of 60 against 53 for GLM-5.2 on the same index version, measured by a third party; list rates identical to the two previous GLM flagships; an 81 percent cache discount that decides the bill on repository-scale prompts; a 1 million token context window; three API protocols including an Anthropic-compatible endpoint; and a vendor willing to state in writing that the gains came from post-training on an unchanged base model.

What gives us pause: the non-reasoning mode is gone with no equivalent; the independently measured cost per completed task is about 55 percent higher despite identical rates; we found no published weights, so self-hosting stops at GLM-5.2; input is text-only; the headline coding and cybersecurity numbers are vendor-run and not externally reproducible; and any organization that has held a Coding Plan subscription is currently limited to a single API protocol.

Use cases

Where the combination of a wide context, permanent reasoning and unchanged rates actually pays:

• Whole-repository refactors that fit inside a 1 million token context in a single pass.
• Long-horizon agentic coding sessions where the model owns a task end to end rather than a single edit.
• Terminal and CLI agent work, the category Z.ai targeted most directly in this release.
• Migrating an existing GLM-5.2 workload to a higher index score without renegotiating a budget.
• Cost-controlled team coding through the points-based Coding Plan, with weekend and off-peak work at half the points.
• White-box code review, treated as a lead generator for human triage rather than as a verdict.
• Long-context document and specification review where reasoning depth is worth paying for.

Who should move, and who should wait

Move if you are already paying GLM-5.2 rates for agentic coding. The same per-token invoice buys a model that measures seven displayed index points higher on an independent evaluation, and the migration is a two-parameter change made in a documented order. Budget for a higher cost per completed task at default settings, and decide deliberately what reasoning_effort you send.

Wait if you rely on a non-reasoning mode. Teams calling GLM-5.2 with reasoning disabled for classification, routing or extraction have no equivalent here. The nearest option is reasoning_effort: low, which is lightweight reasoning rather than no reasoning, and it emits tokens accordingly.

Wait if you need to self-host. Until GLM-5.3 weights appear somewhere verifiable, an on-premises plan cannot be built on this release. GLM-5.2 remains available and is not marked deprecated.

Wait if you need vision. GLM-5.3 is text-only. GLM-5.3-Flash accepts images, video and files, but it is a different model with a different score and a different price card.

Alternatives

GLM-5.2 is the most relevant alternative and the least obvious one, because it is still sold at exactly the same rates. Z.ai has not marked it deprecated anywhere in its documentation, it remains listed on the pricing page, and its MIT-licensed weights are downloadable. If reasoning-disabled calls or self-hosting are in your architecture, staying is a legitimate engineering decision rather than a delay.

Kimi K3 is the closest competitor on the index, displaying the same 60 at max effort and ranked just above GLM-5.3 by the evaluator. Our GLM-5.2 against Kimi K3 comparison was written against the older GLM scores and now understates the GLM side of that matchup.

DeepSeek V4 competes on the cost axis rather than the score axis and remains the cheaper open-weight option; our head-to-head between the two covers that trade-off. Above them, Claude Opus 5, Claude Fable 5 and GPT-5.6 Sol hold the top displayed scores at materially higher rates — see Claude Opus 5 against GLM-5.2 and GPT-5.6 Sol against GLM-5.2 for how that trade prices out. Qwen3.8 Max sits just below on the same leaderboard.

Frequently Asked Questions

What is GLM-5.3?

GLM-5.3 is Z.ai's flagship language model, released August 18, 2026. Z.ai's documentation states that it uses the same base model as GLM-5.2, with all improvements driven by post-training. It accepts text input only, carries a 1 million token context window and a maximum output length of 128,000 tokens, and always runs with reasoning enabled at one of three effort levels: low, high or max.

How much does GLM-5.3 cost?

Z.ai's pricing page, read August 28, 2026, lists GLM-5.3 at $1.40 per million input tokens, $0.26 per million cached input tokens and $4.40 per million output tokens, with cached input storage free for a limited time. Those are the same three rates the page shows for GLM-5.2 and GLM-5.1, so the newer model did not arrive with a price increase.

What breaks when I switch to GLM-5.3?

Any request that sends thinking.type set to disabled. Z.ai's documentation states that disabling reasoning is no longer supported and that such a request will fail. The documented migration path is to change thinking.type to enabled and set reasoning_effort to low before updating the model ID to glm-5.3, so the parameter change is already deployed everywhere when the model swap happens.

Is GLM-5.3 more expensive to run than GLM-5.2?

Per token, no: the rates are identical. Per completed task, yes. Artificial Analysis measured a cost of $0.68 per Intelligence Index task for GLM-5.3 against $0.44 for GLM-5.2, read August 28, 2026. Reasoning is always on and the default effort is max, so a call that omits reasoning_effort buys the deepest tier and emits tokens accordingly.

Does GLM-5.3 replace GLM-5.2?

No. Z.ai has not marked GLM-5.2 as deprecated anywhere in its documentation, and it remains listed on the pricing page at exactly the same rates as GLM-5.3. Both models are sold in parallel. GLM-5.2 also keeps two things GLM-5.3 does not offer: a reasoning-disabled mode and downloadable MIT-licensed weights.

How does GLM-5.3 score on an independent benchmark?

Artificial Analysis displays 60 for GLM-5.3 at max reasoning effort on version 4.1.1 of its Intelligence Index, read August 28, 2026, against 53 for GLM-5.2 at max effort on the same index version. That version aggregates nine evaluations: GDPval-AA v2, Tau-cubed Banking, Terminal-Bench v2.1, SciCode, Humanity's Last Exam, GPQA Diamond, CritPt, AA-Omniscience and AA-LCR.

Is GLM-5.3 the best model from a Chinese lab?

It is inside the leading group rather than alone at the top. On the Artificial Analysis leaderboard read August 28, 2026, Kimi K3 at max effort displays the same 60 and is ranked above GLM-5.3 by the evaluator, since displayed scores are rounded and the underlying values are not equal. Qwen3.8 Max displays 58 and GLM-5.2 displays 53.

Are GLM-5.3 weights open source?

We could not confirm it. Searching Hugging Face on August 28, 2026, we found no GLM-5.3 weights published under the zai-org account that released GLM-5, GLM-5.1 and GLM-5.2. We did find GLM-5.3-Flash published there on August 25, 2026. Artificial Analysis classifies GLM-5.3 as proprietary while classifying GLM-5.2 and GLM-5.3-Flash as open weights. That is the result of a search, not proof that no GLM-5.3 weights exist.

What is the difference between GLM-5.3 and GLM-5.3-Flash?

They are separate models. Z.ai's release notes date GLM-5.3 August 18, 2026 and GLM-5.3-Flash August 26, 2026. GLM-5.3 is text-only; GLM-5.3-Flash is the first native multimodal model in the GLM-5 series, accepting video, image, text and file input, with 320 billion total parameters and 18 billion activated. They carry different prices, and Artificial Analysis displays 57 for the Flash variant against 60 for GLM-5.3.

How much does GLM-5.3-Flash cost?

Z.ai's pricing page lists GLM-5.3-Flash at $0.15 per million input tokens, $0.03 per million cached input tokens and $0.50 per million output tokens. A 50 percent discount is promotional through September 9, 2026, ending at 24:00 in UTC+8, Singapore time, which brings those rates to $0.075, $0.015 and $0.25 while it lasts. Budget against the list rates, since they are what applies once the promotion ends.

Does GLM-5.3 support images or video?

No. Z.ai's documentation states that GLM-5.3 currently supports text-only inputs. Multimodal input runs through separate models in the catalog, including GLM-5.3-Flash and the GLM vision line, and this release does not change that split.

Which API endpoint should I call for GLM-5.3?

Z.ai documents three protocols: an OpenAI Chat Completion endpoint, an OpenAI Response endpoint, and an Anthropic Messages endpoint. There is a restriction attached: anyone who has previously subscribed to a GLM Coding Plan, including an expired subscription, can currently access the model API only through the OpenAI Chat Completion-compatible protocol.

Sources

• Z.ai Developer Documentation — GLM-5.3 (same base model as GLM-5.2, text-only input, context and output limits, reasoning parameters and migration notice, API protocols, Coding Plan points, coding and cybersecurity benchmark figures)
• Z.ai Developer Documentation — Pricing (per-token list rates for GLM-5.3, GLM-5.3-Flash, GLM-5.2, GLM-5.1 and the rest of the line, and the dated Flash promotion)
• Z.ai Developer Documentation — GLM-5.3-Flash (native multimodal input, parameter counts, recommended settings, Coding Plan quota)
• Z.ai Developer Documentation — Release notes (entry dated 2026-08-18 for GLM-5.3 and 2026-08-26 for GLM-5.3-Flash)
• Artificial Analysis — GLM-5.3 (Intelligence Index version 4.1.1 score, cost per task, throughput, total evaluation cost, proprietary classification)
• Artificial Analysis — GLM-5.2 (Intelligence Index version 4.1.1 score, cost per task, throughput, open weights classification)
• Artificial Analysis — GLM-5.3-Flash (Intelligence Index version 4.1.1 score, parameter counts, open weights classification)
• Hugging Face — zai-org (public repository list queried by creation date on August 28, 2026)

Every measurement attributed to Artificial Analysis was read on August 28, 2026 on version 4.1.1 of the Intelligence Index. Independent index scores and throughput measurements are re-run over time and can move; the index version and the reading date are stated wherever a score appears. Prices are vendor list rates read on the same date and exclude any subscription plan discount.