GLM-5.3 claims a 50% coding jump — but there is no API price
Chinese AI company Z.ai released GLM-5.3 on August 14, 2026, saying it lifted coding benchmark scores by 50% without changing the underlying base model and scored 84.5% on the CyberGym security benchmark; access currently runs only through a coding plan starting at $18 a month, with no per-token API rate published
The three lines
- Release — August 14, 2026, by Z.ai of China
- Claimed scores — 50% coding improvement with no base-model change; CyberGym 84.5%, DeepSWE 1.1 66.9%, ExploitBench 54.4%, AutomationBench 48.2%, Terminal-Bench 3.0 28.3%
- Pricing — coding plans at $18 (Lite), $80 (Pro) and $168 (Max) per month, or $12.60, $56 and $117.60 effective monthly on annual billing. No per-token API rate published
Key questions
- What is GLM-5.3?
- A language model released on August 14, 2026 by the Chinese AI company Z.ai, aimed at coding and agentic work. The company says it raised coding performance through post-training alone, without retraining the underlying base model. It follows DeepSeek V4-Pro, covered here on August 14, in a run of Chinese model releases.
- How good is it?
- The published figures are DeepSWE 1.1 at 66.9%, Terminal-Bench 3.0 at 28.3%, CyberGym at 84.5%, ExploitBench at 54.4% and AutomationBench at 48.2%. Security-oriented benchmarks score notably higher than general coding ones. All of these are vendor-reported; no independent third-party verification has been published.
- What does it cost?
- Access currently runs through coding plans only: $18 a month for Lite, $80 for Pro and $168 for Max, or effective monthly rates of $12.60, $56 and $117.60 on annual billing. The per-token API rate — the number any cost calculation actually needs — has not been published.
This page has covered five model releases in the past two weeks: DeepSeek V4-Pro, Gemini 3.7 Flash, LG EXAONE 2, a batch of five Chinese models, and now GLM-5.3 on August 14.
What stands out in this one is not a performance number. It is the blank between the performance numbers and the price list.
1. What was released
| Item | Detail |
|---|---|
| Model | GLM-5.3 |
| Developer | Z.ai (China) |
| Release date | August 14, 2026 |
| Focus | Coding · agentic tasks · security |
| Claim | Performance gains through post-training alone, no base-model change |
The "no base-model change" claim is the center of this release. Rather than training a new foundation model, the company says it layered post-training onto the existing base and lifted coding scores by 50%.
If that holds, the cost structure shifts. That the bottleneck in frontier-model competition is training cost was covered here on August 4 in "What a frontier model is."
2. The published scores
| Benchmark | Score | What it measures |
|---|---|---|
| CyberGym | 84.5% | Security task completion |
| DeepSWE 1.1 | 66.9% | Resolving real software issues |
| ExploitBench | 54.4% | Writing proof-of-concept exploits |
| AutomationBench | 48.2% | Automation task completion |
| Terminal-Bench 3.0 | 28.3% | Multi-step terminal work |
All vendor-reported. No independent verification has appeared. How to read benchmark numbers is set out in this edition's "What a coding benchmark is."
One pattern is worth naming. Security-oriented benchmarks (CyberGym 84.5%, ExploitBench 54.4%) sit far above general coding work (Terminal-Bench 28.3%). The same model posts nearly a threefold spread across task types — in part because each benchmark sets difficulty by its own standard.
3. The blank on the price list
Access today runs through one route: a coding plan.
| Plan | Monthly | Effective monthly on annual billing |
|---|---|---|
| Lite | $18 | $12.60 |
| Pro | $80 | $56 |
| Max | $168 | $117.60 |
What is missing is the per-token API rate — the number any cost calculator needs, and the only unit that permits direct comparison with other models. That unit was covered here on August 15 in "What AI token pricing is."
Publishing subscription tiers while deferring metered rates is a repeating pattern in this market right now. DeepSeek raised API prices on August 16, and peak-hour pricing has appeared alongside it.
4. Why lead with the security score
That GLM-5.3's headline figure is a security benchmark deserves separate attention.
Finding vulnerabilities and writing proof-of-concept code is the same capability on both offense and defense. That is the problem covered here on August 6 in "AI sandbox escape published" and on August 10 in "What a zero-day is."
What safeguards or access restrictions Z.ai has paired with these scores could not be established from published material.
5. What remains unverified
There is no API price. No date has been announced for one.
There is no independent verification. All five benchmark figures are vendor-reported. Until external reproduction appears, they are claims.
The "no base-model change" claim lacks support. Post-training method and data scale were not published alongside it.
Weight availability is unclear. Some weights are described as held back, but the release scope and licensing terms could not be confirmed here.
Earlier coverage: "DeepSeek V4-Pro general release" and "Gemini 3.7 Flash pricing."
Sources
- aireleasetracker — GLM-5.3: Benchmarks, Specs & Release Date
- Memeburn — GLM 5.3 Is Here: Benchmarks, Pricing, Coding and What's New
- Fello AI — GLM 5.3: Benchmarks, Pricing and the Held-Back Weights
- emergent.sh — GLM-5.3 Pricing: What It Costs to Use Z.ai's New Coding Model
- Qubrid AI — GLM-5.3 Benchmarks, Architecture & Pricing (Aug 2026)
- explainx.ai — GLM-5.3 Launch: Benchmarks, Pricing & Access (Aug 2026)