GLM 5.3 benchmarks, pricing, and open weights

What changed from GLM 5.2 to GLM 5.3

Z.ai built GLM-5.3 on the same base model as GLM-5.2. It did not retrain from scratch.

Post-training is the stage after pretraining, where a finished model is trained further on tasks and rewarded for correct results. Z.ai says it spent the month before launch scaling that stage: more training environments, more varied tasks, and more compute spent on them.

The company's headline claim is a 50% improvement over GLM-5.2 on Z.ai Code Bench. That benchmark is Z.ai's in-house test, so the 50% is the vendor's own figure. The other benchmarks named in the same post move the same way:

• Terminal Bench 3.0 rises from 4.6 to 28.3, about six times higher.
• DeepSWE v1.1 rises from 46.2 to 66.9.
• Agents' Last Exam rises from 23.8 to 28.5.

Terminal Bench 3.0 is the big jump. GLM-5.2 scored 4.6 on it and GLM-5.3 scores 28.3 on the same base. The model that came before it is covered in GLM 5.2 benchmarks, pricing, and Claude Code setup.

GLM 5.3 benchmarks against Opus 4.8 and Kimi K3

Z.ai's own table puts GLM-5.3 ahead of Opus 4.8 on five of the seven rows below. The rows cover coding, security and tool use, and every number comes from Z.ai's own runs.

Z.ai's comparison table, selected rows (vendor-run)

Benchmark	GLM-5.3	GLM-5.2	Kimi K3	Opus 4.8
Terminal Bench 2.1	88.2	81.0	88.3	85.0
Terminal Bench 3.0	28.3	4.6	17.4	21.1
DeepSWE v1.1	66.9	46.2	67.5	58.0
SWE-Marathon v1.1	42.5	19.4	48.1	48.8
Toolathlon Verified	73.0	59.9	76.5	76.2
CyberGym	84.5	77.2	80.0	78.1
ExploitBench	54.4	24.4	32.2	40.0

GLM-5.3 beats GLM-5.2 on every row. Opus 4.8 is behind it on five of the seven rows and ahead on SWE-Marathon v1.1 (48.8 against 42.5) and Toolathlon Verified (76.2 against 73.0).

Kimi K3 gives a split result. GLM-5.3 leads it on Terminal Bench 3.0, CyberGym and ExploitBench, and trails on the other four rows.

GPT-5.6 Sol scores higher than GLM-5.3 on Terminal Bench 3.0 (34.6 against 28.3) and DeepSWE v1.1 (72.7 against 66.9), among other rows. Kimi K3 benchmarks, pricing and open weights covers that model.

Does GLM 5.3 use fewer tokens than Opus 4.8?

Yes, on Z.ai's own coding test. At High effort, GLM-5.3 scores 31.4% using about 50K output tokens per task. Claude Opus 4.8 scores 29.5% and uses 120K.

Effort is the setting that tells the model how long to think before answering, and the API offers low, high and max. At Max effort, GLM-5.3 reaches 34.5% at roughly 75K output tokens, while GLM-5.2 reaches 23.4% at 96K.

Z.ai Code Bench score by model and effort level

Bars show the percentage score on Z.ai's in-house benchmark (vendor-run). Labels give output tokens per task where Z.ai states them.

High effort

31.4GLM-5.3 (about 50K tokens)

29.5Opus 4.8 (120K tokens)

0102030

Max effort

34.5GLM-5.3 (about 75K tokens)

23.4GLM-5.2 (96K tokens)

39.5Fable 5

0102030

Output length matters because the Coding Plan counts output tokens in its points. Z.ai also states the limit of the claim: Claude Fable 5 reaches 39.5% at Max effort and stays ahead of GLM-5.3.

Security: leads on CyberGym, trails on ExploitBench

GLM-5.3 posts the best CyberGym score listed, but trails the closed models on ExploitBench. CyberGym gives a model a project's source code and asks it to find a flaw and prove it by triggering a fault.

GLM-5.3 scores 84.5%, up from 77.2% for GLM-5.2. Z.ai calls that the best result on the benchmark, ahead of Mythos 5 at 83.8% and GPT-5.6 Sol at 83.6%.

ExploitBench asks for deeper reasoning about real vulnerabilities and how they are exploited. GLM-5.3 reaches 54.4%, more than double GLM-5.2's 24.4%, while Mythos 5 and GPT-5.6 Sol score 78.0% and 76.5%. Z.ai's reading is that the further along the exploit chain a test sits, the larger the gain over GLM-5.2 and the wider the gap to the closed models.

Security benchmark scores from Z.ai's launch post

Bars show percentage scores (vendor-run). Mythos 5 is the name the security section uses for the model the table calls Fable 5.

CyberGym

84.5GLM-5.3

77.2GLM-5.2

83.8Mythos 5

83.6GPT-5.6 Sol

0255075

ExploitBench

54.4GLM-5.3

24.4GLM-5.2

78.0Mythos 5

76.5GPT-5.6 Sol

0204060

One naming detail: the comparison table calls that closed model Fable 5, while the security section calls it Mythos 5. The two columns carry the same values, 83.8 on CyberGym and 78.0 on ExploitBench, and the post does not explain the difference in names.

Z.ai also says that, after expert review and deduplication, the model found 2,436 vulnerabilities across 269 projects. Of these, 1,097 are rated medium or high severity. Many had gone unnoticed for years or even decades, the oldest for roughly 40 years.

Independent results from the Ed-o-meter

On the independent Ed-o-meter, GLM-5.3 is the first model on the board to clear all five task categories. Reinvently runs the Ed-o-meter and puts several models through the same 28 real-world tasks, split into coding, data, real work, security and tool use. Unlike the other figures, it does not come from the model's maker.

GLM-5.3 passed 100% of the tasks with a rubric score of 9.3 out of 10. The full run cost $0.28, about a fifth of gpt-5.5's $1.43. The catch is a slower start: a median of 16.3 seconds to the first token, against 13.2 seconds for gpt-5.5.

Reinvently adds its own warning. It ran a limited number of trials, so the confidence interval is wide and the scores do not separate models to the decimal. Treat the result as a signal that GLM-5.3 holds up outside Z.ai's own tests, not as a ranking.

GLM 5.3 pricing: the points-based Coding Plan

The launch post gives no per-token price for GLM-5.3. It gives a quota system instead.

The GLM Coding Plan now runs on points, counted separately for input, cached input and output tokens. Z.ai rolled GLM-5.3 out to all Coding Plan users and says it can be tried in ZCode, Claude Code, OpenCode and other coding agents.

Time of day changes the cost: calls outside peak hours use half the standard points. Peak is 14:00-18:00 UTC+8, Monday through Friday. That is four hours a day on weekdays, so every other hour, weekends included, costs half.

For ZCode, the post claims a cache hit rate of 98% or more. Repeated context is billed at the lower cached rate, which Z.ai says gives about 30% more effective tokens.

API changes: thinking can no longer be disabled

GLM-5.3 offers three thinking effort levels, low, high and max, and no longer lets you turn thinking off. A client that still sends thinking.type set to disabled will fail once it points at the new model. Migrate in this order:

1. Set thinking.type to enabled and reasoning_effort to low.
2. Change the model ID to glm-5.3.
3. For coding work, move reasoning_effort to max, which Z.ai recommends.

The order matters. A request that still disables thinking fails once the model ID changes. Z.ai sets low as the level for the switch and recommends max for coding.

Are the GLM 5.3 weights available?

Yes. Hugging Face lists a zai-org/GLM-5.3 repository and a BF16 build of it, both created on August 25, 2026.

Their license tag is "other", not MIT, so the license text on the model page is the document to read before commercial use. Z.ai's launch post promised the weights two weeks after launch, and August 25 falls inside that window.

Don't confuse Flash with the flagship. GLM-5.3-Flash is a separate model with its own repositories, also created on August 25, 2026, under the MIT license. When a report says "the GLM-5.3 weights are out", check whether it means the flagship or Flash.

Use GLM 5.3 through the API or host the weights yourself

If you pay per token for agentic coding, the API is the simpler route. Z.ai's tests show GLM-5.3 reaching a higher score than Opus 4.8 on less than half the output, and the Ed-o-meter put a full 28-task run at $0.28. Both point to a low cost per finished task, and both come with the caveats above.

If you need to run the model on your own machines, the published weights make that possible. Flash is a second option, released under the MIT license.

Check two things on the day you decide. First, latency: the Ed-o-meter measured a slower first token for GLM-5.3 than for gpt-5.5. Second, the license on the Hugging Face page, which is tagged "other" for the flagship and MIT for Flash.