GLM-5.3-Flash vs Claude Opus 4.8: Every Benchmark Explained

Ox Alpha finally has a name: GLM-5.3-Flash.

Z.ai's new open-weight model has 320B total parameters, only 18B active at a time, native image understanding, and a 1M-token context window. It is also released under the MIT license. Z.ai says it beats GLM-5.2 across its reported evaluations while getting surprisingly close to Claude Opus 4.8 on several hard coding and agent tests.

The benchmark numbers are the interesting part.

Here is what they actually mean.

GLM-5.3-Flash benchmark cheat sheet

The simple rule for reading this table:

Don't ask whether GLM has a high score. Ask what job produced that score.

Terminal-Bench 2.1: 84.3 vs Opus 4.8 at 85.0

This is probably the easiest result to understand.

GLM-5.3-Flash: 84.3
Claude Opus 4.8: 85.0

That is only a 0.7-point gap in Z.ai's published comparison.

Terminal-Bench gives an AI agent difficult jobs that must actually be completed inside a terminal. The benchmark covers areas such as software engineering, machine learning, security, data processing, and system work.

Think of a job like this:

A software project will not start.

The model needs to inspect the files, find the problem, change the right thing, run the project again, and check whether the fix worked.

That is very different from simply asking:

“What command should I use?”

The model has to do the job.

An 84.3 result puts GLM-5.3-Flash extremely close to Opus 4.8 in Z.ai's reported setup.

One important detail: Z.ai ran Terminal-Bench using Claude Code as the agent around the model. The model matters, but the surrounding agent setup matters too. So a 0.7-point difference does not mean the two models will behave exactly the same everywhere.

DeepSWE: 63.4 vs Opus 4.8 at 58.0

This result may be even more interesting.

GLM-5.3-Flash: 63.4
Claude Opus 4.8: 58.0
GLM-5.2: 46.2

In Z.ai's comparison, Flash is 5.4 points ahead of Opus 4.8 and 17.2 points ahead of GLM-5.2.

DeepSWE is designed around long software-engineering jobs, not small coding questions.

Its current version contains 113 tasks across 91 real repositories and five programming languages. The tasks are written from scratch and checked with tests that verify whether the software actually behaves correctly.

A simple example:

There is a bug somewhere inside a large project. Find the right files, understand how they connect, make the fix, and avoid breaking another part of the software.

That is why this benchmark is useful.

A model can be great at generating a clean function from scratch and still struggle when it has to understand somebody else's project for a long time.

63.4 suggests GLM-5.3-Flash is strong at this longer kind of coding work.

It does not prove GLM is better than Opus at every coding task. It means GLM wins this specific comparison under Z.ai's reported evaluation setup.

AutomationBench: 26.2 → 48.8

This may be the clearest sign that Flash improved at more than coding.

Z.ai reports:

GLM-5.2: 26.2
GLM-5.3-Flash: 48.8

That is a 22.6-point jump.

AutomationBench tests whether an AI agent can finish a real multi-step business workflow. Zapier built the benchmark around jobs across sales, marketing, operations, support, finance, and HR.

Imagine this task:

Find the correct customer → check their account → find an open support issue → update the CRM → send the right message to the right team.

The model cannot just write a convincing answer saying the work is finished.

AutomationBench checks what actually changed inside the simulated business systems.

This is important because long AI jobs often fail in boring ways. The model finds the wrong person, updates the wrong record, skips one step, or says it finished before the work is actually done.

So the useful message behind 48.8 is:

GLM-5.3-Flash appears much better than GLM-5.2 at keeping track of long, connected jobs.

One caution: Z.ai reports the 48.8 result from its own AutomationBench evaluation. When comparing against Zapier's public leaderboard, make sure the exact evaluation setup and scoring metric match before comparing numbers directly.

Z.ai Code Bench: 29.0 vs Opus 4.8 at 29.5

Another close result:

GLM-5.3-Flash: 29.0
Claude Opus 4.8: 29.5

That looks impressive, but there is an important label attached to it:

Z.ai Code Bench is Z.ai's own internal benchmark.

The test is designed around realistic coding-agent jobs in larger development environments.

Think:

understand the project → find the relevant code → make several changes → test the result → finish the job

rather than:

write one function.

Z.ai reports Flash almost matching Opus 4.8 at its highest effort setting.

That is useful evidence, but because Z.ai created the benchmark itself, it should be treated as supporting evidence, not independent proof that GLM-5.3-Flash equals Opus overall.

GLM can see the work too

GLM-5.3-Flash is also natively multimodal.

That means its useful work is not limited to plain text.

OfficeQA Pro is a good example of why this matters. The benchmark uses real U.S. Treasury Bulletin documents containing dense text, financial tables, and charts. Its Pro set contains 133 difficult questions.

A real task might look like:

Read a long financial PDF, find the correct historical table, identify one number from a specific year, and return the answer.

Z.ai reports 62.4 for GLM-5.3-Flash on OfficeQA Pro.

That gives Flash another useful direction beyond coding: documents, tables, screenshots, interfaces, and other work where important information is visual.

320B parameters, but only 18B active

One specification looks strange at first:

320B total parameters
18B active parameters

The easiest way to picture it is a large team.

GLM has a huge team available, but the entire team does not work on every small step. Only a smaller part is called in each time.

That helps explain how such a large model can be designed for lower running costs.

But GLM-5.3-Flash is not a normal 18B model. The full model is still around 320B parameters. “18B active” describes how much of the model is being used at each step, not the total size of the model.

What GLM-5.3-Flash looks strongest at

The benchmark picture is fairly clear.

Terminal work: very strong.
84.3 puts it almost level with Opus 4.8 in Z.ai's comparison.

Long coding jobs: very strong.
63.4 on DeepSWE beats the Opus 4.8 result reported by Z.ai and is a big jump over GLM-5.2.

Long agent workflows: major improvement.
AutomationBench rises from 26.2 to 48.8 in Z.ai's evaluation.

Documents and visual work: promising.
Native multimodal support means Flash can work with more than plain text.

None of this means:

GLM-5.3-Flash is better than Opus 4.8 at everything.

It beats Opus on some published tests, sits almost level on others, and benchmark results depend on the task and the agent running around the model.

That is the better way to read the launch.

The benchmark rule worth keeping

84.3 means almost nothing until you know what Terminal-Bench asks the model to do.

Read benchmark launches like this:

Terminal-Bench → terminal jobs

DeepSWE → long software jobs

AutomationBench → multi-step business work

OfficeQA → difficult documents

Once you translate the benchmark into a real job, GLM-5.3-Flash becomes much easier to judge.

And the early picture is impressive: a lower-cost, open-weight model is now landing close to flagship closed models on several difficult coding and agent tasks.

Save the cheat sheet. The next time a model launches with a wall of benchmark numbers, read the job before you read the score.