OpenAI’s ChatGPT desktop client for Linux introduces an interesting possibility for users who have traditionally relied on the browser to access ChatGPT.
At DDR Lab, however, the relevant question was not simply whether the application could be installed on Linux.
That would have been a very short test.
We wanted to know something else:
Can the ChatGPT desktop client remain genuinely useful on constrained or ageing Linux hardware, and can it become part of a complete professional workflow rather than merely run as a technical demonstration?
To answer that question, we tested the client in two very different environments.
The first was deliberately extreme: a lightweight antiX installation on a computer with only 2 GB of RAM.
The second was Linux Trabant 1.1 running on a Lenovo ThinkPad T430 configured as an actual legal workstation.
The results were rather different — and both were useful.
1. First test: ChatGPT with 2 GB of RAM
Our first machine was an HP Stream running antiX Linux with approximately 2 GB of RAM.
This is not hardware anyone would reasonably choose today for a conventional AI workstation.
That was precisely the point.
ChatGPT had previously been accessible through Firefox ESR on the machine. It worked, but the limitations of the hardware became apparent as browser sessions grew more complex, particularly with long conversations.
The Linux desktop client offered an opportunity to test whether separating ChatGPT from the general-purpose browser could improve the practical experience.
Installation succeeded.
More importantly, the application was functional.
Authentication was performed through the browser and control subsequently returned to the desktop client. Once authenticated, ChatGPT could be used for actual tasks rather than merely opened as a demonstration.
The result should not be misunderstood.
We are not recommending 2 GB of RAM as a sensible specification for an AI workstation.
The significance of the test is narrower:
The ChatGPT Linux desktop client proved functional on our 2 GB antiX test platform.
This establishes a useful lower experimental boundary.
It also suggested that the desktop client deserved to be tested in a much more realistic environment.
2. Second test: Linux Trabant 1.1
The second machine was a Lenovo ThinkPad T430.
The T430 belongs to a generation of business laptops introduced in the early 2010s. Our unit has subsequently been upgraded and currently uses:
- 12 GB of RAM in a 4 + 8 GB configuration;
- a 512 GB SSD;
- Linux Trabant 1.1, 64-bit;
- LibreOffice;
- PDF tools;
- Firefox;
- Firefox ESR;
- pCloud integrated into the filesystem;
- professional cryptographic tools;
- and the ChatGPT desktop client.
This was deliberately not a clean benchmarking installation.
It is a working computer.
What is Linux Trabant?
Linux Trabant is an experimental Linux distribution developed within DDR Lab / Trabant Systems.
Version 1.1 is based on Linux Mint but has been adapted and configured around a particular objective: providing a stable, durable and observable environment for professional legal work.
It therefore combines two principles that can initially appear contradictory.
At the user level, the desktop remains deliberately conventional. Files, folders, applications, cloud storage, browsers and office software behave much as a user coming from a traditional Windows workstation would expect.
Below that layer, however, Linux Trabant deliberately exposes more of the machine.
Technical boot messages remain visible. System initialization is observable rather than hidden behind a graphical splash screen. During normal operation, Conky continuously displays processor load, temperature, memory, swap, disk activity and network information.
The intention is not complexity for its own sake.
It is observability.
A professional user should not need to become a Linux administrator to write a document. But when something behaves unexpectedly, the computer should not unnecessarily conceal what it is doing.
3. A legal workstation is not an easy environment
Building a Linux workstation for legal practice in Spain involves more than installing an office suite.
Professional work may require a chain of components involving digital certificates, cryptographic cards, AutoFirma, electronic public-administration services and professional platforms such as LexNET.
Linux Trabant 1.1 includes a working environment configured around those requirements.
In our installation, Firefox ESR has been retained specifically as part of the validated professional environment used with the cryptographic infrastructure, including the ACA professional card.
This matters because a configuration that works reliably should not be changed merely for the sake of using the newest available component.
Stability is itself a feature.
4. ChatGPT authentication and the native browser
An interesting behaviour appeared during authentication.
On Linux Trabant, the ChatGPT client delegated authentication to the normal Firefox installation available through the operating environment rather than specifically invoking Firefox ESR.
After completing authentication, Firefox displayed a request asking whether chatgpt.com should be permitted to open a codex: link with ChatGPT.
Accepting it returned control to the desktop application.
The observed sequence was therefore approximately:
ChatGPT desktop client → system browser → OpenAI authentication → codex: link → ChatGPT desktop client
The earlier antiX test produced a compatible observation. On that machine Firefox ESR was the available browser and was used during authentication.
This suggests that the application can delegate the web authentication stage to the browser made available by the Linux environment rather than requiring one particular Firefox edition.
This description reflects our observed test behaviour. It is not intended as a statement about the application’s undocumented internal architecture.
Authentication behaviour itself also differed between the two machines.
One test involved additional telephone verification. The other completed authentication through the conventional email-code flow.
Again, we regard this as an observation rather than evidence of any Linux-specific authentication requirement.
5. The test stopped being a test
Synthetic benchmarks have their uses.
For this experiment, however, we were more interested in something simpler:
Can the machine actually be used for work?
We therefore used ChatGPT on the ThinkPad for a real, document-heavy legal workflow.
A long existing conversation containing substantial context and documentation was opened. ChatGPT was used to process material relating to a tax matter and to assist with the preparation of a response to the Spanish tax administration.
This was not a short prompt followed by a two-paragraph answer.
It was a long professional conversation containing accumulated context and extensive documentary material.
While that work continued, the computer was also used with LibreOffice Writer, PDFs and browser windows.
Later, the same working session included access to electronic public-administration services using the professional cryptographic environment and ACA card.
pCloud remained integrated directly into the Linux filesystem and available to applications in the usual way.
Several windows remained open.
At some point during this process we stopped consciously testing the computer.
We simply started working on it.
That turned out to be one of the more useful results of the experiment.
6. Resource usage
Linux Trabant’s permanently visible Conky monitor made it possible to observe the machine throughout the test without repeatedly opening a separate system-monitoring application.
The figures below represent snapshots rather than laboratory-grade performance measurements. They should therefore be read as observations of a real working session, not as universal benchmarks for the ChatGPT client.
Fresh working environment
After a fresh start, Linux Trabant used approximately:
1.11 GiB of 11.4 GiB available RAM
with no swap in use.
This provides a useful reference for the underlying workstation before accumulating the normal professional workload.
ChatGPT under active load
During the first controlled ChatGPT workload, total system memory generally remained around:
2.75–3.00 GiB
One snapshot during active processing showed approximately:
- CPU: 16%
- RAM: 2.78 GiB
- swap: 0
- temperature: 75°C
The processor and temperature reacted to activity, while memory remained relatively stable.
Once ChatGPT completed the task, the system returned to approximately:
- CPU: 3%
- RAM: 2.75 GiB
- swap: 0
- temperature: 63°C
There was no observable accumulation of swap during the operation.
ChatGPT, Firefox and LibreOffice
LibreOffice was then added to the active workload while ChatGPT and browser processes remained available.
A transient measurement showed approximately:
- CPU: 38%
- RAM: 3.00 GiB
- swap: 0
- temperature: 73°C
Once activity settled, memory returned to approximately 2.8–2.9 GiB and processor load fell substantially.
Again, swap remained unused.
Extended professional session
The most interesting measurement came later.
By then the ThinkPad was no longer being treated as a benchmark machine. It had accumulated a normal collection of open applications, documents and windows during an extended working session.
The system included ChatGPT, LibreOffice, browser activity, PDFs, cloud storage and professional web services.
Memory usage reached approximately:
3.94 GiB of 11.4 GiB
while swap remained:
0
At that point CPU utilization was low and the computer remained responsive.
This figure is arguably more representative of professional use than the lower controlled measurements.
A freshly prepared benchmark and a workstation that has been actively used for more than an hour are different things.
7. What the numbers do — and do not — tell us
It would be tempting to subtract the fresh-system RAM figure from one of the later measurements and announce that the difference represents the memory consumption of ChatGPT.
We do not think that would be technically responsible.
Linux dynamically uses memory for applications, buffers and caches. Browser processes, document viewers and other components also change their memory behaviour over time.
The measurements therefore describe total system behaviour under particular workloads, not the isolated memory footprint of the ChatGPT executable.
What we can say is simpler.
- Long ChatGPT conversations remained usable.
- Professional document work could continue simultaneously.
- Processor load appeared primarily as transient activity rather than permanent saturation.
- Memory remained manageable.
- The 12 GB workstation never needed to use swap.
For our purposes, those observations are more useful than an artificial single-process benchmark.
8. What about 4 GB?
The obvious question is whether 4 GB would be sufficient.
We do not currently have a 4 GB Linux test machine available with an equivalent configuration.
We therefore did not test it.
The controlled measurements around 2.8–3.0 GiB suggest that a carefully configured 4 GB system might be viable for some workloads.
The extended-session measurement of approximately 3.94 GiB, however, provides an important warning against making that conclusion too casually.
A computer physically limited to 4 GB would manage memory and caches differently and might begin using swap substantially earlier.
For that reason, we make no claim about 4 GB.
The experiment establishes two observed points:
2 GB — functional under severe constraints.
12 GB — comfortable for the complete professional workflow tested on the ThinkPad T430.
Everything between those two configurations remains an open experimental question.
9. Does the SSD explain the result?
The ThinkPad T430 has been upgraded with a 512 GB SSD.
That obviously matters for the overall experience.
An SSD improves boot time, application loading, filesystem operations and responsiveness compared with the mechanical disks originally associated with this generation of hardware.
But it does not explain the observed memory behaviour.
Throughout the tests described above, swap remained at zero.
The SSD was therefore not compensating for exhausted physical memory during these measurements.
It made the workstation faster as a workstation.
It was not rescuing ChatGPT from insufficient RAM.
10. Desktop client versus browser
Before installing the desktop client, ChatGPT was already usable on our Linux systems through conventional browsers.
On the ThinkPad, Chromium provided a generally workable experience.
Long conversations, however, could sometimes become less responsive as sessions grew.
The desktop application handled the long professional conversation used during this test without producing the same degree of noticeable degradation.
That observation should be treated cautiously.
It is based on our particular machines, our particular conversations and our particular Linux configurations. It is not a controlled browser-versus-client performance study and should not be generalized into a universal claim about Chromium.
Nevertheless, it matters in the context of older hardware.
A general-purpose browser on a professional workstation may already be responsible for public-administration portals, PDFs, cloud services and numerous web applications.
Separating the AI workspace from that environment can therefore have practical advantages even where the total hardware resources are not particularly constrained.
11. Cloud storage without changing the workflow
pCloud is integrated into Linux Trabant as part of the filesystem.
From the user’s perspective, cloud files therefore appear naturally within the normal file manager rather than requiring the browser to become another intermediary in everyday document management.
A document can be opened from the pCloud environment in LibreOffice, edited locally and saved through the same filesystem integration while ChatGPT and other applications remain available.
This is not particularly spectacular technology.
That is precisely why it works well in a professional environment.
Infrastructure becomes most useful when the user stops having to think about it.
12. Old hardware, contemporary work
The ThinkPad T430 is old.
The work performed on it during this experiment was not.
The workstation handled:
- generative AI;
- long-context professional conversations;
- cloud storage;
- office documents;
- PDFs;
- digital certificates;
- cryptographic authentication;
- electronic government services;
- and ordinary multitasking.
This is a 2026 professional workflow running on hardware from a much earlier generation.
That distinction is central to the Linux Trabant project.
DDR Lab is not interested in preserving obsolete computers merely because they are old.
The more useful question is whether technically sound hardware becomes functionally obsolete because the work has exceeded its capabilities, or because the software environment has accumulated requirements unrelated to the work itself.
Those are not the same thing.
Linux Trabant attempts to explore that difference.
13. Familiar above, observable below
Linux Trabant does not attempt to reinvent the desktop.
A professional user finds a conventional graphical environment: windows, folders, files, a taskbar, browsers, office applications and cloud storage.
Someone accustomed to a traditional Windows workstation should understand the basic interaction immediately.
The more unusual decisions are underneath.
During boot, Linux Trabant deliberately leaves technical initialization visible rather than replacing it entirely with a graphical animation.
Successful service initialization can be observed. Hardware detection can be seen. An abnormal delay or failure therefore has somewhere visible to occur.
Once the desktop is running, Conky continues that philosophy by exposing basic operational information permanently.
There is admittedly a retro-computing aesthetic to all this.
Some of us grew up with DOS and Windows 3.1. Watching a computer tell us what it is doing does not necessarily feel primitive.
But the decision is functional rather than nostalgic.
During this experiment, the supposedly decorative numbers told us exactly what we needed to know.
We watched CPU utilization rise when ChatGPT was active.
We watched temperature follow it.
We saw memory remain comparatively stable during the controlled workload.
We saw memory increase later as a real working session accumulated applications.
And throughout the experiment we could see that swap remained unused.
The machine was reporting its own behaviour while we worked.
14. A computer does not need to hide that it is a computer
Contemporary operating systems increasingly attempt to abstract the machine from the user.
There are good reasons for this.
Most people do not need kernel messages, memory figures or filesystem details in order to write an email.
Linux Trabant takes a slightly different position.
Complexity that serves no purpose should be removed. Complexity that remains relevant should be observable.
This does not mean forcing technical information on the professional user.
It means not deliberately making useful information difficult to obtain.
The desktop should remain familiar.
The machinery underneath it may remain visible.
That combination — familiar above, observable below — has become one of the design principles of Linux Trabant.
15. The real test
The original experiment was supposed to answer a relatively narrow question:
Does the ChatGPT desktop client work satisfactorily on Linux, including older hardware?
The 2 GB antiX machine answered one part of that question.
Yes, it can be functional under remarkably constrained conditions.
The ThinkPad T430 answered a different question.
On Linux Trabant 1.1, ChatGPT became part of a complete professional environment alongside LibreOffice, PDFs, cloud storage, browsers, digital certificates, the ACA cryptographic environment and electronic public-administration services.
The machine did not merely run ChatGPT.
It ran the work.
Eventually, there was no reason to continue inventing benchmark scenarios because the workstation had already become the workstation on which actual legal work was being performed.
That is the result we consider most significant.
A professional operating environment succeeds when the user stops thinking about the operating environment and returns to thinking about the work.
For this test, Linux Trabant 1.1 reached that point.
Built to last.
Test information
Date: 24 August 2026
Research environment: DDR Lab / Trabant Systems
Test platform 1
- HP Stream
- antiX Linux
- Approximately 2 GB RAM
- Firefox ESR
Test platform 2
- Lenovo ThinkPad T430
- Linux Trabant 1.1, 64-bit
- Linux Mint-based
- 12 GB RAM (4 + 8 GB)
- 512 GB SSD
- Firefox + Firefox ESR
- LibreOffice
- pCloud filesystem integration
- Professional cryptographic environment / ACA
- ChatGPT desktop client
Availability note
At the time of this test, the ChatGPT desktop application for Linux was being distributed as a public-preview product, and the officially documented Linux compatibility matrix did not necessarily encompass every distribution used in our experiments.
The antiX and Linux Trabant results described above should therefore be understood as observed compatibility in our test environments, not as a statement of official support by OpenAI.
Where the ChatGPT web interface does not itself present the Linux download option, users should consult OpenAI’s current official download and support information before installation, since availability, supported distributions and package formats may change during the preview period.
This article documents an independent DDR Lab field test. Trabant Systems and Linux Trabant are not affiliated with or endorsed by OpenAI. ChatGPT and OpenAI are trademarks of their respective owner.