private-local.ai vs ChatGPT: a comparison for business use
Benjamin Spinola, Founder
Published Updated 14 minutes read
private-local.ai vs ChatGPT is, at its core, a choice about where artificial intelligence (AI) runs: on OpenAI's servers or on a machine that your company owns. ChatGPT is a finished product. Its current models are GPT-6 Astra, released on 3 September 2026, and GPT-5.6 Sol; it now works inside Microsoft Excel, PowerPoint and Word, and OpenAI improves it continuously. Everything a user types or uploads, however, is processed by OpenAI, and many users and business owners have serious concerns about whether their intellectual property and client information are safe there.
The difference is not a matter of individual features. ChatGPT is a service that a company uses, whereas a local setup is infrastructure that a company owns: open models running on a machine it has bought, with its material remaining inside the company. That is a larger decision, and our test drive exists so that it can be made on evidence. We run an NVIDIA DGX Spark with an open model and provide a web address and an access key for it, so that you can connect your own tools and run your own tasks before any equipment is bought. The test drive shows whether owning that setup would deliver the results you need.
This page compares seven tasks that companies regularly assign to AI, from Excel workbooks to web research, on the basis of current evidence: what ChatGPT does today, where its limits lie, and how open models perform. Two public benchmarks frame the comparison: GDPval-AA, an independent repetition by Artificial Analysis of OpenAI's own benchmark of professional work, and the Artificial Analysis Intelligence Index.
private-local.ai vs ChatGPT: the main differences
| private-local.ai | ChatGPT | |
|---|---|---|
| Who operates the system, and where | Your company operates it, on a machine that it buys and keeps in a location of its choosing; an employee or a supplier maintains it. For the test drive, we operate the machine on your behalf by the hour, so that you can assess the hardware before buying any. | OpenAI operates it on its own servers. Users sign in and use it; nothing needs to be installed or maintained, and no hardware needs to be replaced later. |
| What happens to your data | On a machine that your company owns, prompts, files and results remain on that machine. | On consumer plans, OpenAI uses conversations for training unless the user opts out; on Business, Enterprise and the application programming interface (API), it does not do so by default. Processing inside Europe is available for Enterprise, Edu and the API, whereas Business keeps a copy in the United States for abuse monitoring. In the New York Times copyright case, a court ordered OpenAI in January 2026 to hand over 20 million de-identified chat logs. |
| What can be verified and repeated | The exact machine, model version and serving software are documented, so that a result can be reproduced later on the equipment you would buy. The whole configuration can be inspected. | OpenAI names and documents its models, but the hardware and the serving software belong to OpenAI and can change. Users can assess the answers, but not the configuration that produced them. |
| Which models are available | Open models that you choose and fix to a specific version. We begin with three model families: Qwen, DeepSeek and GLM. Which models fit depends on the memory of the machine you buy, which is why the test drive uses real hardware. | OpenAI's own models, GPT-6 Astra, GPT-6 Sol and GPT-5.6 Sol, which remain on OpenAI's servers. OpenAI also publishes two open models, gpt-oss-120b and gpt-oss-20b, which rate far below its current models. |
| How the costs are structured | A one-off purchase of equipment, plus the cost of operating it: electricity, space, updates and staff time. The costs are capital and operating expenditure rather than a monthly fee per person. The test drive that informs this decision is charged by the hour, and the rate is published on our pricing page. | A subscription billed per person per month, or billing by volume through the API. There is no equipment to buy or operate, and the cost grows with the number of users; current prices are listed below. |
| What is needed to begin | You select a setup from a short list and receive a web address and an access key for it. Self-service booking is under development; until it is available, setups are requested through a form. | A user creates an account and can begin immediately; no further setup is required. |
What the public benchmarks show
GDPval-AA, published by Artificial Analysis, runs the 220 public tasks of OpenAI's GDPval benchmark in a single neutral environment. The tasks are finished pieces of work from 44 occupations, such as workbooks, reports and presentations. Experts compare two results without knowing which model produced which, and the outcome is expressed as a rating on the Elo scale used in chess: a lead of 50 points means that graders prefer the leading model's work in about 57 of 100 comparisons. The Artificial Analysis Intelligence Index combines ten evaluations of reasoning, knowledge, programming and professional tasks into a single score.
| Model | Developer | GDPval-AA rating | Intelligence Index |
|---|---|---|---|
| GLM-5.3 Flash | Z.ai, open weights | 1641 | 42 |
| Qwen3.8-Flash-Next | Alibaba, open weights | 1612 | 40 |
| DeepSeek V4.1 Flash | DeepSeek, open weights | 1600 | 39 |
| GPT-5.6 Sol | OpenAI | 1588 | 47 |
| GPT-6 Astra | OpenAI | 1542 | 53 |
| GPT-6 Sol | OpenAI | 1487 | 48 |
| gpt-oss-120b | OpenAI, open weights | 596 | 12 |
On finished work products, the best open models rate above OpenAI's current models. On the broader index, OpenAI's newest model leads.
Figures read from Artificial Analysis on 28 September 2026, each at the model's highest reasoning setting and with the full-size model running on its provider's servers. Anthropic's Claude Opus 5.5 leads both measures, with 1846 and 58.
Seven everyday tasks compared
For each task, this section states what ChatGPT does today and how well it does it, and what an open model on a machine that you own achieves. Every figure is dated in the sources at the end of the page.
Excel workbooks
ChatGPT leads on demanding spreadsheet work.
Open models on your own machine
An open model with a Python tool writes working formulas and native charts, and LibreOffice recalculates and checks them; pivot tables are not supported. On SpreadsheetBench 2, open models solved 7 to 17 per cent of the tasks, and changing only the tool configuration moved one model between 9 and 16 per cent, which shows that the connected tools matter as much as the model. In our own AI accounting test, the open model GLM-5.3 Flash built a reconciled workbook from 66 invoices and scored 90 of 95 points, against 82 and 72 points for GPT-5.6 Sol in two runs.
ChatGPT today
Since May 2026, ChatGPT for Excel has worked in a side panel inside Excel on every plan and builds workbooks with several sheets, working formulas and charts. OpenAI states that complex formulas may require manual work and that macros may not be fully supported. On SpreadsheetBench 2, a benchmark of 321 real business workbooks, the best proprietary models solved 24 to 35 per cent of the tasks.
PowerPoint presentations
The quality is comparable; ChatGPT is more convenient inside PowerPoint.
Open models on your own machine
On AA-Briefcase, which grades projects lasting several weeks that involve spreadsheets, presentations and memoranda, Qwen3.8-Flash-Next rates 1588, slightly ahead of GPT-6 Astra at 1569 and above GPT-5.6 Sol at 1487. An open model fills the placeholders of your own template and adds editable charts through Python.
ChatGPT today
Since July 2026, ChatGPT for PowerPoint has drafted and revised editable presentations on every plan. OpenAI's help page states that slides may not match a template exactly and that charts and shapes may be limited.
Word documents and contracts
The results are comparable; the best open models rate higher on reports and memoranda.
Open models on your own machine
On GDPval-AA, which includes reports and memoranda, GLM, Qwen and DeepSeek rate above both GPT-5.6 Sol and GPT-6 Astra. A Python tool writes Word files with comments; tracked changes require additional work.
ChatGPT today
ChatGPT for Word was released on 17 September 2026. It drafts, summarises and revises text inside Word and offers reusable instructions for memoranda and contract review. OpenAI states that complex formatting and tables may require manual correction, and its documentation does not mention tracked changes.
The models perform equally well; ChatGPT is more convenient.
Open models on your own machine
The model writes the text, and the company's own email client or automation software reads and sends the messages. On AutomationBench-AA, which simulates business workflows across Gmail, Slack and Salesforce, DeepSeek V4.1 Flash completed 68.9 per cent of the objectives and GPT-6 Astra 68.5 per cent.
ChatGPT today
With Gmail or Outlook connected, ChatGPT drafts and sends email and can start a task when a new message arrives, pausing for approval where required. From shared mailboxes, it sends plain text only and cannot read attachments.
Web research
ChatGPT leads.
Open models on your own machine
The best open models that fit on a single machine score 64 to 68 per cent on BrowseComp. With either approach, every web search passes the query to an external search engine.
ChatGPT today
Deep research plans a search, reads sources and connected applications, and writes a report with citations. OpenAI reports a score of 91.5 per cent for GPT-6 Astra on BrowseComp, a test of finding hard-to-locate facts on the web, and states that search results and citations can be incomplete, outdated or incorrect.
Image generation
ChatGPT leads clearly.
Open models on your own machine
In public comparisons, voters prefer GPT Image 2.5 to the best open image model in about three of four cases. Most strong open image models require a paid licence for commercial use, and on a DGX Spark a high-quality image takes minutes rather than seconds.
ChatGPT today
GPT Image 2.5, released on 8 September 2026, ranks first on both public image leaderboards. It edits parts of an image and combines several reference images. OpenAI's own guide states that its image models still struggle with precise text placement and with consistent brand elements.
Analysing data files
Performance is comparable on simple questions; ChatGPT leads on difficult ones.
Open models on your own machine
In the same study, gpt-oss-120b answered 71 per cent of simple questions and 8 per cent of difficult ones correctly. On a machine that you own, files are analysed where they are stored and are never sent to an external provider.
ChatGPT today
ChatGPT analyses Excel, comma-separated values (CSV), JavaScript Object Notation (JSON), Extensible Markup Language (XML) and Portable Document Format (PDF) files with Python, up to 512 megabytes per file. In the independent DSGym study, GPT-5 answered 75 per cent of simple and 28 per cent of difficult data-analysis questions correctly. Its analysis environment has no internet access and may misread scanned tables.
What ChatGPT costs
ChatGPT is sold per person per month, and the API, which companies use for their own software, is billed according to the volume of text processed. Our test drive costs $10 per hour, and a machine of your own is a one-off purchase plus the cost of operating it.
| Plan | Price | Models |
|---|---|---|
| Free | $0 | GPT-5.6 Luna, with usage limits |
| Go | $8 | GPT-5.6 Luna |
| Plus | $20 | GPT-5.6 Sol; GPT-6 Astra only in the Work agent and in Codex, with usage limits |
| Pro | $100 | GPT-5.6 Sol, and GPT-6 Pro on GPT-6 Astra for up to 50 messages per week |
| Business, standard seat | $25$20 with annual billing | GPT-5.6 Sol, and GPT-6 Pro for up to 15 messages per month |
| Business, premium seat | $125$100 with annual billing | GPT-5.6 Sol, and GPT-6 Pro for up to 50 messages per week |
| Enterprise | Priced individually | GPT-6 Astra, once an administrator enables it |
| Model | Input | Output |
|---|---|---|
| GPT-6 Astra | $10 | $50 |
| GPT-6 Sol | $2 | $10 |
| GPT-6 Luna | $0.10 | $0.50 |
| GPT-5.6 Sol | $4 | $20 |
United States list prices before tax, read on 28 September 2026. Pro also had a $200 tier, which has been closed to new subscribers since 10 September 2026, and GPT-5.6 Sol is offered at a promotional API price until at least 21 November 2026. A token corresponds to roughly three quarters of an English word.
Which option suits which company
ChatGPT is the better choice if
- You need the strongest image generation and web research, and nothing that you use AI for is confidential.
- Nobody in your company is able to maintain a machine. A local setup must be installed, updated and maintained.
- Your use is light or occasional. A subscription is difficult to beat when your own machine would stand idle for most of the day.
Your own hardware is worth considering if
- You handle material that should not be sent to an external provider at all, such as client files, contracts, source code, patient or case records, or anything covered by an agreement about where data may be processed. A machine of your own keeps that material inside the company, and the test drive shows whether such a machine would do the work.
- The same kind of task runs frequently and in large volumes. Steady, repeated work is where owning equipment begins to pay off.
- You need to know exactly what produced an answer and to be able to repeat it later on the same setup.
Sources and notes
The ChatGPT column is based on OpenAI's own documentation and release notes, read on 28 September 2026. The ratings come from Artificial Analysis (GDPval-AA version 2.1, AA-Briefcase version 1.1, AutomationBench-AA, the Intelligence Index version 4.3.2 and its text-to-image leaderboard), read on the same day. The prices come from OpenAI's ChatGPT and API pricing pages, also read on 28 September 2026. Further figures come from SpreadsheetBench 2 (29 June 2026), the DSGym study (22 January 2026) and BrowseComp results published by OpenAI and by the Steel leaderboard (4 September 2026).
Our own measurements from named configurations will be added here, stating the machine, the model version, the task and the date. If any information on this page is out of date, please let us know and we will correct it.
Still to be added
- Dated results from a named machine, model and serving configuration, showing the tasks that failed alongside those that succeeded.
- A repetition of the seven tasks on a DGX Spark with the models we offer, with dated results.
- A note on the machine that produced our figures, so that it can be compared with the equipment you are considering.
Further reading: the other four comparisons cover Palantir, the open-source engine LocalAI, the redaction company Private AI and the on-premises assistant Seesa, and the guide how to run AI locally explains the technical route.
Questions about private-local.ai and ChatGPT
private-local.ai vs ChatGPT: which one should we choose?
That depends on the work. Your own setup is worthwhile when the same kind of task recurs in large volumes, when the material is sensitive, or when you need to know and repeat exactly what produced a result, and on finished work products the best open models now rate above OpenAI's current models. ChatGPT remains ahead on image generation and web research. The decision is made per task rather than once for the whole company, and the test drive allows you to make it with your own workload.
Can we run ChatGPT's models on our own hardware?
Not the models that ChatGPT uses: GPT-6 Astra, GPT-6 Sol and GPT-5.6 Sol remain on OpenAI's servers. OpenAI does publish two open models, and gpt-oss-120b runs on a DGX Spark, but it rates far below current models, with 596 on GDPval-AA against 1542 for GPT-6 Astra. Running AI on your own equipment therefore means using open models such as GLM, Qwen or DeepSeek, and the decisive question is whether one of them performs your particular task well enough. A test drive answers exactly that question.
What does a test drive cost, and how do we book one?
$10 per hour of endpoint access, and most questions are answered within one or two hours. Self-service booking is under development; until it is available, you request the setup you have chosen through the contact form.
Which ChatGPT models does this comparison cover?
GPT-6 Astra, released on 3 September 2026, is OpenAI's most capable model; GPT-5.6 Sol is the model that Plus subscribers use in ChatGPT; and GPT-6 Sol followed on 22 September 2026 for ChatGPT's Work agent and for Codex. Every rating on this page names the model to which it refers.
Find out which option suits your work
Choose a setup, connect the tool you already use and run the task that matters to you. Within one or two hours, you will know whether a machine of your own can do the work, before you buy one.
Request an endpoint