Published: November 5, 2026.
If you want a free assistant to explain a piece of code, work through a math problem, or read a screenshot, DeepSeek is worth trying before you pay for ChatGPT. If you want one app that can also conduct a practice interview by voice, generate an illustration, and calculate totals from a spreadsheet, start with ChatGPT Free. You can try those tools before deciding whether their usage limits justify a subscription.
The DeepSeek vs ChatGPT comparison changes when you move beyond the chat window. A writer may care most about keeping a style guide across assignments. A developer may care about the cost of processing 10,000 requests. A company handling confidential files may need to control where the model runs. Those are three different buying decisions, even when the same two names appear on the shortlist.
DeepSeek and ChatGPT at a glance
| What you need | DeepSeek | ChatGPT |
|---|---|---|
| Everyday chat | Free web and mobile chat with reasoning, web search, and file uploads. | Free web and mobile chat with search. Uploads, voice, image creation, and data analysis have separate free allowances. |
| Help with a visual | Upload a photo, diagram, or screenshot and ask questions about it. | Ask about an existing visual, generate a new image, or request an edit. |
| A research report | Use web search to find sources and ask follow-up questions. | Use search for a quick answer or deep research for a cited report. You can restrict the research to selected websites. |
| Work across a codebase | DeepSeek Harness can edit local files and run tests. Its default model service requires a prepaid balance. | Codex can work with a repository and run tests. Included usage depends on your ChatGPT plan. |
| A model you can host | DeepSeek publishes downloadable weights for the V4.1 Flash model used in its official chat. | OpenAI offers separate downloadable gpt-oss models. ChatGPT itself is a hosted service. |
Which models are you actually comparing? DeepSeek updated its official web and mobile apps to DeepSeek-V4.1-Flash-1017 on October 17, 2026. In the main ChatGPT chat, OpenAI identifies GPT-6 Luna for Free and Go, GPT-6 Sol for Plus and Pro, and GPT-6 Astra for the Pro reasoning option. Work and Codex have their own model choices. A result from a paid coding agent tells you little about what to expect when you paste the same question into a free chat. DeepSeek model updates and ChatGPT release notes.
What you get for free and when a paid plan is worth it
DeepSeek’s official assistant is free on the web, iPhone, and Android. You can upload a document, ask it to extract information, and use web search without buying API credits. ChatGPT Free includes unlimited everyday text chat subject to abuse safeguards. Its tool limits work independently, so you can reach your image-generation allowance and still carry on a text conversation. DeepSeek app features and ChatGPT Free allowances.
As of November 6, 2026, the US monthly price for ChatGPT Go is $8. It makes sense if you like the free experience but regularly need more uploads, images, or voice time. Plus is $20 and adds advanced reasoning and expanded deep research and Codex usage. Pro has $100, $200, and $500 monthly tiers for people who need Pro reasoning or substantially more capacity. For occasional writing or homework explanations, try a representative task on both free services first. For a week spent analyzing reports, the limit you actually hit is a better guide to what to buy. US plan prices, ChatGPT feature comparison, and Pro tier details.
Writing and research reward different features
For a single email or rewrite, either free chatbot is a reasonable choice. One revealing test is to give both the same two facts: a shipment is expected Friday, and the carrier has yet to confirm pickup. Ask for a two-sentence customer update. A polished reply that promises Friday delivery has changed the meaning. The better draft is the one you can send with the facts intact. For a newsletter or other recurring assignment, ChatGPT has a practical advantage. A Project can keep your style guide, examples, instructions, and related chats together for the next issue. How ChatGPT Projects work.
For a current factual question, both services offer web search. Suppose you want to know whether a manufacturer’s US warranty covers a replacement battery. Give the assistant the exact model number and ask for the warranty document and the clause that answers the question. Open the document and check the model and region yourself. A citation to the right manufacturer can still point to the wrong product. DeepSeek web search and ChatGPT search and citations.
Choose ChatGPT deep research when you need to compare several sources in one report. For example, you can limit a warranty comparison to three manufacturers’ websites, review the proposed research plan, and get a report with links. DeepSeek search is a good fit when you have one policy to locate and would rather guide the follow-up questions yourself. ChatGPT deep research and source selection.
Reading a PDF, checking a spreadsheet, and making an image
For a screenshot or diagram, either assistant can help. DeepSeek’s V4.1 Flash has native visual understanding. If you upload a chart, ask it to identify the axis labels, units, and two plotted values before explaining the trend. That quick check can reveal a misread scale. If you want the assistant to create an illustration or edit a photograph, ChatGPT offers those actions inside the conversation. DeepSeek’s Janus-Pro can generate images, but it is a separate model from the Flash model in the consumer app. DeepSeek Flash model card, Janus-Pro model card, and image creation in ChatGPT.
For a PDF with important charts, how you upload it matters. DeepSeek supports file uploads and text extraction. If the answer depends on a figure, attaching a clear image of that figure lets you ask about it directly. In ChatGPT, attaching a PDF directly to a chat supports questions about its text and images. A PDF saved as a Project file or as knowledge for a custom GPT can be handled differently, and embedded visuals may be skipped. Attach the PDF to the conversation or upload the relevant page if the chart is what matters. DeepSeek file uploads and ChatGPT’s PDF and image handling.
For a spreadsheet you need to reconcile, choose ChatGPT. Its data analysis tool can run Python against an uploaded CSV or workbook and show the calculation behind a result. If you ask for monthly sales by store, have it report the rows it excluded and reconcile the store subtotals to the original total. A missing store is much easier to catch in that check than in a finished bar chart. For a spoken task, such as rehearsing a job interview and answering follow-up questions aloud, ChatGPT Voice is also a straightforward choice. ChatGPT data analysis and ChatGPT Voice.
Which is better for coding, reasoning, and math?
For a self-contained bug with a reproducible input, DeepSeek is a good place to start. Paste the function, the failing input, and the expected output. For example, if a CSV importer turns the product ID 00127 into 127, ask for the smallest fix and a test that preserves the leading zeros. You can judge either assistant’s answer by running that test and checking that ordinary numeric fields still import correctly.
There is independent evidence that the choice of task can reverse the result. In Artificial Analysis’s comparison available November 6, 2026, DeepSeek V4.1 Flash 1017 scored 34% and GPT-6 Luna scored 13% on Terminal-Bench 4.0, a set of 66 terminal tasks scored by running tests. On the hard subset of CritPt v2, which tests research-level physics reasoning, Luna scored 43% and Flash scored 25%. Both models ran at their maximum reasoning setting. Those results help you choose a model to try for a similar task. They do not measure the default chat experience, ChatGPT’s paid models, or how well an explanation will teach you algebra. Side-by-side model results and benchmark methodology.
For learning math, use a problem where you can check the reasoning. If an $80 item needs a 20% gross margin, the price is $100 because the $20 profit is 20% of the sale price. An answer of $96 applies a 20% markup to cost. Ask each assistant to explain why the denominator changes and give you a second problem to solve. The explanation that helps you catch the distinction yourself is the more useful tutoring answer.
For a bug spread across several files, compare the coding tools directly. DeepSeek Harness is a separate desktop application in public preview for Windows and Apple silicon Macs. It can inspect a repository, edit files, and run tests. Using its default hosted model requires prepayment on the DeepSeek Open Platform. OpenAI Codex can also work with repositories and execute tests, with limited access on ChatGPT Free and higher allowances on paid plans. If you already subscribe to ChatGPT, try the included Codex allowance on an actual issue before funding another service. DeepSeek Harness, Harness billing terms, and Codex plan access.
Privacy depends on where you send the conversation
DeepSeek’s hosted service. DeepSeek says it collects, processes, and stores personal data in the People’s Republic of China. That data can include prompts, uploaded files, photos, and chat history. Retention depends on the purpose and type of data. Its terms describe an Improve the model for everyone setting you can turn off to opt out of the stated use of de-identified inputs and outputs for improvement. Turning it off and deleting a conversation are separate actions. DeepSeek privacy policy and DeepSeek terms, section 4.3.
Personal ChatGPT accounts. OpenAI lets you turn off Improve the model for everyone under Settings → Data controls. New chats can then remain in your history while being excluded from training. A Temporary Chat stays outside your history and is excluded from model improvement while it remains temporary. OpenAI may retain a copy for up to 30 days for safety. Its US privacy policy describes processing in multiple jurisdictions, including the United States. ChatGPT data controls, Temporary Chat retention, and OpenAI’s US privacy policy.
Work documents. If you need to analyze a client contract, first check which service and account your organization has approved. OpenAI excludes ChatGPT Business, Enterprise, and API inputs and outputs from training by default. For US federal civilian agencies, a separate procurement and IT rule matters. An October 2026 Department of Homeland Security order covers DeepSeek products and also requires removal of third-party products and services developed using DeepSeek implementations from civilian agency IT. OpenAI business data commitments and the federal order and its scope.
Can you run DeepSeek or ChatGPT locally?
DeepSeek publishes the V4.1-Flash-1017 model weights under the MIT license. That gives a team the option to run inference on hardware it controls. The official repository totals about 510 GB, so the full release is a substantial deployment. That figure is the repository’s storage size. Working memory depends on the model format, runtime, and context length. For a personal experiment, a smaller release such as DeepSeek-R1-Distill-Qwen-1.5B is a more practical starting point, with different capabilities from the hosted Flash model. Flash license and deployment notes, published model files, and the smaller DeepSeek model.
OpenAI also publishes downloadable models, including gpt-oss-20b, under Apache 2.0. These are separate releases from the GPT-6 models served in ChatGPT. If your goal is to keep documents on your own machine, check where the software sends each request. A desktop app can read a local file and still send its contents to a hosted model. For fully local processing, the model, document processing, and any tools that receive the content all need to run within your chosen environment. OpenAI’s guide to its open-weight models.
For developers, which API costs less?
For developers evaluating ChatGPT vs DeepSeek, the relevant prices are the rates for the specific OpenAI and DeepSeek API models your application will call. DeepSeek’s free chat and ChatGPT Plus are consumer offerings. Plus bills API usage separately. Eligible ChatGPT Pro and Business subscribers can activate promotional API credits for a linked API organization, and unused credits expire according to the billing cycle. ChatGPT Plus billing and API credit eligibility and terms.
Consider a support app making 1,000 short requests, each using 1,000 uncached input tokens and 100 billable output tokens. That totals 1 million input tokens and 100,000 output tokens. At the US dollar rates published on November 6, 2026, the model charges would be:
| Model and billing rate | Uncached input per 1M tokens | Output per 1M tokens | Example total |
|---|---|---|---|
| OpenAI GPT-6 Luna, standard | $0.10 | $0.50 | $0.15 |
| DeepSeek Flash, off-peak | $0.15 | $0.60 | $0.21 |
| DeepSeek Flash, peak | $0.30 | $1.20 | $0.42 |
| OpenAI GPT-6.1 Sol, standard | $2.00 | $10.00 | $3.00 |
GPT-6 Luna costs least for this particular token mix. A different workload can change the order. DeepSeek offers lower rates for cached input, and both the length of the answer and the number of retries affect the bill. The example assumes short-context OpenAI requests, no cache hits, no paid tools, and no credits. For a production decision, run a sample of real requests and divide the total cost by the number of answers that pass your quality checks. DeepSeek API rates and peak hours and OpenAI API rates by model and tier.
Also check whether the model accepts your input. DeepSeek lists image input for deepseek-flash and text input for deepseek-v4-pro. Its API documentation lists a 1-million-token context window and up to 393,216 output tokens for those models. Those figures describe developer requests. Use the consumer product’s own upload and usage limits when deciding whether it can handle a particular document in chat. DeepSeek’s model specifications.
Should you switch from ChatGPT to DeepSeek?
Look at the last few tasks you actually used ChatGPT for. If they were explaining an error, revising a message, and answering a question about a screenshot, try the hardest one in DeepSeek’s free app. If its answer needs no more correction and you rarely use paid tools, you have a concrete reason to reconsider a subscription.
If those tasks involved talking through interview questions, returning to a Project full of reference material, or checking the calculation behind a chart, keep those workflows in ChatGPT and see whether Free covers your usage. Using both also makes sense for an occasional second opinion. For a disputed answer, settle the question with the original document, a calculation, or a passing test. Agreement between two chatbots is a poor substitute for any of the three.