DeepSeek vs ChatGPT in 2026: Where the Cheap One Wins, and the One Thing No Page Documents

Disclosure: NOXI publishes this page. NOXI is a presentation maker, not a chat assistant, so it sits outside the DeepSeek vs ChatGPT head-to-head and appears once, as a reference row in the slides section, scored on the same seven presentation rules as the two assistants. That is why the rules, the weights and the source of every cell are printed, along with the rule where ChatGPT finishes ahead of NOXI.

Most DeepSeek vs ChatGPT comparisons still quote API prices from a tariff that changed, and none of the ones we read this morning mentions what DeepSeek did at midnight Beijing time. This page reads both vendors' own pages on one day, signed out, and scores them on the same seven rules as our ChatGPT vs Claude and ChatGPT vs Gemini pages. The short version: the cheap one really is cheap, and the gap is almost entirely about tools and files.

DeepSeek vs ChatGPT at a glance (2026)

ChatGPT leads DeepSeek by 42 points as an assistant, 90 to 48, and by 56 points as a slide maker, 90 to 34, while the two tie on the entry price and on add-ons for individuals.

In the slides section, NOXI, a dedicated deck maker you can use inside ChatGPT and as a custom connector in Claude, is a reference row at 98 on the same seven presentation rules, outside this head-to-head. ChatGPT finishes ahead of NOXI on where a tool runs, 5 against 4. Prices are US list prices before tax, as each vendor's page served them on 1 November 2026.

Who should pick what

The honest split is by job, not by total: DeepSeek is a writing and reasoning surface with a price that is hard to argue with, and ChatGPT is the one that hands you a file. Each line rests on one fact from a vendor page opened on 1 November 2026.

• If the cost has to be zero and the output is text, pick DeepSeek, because its download page says "Chat with leading AI models for free." and no DeepSeek page sells a paid consumer tier at all [13].
• If you are paying per token in your own code, pick DeepSeek, because its pricing page lists off-peak output at $0.6 per 1M tokens on flash and $1.98 on v4-pro, with cache hits at $0.003 and $0.022 [15].
• If you want the most for $0 in a chat window, pick ChatGPT, because its signed-out plan table lists search, file uploads, image generation, plugins and ChatGPT Work on Free, five of the five capabilities this rule counts [2].
• If the assistant has to produce a document, spreadsheet or deck, pick ChatGPT, because ChatGPT Work is documented for creating files and the PowerPoint add-in is "available globally to Free, Go, Plus, Pro, Business, Enterprise" accounts [6].
• If your chats must stay out of training without writing to anybody, pick ChatGPT, because "Improve the model for everyone" can be switched off and Temporary chats "are not used to improve OpenAI models"; DeepSeek's opt-out runs "by emailing privacy@deepseek.com" [5][18].
• If the deliverable is a designed deck, and especially an Arabic one, use a dedicated deck maker such as NOXI, because its reference row takes full marks for right-to-left slide layout, where no official DeepSeek page names Arabic at all [22].

Key numbers

The two scores count documented capability on printed rules, not answer quality. The two prices count different things: $0.75 buys a thousand successful search or fetch requests on DeepSeek's API, $20 buys a month of a consumer chat plan.

DeepSeek vs ChatGPT side by side

DeepSeek and ChatGPT agree on more than the totals suggest, including a free entry point, web search and file uploads; the documented differences are tools, file output, connectors and the published context window.

How we scored DeepSeek and ChatGPT

Both assistants are scored on documented capability from pages opened signed out on 1 November 2026, using the same seven rules as our ChatGPT vs Claude and ChatGPT vs Gemini pages. It is not a new hands-on test, and no rule scores answer quality.

Changed recently: DeepSeek started charging for Search and Fetch at 00:00 Beijing time on 1 November 2026, two weeks after its in-house search API opened on 17 October 2026; OpenAI rewrote and renamed its custom-MCP article five days ago, refreshed the Free-tier FAQ fourteen hours ago and the data-controls FAQ two days ago, and still describes ChatGPT Slides as "rolling out to eligible accounts" on Pro, Business and Enterprise.

One cell moved since our October assistant pages, where DeepSeek scored 39. DeepSeek now documents connecting any custom MCP server through its own published client in Harness, so A5 rises from 1 to 4 and the total from 39 to 48. The anchor is 4 and not 5 because the route runs through Harness, which needs a paid DeepSeek API key, and the reference configurations ship switched off [19].

The scored table: ChatGPT 90, DeepSeek 48

ChatGPT scores 90 and DeepSeek 48 out of 100, a gap of 42 points that comes from five rules: tools on the $20 plan (+16), the free plan (+12), training control (+6), surfaces (+6) and the published context window (+2).

The arithmetic, row by row

Each weighted score is the rating divided by 5, times the criterion's weight, and each total is the sum of the seven weighted scores.

Is DeepSeek better than ChatGPT?

Not on these rules: ChatGPT leads on five of seven and the two tie on the rest, so DeepSeek leads on none. That is a narrower verdict than it sounds. Every rule DeepSeek loses is about tools, files, connectors or settings, and none is about the writing or the reasoning, which no vendor publishes in a form these rules can read. If you mostly type a question and read the answer, the 42 points buy you less than the arithmetic suggests. For the wider field, the same seven rules put Google Gemini at 94 and Microsoft Copilot at 77 in our Copilot vs Gemini comparison, and ChatGPT at 90 against Copilot in ChatGPT vs Copilot.

What would change the result

• DeepSeek 39 if the Harness documentation site is not accepted as an official DeepSeek page, or if custom MCP behind a paid API key is not read as an add-on for individuals (A5 1).
• DeepSeek 51 if the generic MCP row is read as the same free-route custom MCP that scores 5 elsewhere on this rubric (A5 5).
• DeepSeek 50 if Harness's command-line tool and Python SDK count as DeepSeek's own coding agent in addition to the desktop app (A7 3).
• DeepSeek 56 if Harness's document creation counts on the free plan despite the API key (A1 3). We did not take this reading, because the quickstart requires the key.
• ChatGPT 86 if free image generation is not counted, or if ChatGPT Work file creation on Free is not counted (A1 4); 82 if neither is counted.

At the closest pairing, ChatGPT 82 against DeepSeek 56, the order is unchanged and the gap is still 26 points.

Where DeepSeek wins: price, and what that buys

DeepSeek takes full marks on the entry-price rule because no DeepSeek page sells a paid consumer tier, and its API prices are the lowest here by an order of magnitude: off-peak output runs at $0.6 per 1M tokens on flash and $1.98 on v4-pro.

DeepSeek publishes two models, "deepseek-flash" and "deepseek-v4-pro", priced "in units of per 1M tokens". Off-peak is half the peak rate, and peak hours are 01:00 to 04:00 and 06:00 to 10:00 UTC, Monday through Friday, excluding Chinese public holidays. Context length is listed as 1M and maximum output as 384K [15].

What this does not buy is the rest of the rubric. The free chat app documents web search and little else, so the tools rule scores 1. The price is a reason to keep DeepSeek in your stack rather than to make it the only thing in it.

What changed today: DeepSeek started billing Search and Fetch

From 00:00 Beijing time on 1 November 2026, DeepSeek charges $0.75 per 1,000 successful requests for Search and $0.75 per 1,000 for Fetch, and the same page confirms that the server-side tool calls those endpoints serve were free until that moment.

The sentence to read carefully sits on DeepSeek's API pricing page: when you use the DeepSeek API through DeepSeek Harness, Claude Code or Codex, server-side tool calls already hit the Search and Fetch APIs, and those calls were "currently free of charge; per-request billing starts at 00:00 Beijing time on November 1, 2026" [15]. Two details matter. First, "currently free of charge" covers server-side tool calls through those three clients, not every search call: the standalone Search and Fetch APIs are described on the same page as billed per request. Second, 00:00 Beijing time is 16:00 UTC on 31 October, so the change is already in force.

The context is two weeks old. DeepSeek's changelog entry of 17 October 2026 says its web and API search capabilities are powered by its own in-house search engine, an AI-native general-purpose engine that filters SEO spam and low-quality machine-generated pages [16]. The search engine arrived free, and today it has a price list.

For an individual the amounts are small: a thousand searches cost less than a coffee. For anyone running an agent in a loop, a free call just became a metered one, which is worth a line in the cost model.

Is DeepSeek free, and on what

The DeepSeek chat app is free, in one sentence on its own download page, and the paid part is the API: the app carries no price and no documented usage limit, while the API is billed per token and now per search request.

DeepSeek's download page says "Chat with leading AI models for free." and offers the app for iOS and Android alongside the Harness desktop app [13]. That is the whole public statement: no consumer pricing page, no tier list, no usage cap. It is why the entry-price rule gives DeepSeek full marks and why the free-plan rule still gives it 2. A free app with two documented capabilities is not a free plan with five.

Three things that are not free, and are easy to assume otherwise:

• Harness. The quickstart in its own documentation says to open Settings, then Models, enter a DeepSeek API key and save it. API usage is billed per token, so the document and slide work Harness advertises runs on a paid key [23].
• Search and Fetch through an agent. Free until this morning through Harness, Claude Code and Codex, now $0.75 per 1,000 successful requests [15].
• The API itself. Both models are billed per million tokens, with peak rates at double the off-peak ones [15].

ChatGPT's free plan is wider on paper and also metered: search "Yes", uploads and image generation "Limited", plugins "Yes", ChatGPT Work "Limited (desktop app)" [2]. Go costs $8 a month, Plus $20, Pro from $100 [1]. Our best AI for PPT ranking scores ten tools on what a free plan actually hands you at the end.

Can DeepSeek make a PowerPoint? What its own pages document

On the seven presentation rules, ChatGPT scores 90 and DeepSeek 34, and the precise finding is narrower than the usual claim: DeepSeek's own Harness page documents drafting docs and creating slides, and no page on DeepSeek's site documents a PowerPoint (.pptx) file in any of them.

We re-applied the rules of our AI presentation tools ranking on 1 November 2026. Points are the rating divided by 5, times the weight: P1 editable PowerPoint export (20), P2 editing depth (20), P3 research with the source on the slide (15), P4 import of PDF, Word, Excel or CSV, PowerPoint and images with a stated limit (15), P5 languages including documented right-to-left layout (10), P6 free plan reality (10), P7 where it runs, web plus inside PowerPoint or Google Slides plus an API, MCP or connectors (10).

The gaps. ChatGPT minus DeepSeek = 56, made of P1 +16, P2 +16, P4 +12, P6 +6, P5 +4 and P7 +2. They tie on P3, research with sources, where both document research and neither documents a source line on the slide itself. NOXI minus ChatGPT = 8, made of P3 +6, P5 +2 and P6 +2 against P7 minus 2, with P1, P2 and P4 level: ChatGPT finishes ahead on where it runs, because its add-in works inside PowerPoint, while NOXI runs on the web, as a plugin in ChatGPT and as a custom connector in Claude.

What DeepSeek's pages actually say about slides

The page to read is deepseek.com/en/harness/, which calls Harness a public preview worldwide and open source and lists "Organize files, analyze data, draft docs, and create slides" in the assistant's own description of its everyday work [14]. The same page names the files Harness writes: "Brief.docx", "Data.xlsx", "index.html", "label.ts", "analyze.py", "Report.pdf" and a README in Markdown. There is no PowerPoint icon and no .pptx on it. We checked the rest of the surface the same way, including the home and download pages, the API docs, the changelog and the full sitemap. The Files API is about images: "Supported formats: JPEG, PNG, GIF, and WebP" [17].

So the accurate claim is that no DeepSeek page documents a PowerPoint file, not that DeepSeek says nothing about slides. P1 stays at 1 because the rule asks for a documented deck file and "create slides" names no format. P2 is 1 because Harness offers to refine documents, spreadsheets and code through conversation, with no slide editing documented. P6 is 1 because the one place slides are mentioned needs a paid API key. P7 is 4, level with NOXI's cell. Sensitivity: DeepSeek 38 if Harness's PDF output is read as a deck route (P1 2), and 38 if its slide creation is read as free deck generation despite the API key (P6 3).

How ChatGPT gets to 90, and what NOXI's reference row rests on

• ChatGPT. The PowerPoint add-in is "a PowerPoint-native AI experience that lives in a sidebar inside Microsoft PowerPoint" and edits "while preserving editable slide structure", from the Free plan with limited usage [6]. ChatGPT Slides adds element-level edits and is "available on ChatGPT Pro, Business, and Enterprise plans with Space access" while still "rolling out to eligible accounts" [7]. "Documents and presentations have a maximum size of 512 MB per file" [8]. Arabic appears in OpenAI's language list with no right-to-left statement, which is why P5 is 4 [9].
• NOXI (reference). NOXI builds a designed deck from a topic, brief or outline. Decks started from a request research the web first and carry a source line on slides with figures. You can change the whole deck, one slide or one element you click by chat, or upload a file as the style reference. It imports PDF, Word, Excel or CSV, PowerPoint and images up to 200 MB, builds right-to-left Arabic slides with four Arabic fonts, and exports an editable PowerPoint (.pptx) with real text boxes and embedded fonts, or a PDF, with no watermark. A free account gets 1,000 credits a month and a deck uses about 25 [22].

Two alternative readings of the reference row, printed by the same rule as the others: NOXI 89 when the deck starts from an attached file, because research runs on prompt-started decks and an AI check decides with research off by default; and NOXI 94 when the ChatGPT plugin and the Claude connector are set aside and only the web app counts [22].

NOXI is also listed in ChatGPT's plugin directory: open Plugins in the sidebar, search "NOXI", choose Install plugin, then Connect and sign in, and type @NOXI in any chat. OpenAI says that installing or using a plugin "depends on your plan, workspace, role, region" [10][21]. Our plugin launch page walks through the steps. For the shortest path from a chat answer to a file, which AI can make a PowerPoint compares the routes.

DeepSeek vs ChatGPT privacy: who trains on your chats

ChatGPT scores 4 of 5 on training control and DeepSeek 2 of 5, because OpenAI documents a switch and a chat mode excluded from training while DeepSeek's opt-out is exercised by sending an email.

OpenAI's data-controls article tells you to "Turn off Improve the model for everyone" in settings, and says Temporary chats "are not used to improve OpenAI models" [5]. That is an opt-out setting plus a training-free mode, which is anchor 4. Anchor 5 would need the vendor to state that consumer chats are not used for training at all, which OpenAI does not.

DeepSeek's privacy policy grants "the right to opt-out of using your Personal Data for training our models", exercised "by emailing privacy@deepseek.com" [18]. No setting is named and no training-free chat mode is documented, which is anchor 2. The same policy lists what the service collects, including "text input, voice input, prompt, uploaded files", and says it processes and stores personal data in the People's Republic of China. Where a vendor stores data is not one of the seven rules, and for many readers it is the first question.

Neither reading is a security assessment. Whichever you pick, the rule that matters is your own: keep identifiable client material out of any consumer AI account. Our page on ChatGPT and Microsoft Copilot sets out the same rule against a vendor that scores 5 here.

Languages, and what no DeepSeek page says about Arabic

ChatGPT scores 4 on the language rule and DeepSeek 2, because OpenAI publishes a supported-language list that includes Arabic while the only multilingual statement on any DeepSeek page is that its search "supports multilingual search".

We looked for the word "Arabic" on eight official DeepSeek pages on 1 November 2026: the English home page, the download page, the Harness page, the API docs index, the pricing page, the changelog, the search guide and the Files API guide, plus the full documentation sitemap and the privacy policy. It does not appear on any of them, and the policy's only use of the word "language" is the device's system language [12][13][14][15][16][17][18].

That is an absence, and it should be read as one. It says nothing about how well DeepSeek writes Arabic, only that DeepSeek makes you no commitment about it, and the language rule scores commitments. OpenAI's list names Arabic but states nothing about right-to-left layout, which is why ChatGPT's cell is 4 rather than 5 [9].

For a deck the distinction gets sharper, because right-to-left layout means more than right-aligned text: the reading order of a timeline, the direction a process chevron points, the column order of a table. NOXI's reference cell takes full marks because its slides run right to left, with timelines, Gantt charts and chevron processes mirrored, direction-aware tables and four Arabic fonts [22].

What these rules don't measure

Seven documented rules leave out most of what people argue about when they compare DeepSeek and ChatGPT, and several omissions favour DeepSeek.

• Price, which DeepSeek wins outright, counts only once. The entry-price rule is worth 10 points of 100. DeepSeek's API prices are the lowest on this page by a wide margin and the chat app is free, and the rubric records all of that as a single 5.
• Open weights and self-hosting. Harness is open source under the MIT licence, with a public repository, a plugin architecture and a Python SDK. None of that is a rule, and for some teams it is the whole decision [14].
• Reasoning, coding and benchmark performance, and speed. Neither vendor publishes them in a comparable form, so they are not scored. Give both the same brief and judge for yourself.
• DeepSeek's own search engine. Its changelog describes an index of hundreds of billions of web pages that filters SEO spam. The rule only asks whether web search exists [16].
• ChatGPT's ecosystem depth. Surfaces are counted, but not what sits on them: a plugin directory, Codex, office add-ins, Spaces and Skills.
• Where each company stores data, and what your regulator says about it. DeepSeek's policy names the People's Republic of China; restrictions in your own country are outside this rubric [18].

For a comparison of the tools that connect an assistant to the rest of your work, our best MCP servers ranking scores twelve of them on a rubric built for that job.

Worked example: a startup analyst with a Friday deadline

For a fictional analyst who drafts on DeepSeek and needs a 15-slide investor update by Friday, the drafting costs about four cents on the API or nothing in the app, and the deck file comes from a tool whose own pages document one.

His session: about 120,000 input tokens, counted as cache misses because every prompt differs, about 30,000 output tokens, and about 40 searches while the agent checks market figures.

The arithmetic. Drafting on flash off-peak comes to $0.036, which rounds to four cents, and the same session on v4-pro is $0.1386, or 3.85 times as much. The 40 search calls cost 0.04 × $0.75 = $0.03 today and nothing yesterday, and the deck uses 25 ÷ 1,000 = 2.5% of a month's credits. His whole Friday, on the API at off-peak rates, comes to $0.066.

Why this split. Rahul keeps DeepSeek for the writing, because that is what he actually uses it for and the price rule is real. He takes the text somewhere that documents a deck file, because his partners have to open the attachment. A deck asked for in a request researches the web first and carries a source line on slides with figures.

Decision: draft in DeepSeek, build the file elsewhere, pay nothing on either side this week. Review on 1 February 2027: if the agent's search calls pass 10,000 a month, which is $7.50 at today's rate, move the research into a plan that bundles it.

Build it: a DeepSeek draft into a deck, through Claude's connector

To turn a DeepSeek draft into a file, add NOXI to Claude as a custom connector with the address on noxi.ai/connect, sign in once, then paste your draft into the request and ask for the deck; open the result in NOXI to edit, present or download it. One flow, six steps:

DeepSeek vs ChatGPT: FAQ

Is DeepSeek better than ChatGPT?

Not on documented capability. On seven printed rules checked on 1 November 2026, ChatGPT scores 90 and DeepSeek 48. ChatGPT leads on five rules and the two tie on entry price and add-ons for individuals. The rules do not score writing or reasoning quality, which neither vendor publishes in a comparable form.

Is DeepSeek free?

The chat app is. DeepSeek's download page says "Chat with leading AI models for free." and no DeepSeek page publishes a paid consumer tier, a usage cap or a pricing table for the app. The API is billed per million tokens, and Harness needs a paid API key, so its slide work is not free. Checked 1 November 2026.

Is DeepSeek cheaper than ChatGPT?

Yes on both counts we can check. The DeepSeek chat app costs nothing against ChatGPT Go at $8 a month or Plus at $20. On the API, DeepSeek lists off-peak output at $0.6 per 1M tokens on flash and $1.98 on v4-pro, with cache hits at $0.003 and $0.022. Prices read on 1 November 2026.

Can DeepSeek make a PowerPoint?

No page on DeepSeek's own site documents a PowerPoint (.pptx) file. Its Harness page documents drafting docs and creating slides, and names the files it writes: .docx, .xlsx, .pdf, .html, .md, .py and .ts. We checked the home, download and Harness pages, the full API docs sitemap and the Files API on 1 November 2026.

Can DeepSeek make slides from my text?

DeepSeek Harness lists "create slides" among its everyday work, but names no slide file format and documents no slide editing, and it needs a paid DeepSeek API key to run. On the seven presentation rules DeepSeek scores 34 against ChatGPT's 90. The practical route is to draft in DeepSeek and build the file in a dedicated deck maker.

Which is better for coding, DeepSeek or ChatGPT?

Neither vendor publishes coding benchmarks these rules can compare, so this page does not score it. What is documented: ChatGPT ships Codex as its own coding agent, and DeepSeek ships Harness, which is open source under MIT with a command-line tool and a Python SDK, and which runs on a paid DeepSeek API key. Checked 1 November 2026.

Which is better for research, DeepSeek or ChatGPT?

ChatGPT, on documented tools. Its $20 plan lists deep research alongside web search, while DeepSeek documents web search with no named deep-research feature, which is why the tools rule scores 5 against 1. DeepSeek's search runs on its own in-house engine, announced on 17 October 2026, and its Search calls now cost $0.75 per 1,000 requests.

Does DeepSeek support Arabic?

No official DeepSeek page names Arabic. We checked eight of them on 1 November 2026, plus the documentation sitemap and the privacy policy, and the word does not appear; the only multilingual statement is that DeepSeek's search "supports multilingual search". That is an absence of a commitment, not a measurement of quality. OpenAI's language list does name Arabic.

What changed in DeepSeek's prices on 1 November 2026?

DeepSeek started billing its Search and Fetch endpoints at $0.75 per 1,000 successful requests each. Its API pricing page says "per-request billing starts at 00:00 Beijing time on November 1, 2026" and confirms that the server-side tool calls made through Harness, Claude Code or Codex were free of charge until then. Token prices were unchanged.

How do I turn DeepSeek's text into a deck?

Copy the draft, then use a tool whose own pages document a deck file. In Claude, add NOXI as a custom connector under Customize then Connectors, using the address on noxi.ai/connect, sign in once, paste the draft into your request and ask for the deck. Open the result in NOXI to edit, present or download it.

Sources & methodology

Every price, plan feature, file type and setting here comes from DeepSeek's or OpenAI's own pages, opened signed out on 1 November 2026. Scores are documented capability on printed rules, not a hands-on test. Rahul Menon is fictional, and every calculation in his example was recomputed twice. NOXI's reference cells come from its product documentation, its ChatGPT plugin listing and the live noxi.ai/connect page.

• [1] Free $0, Go $8, Plus $20, Pro "From $100 / month" with "Plans at $100, $200, or $500 USD per month", Business $20 per user annual and $25 monthly; "ChatGPT Work and Codex are included in your ChatGPT Free, Go, Plus, Pro, Business, Edu, or Enterprise plan" OpenAI, ChatGPT pricing (checked 1 November 2026).
• [2] Plan table read signed out: Free "Search, Yes", "File uploads, Limited", "Image generation, Limited", "Discover & use Plugins, Yes", "ChatGPT Work, Limited (desktop app)"; Plus "GPT Reasoning total context window, 256K"; "Excel, Word, PowerPoint, and Google Sheets extensions" OpenAI, ChatGPT plans (checked 1 November 2026; the page served no price figures to a signed-out fetch, so no price on this page comes from it).
• [3] "Search the web and obtain up-to-date information"; "Upload images or files in your prompts"; "Create images in ChatGPT" OpenAI Help Center, ChatGPT Free Tier FAQ (checked 1 November 2026).
• [4] "Custom MCP server connections and full MCP support are available for ChatGPT Business and Enterprise/Edu customers"; FAQ: "Pro users can connect custom MCP servers with read/fetch permissions." OpenAI Help Center, Connecting custom MCP servers in ChatGPT (checked 1 November 2026).
• [5] "Turn off Improve the model for everyone"; Temporary chats "are not used to improve OpenAI models" OpenAI Help Center, Data controls in ChatGPT (checked 1 November 2026).
• [6] "a PowerPoint-native AI experience that lives in a sidebar inside Microsoft PowerPoint"; "while preserving editable slide structure"; "available globally to Free, Go, Plus, Pro, Business, Enterprise, ChatGPT Edu, and K-12 users"; "Free and Go include limited usage access." OpenAI Help Center, ChatGPT for PowerPoint (checked 1 November 2026).
• [7] "Slides is available on ChatGPT Pro, Business, and Enterprise plans with Space access."; "Slides is rolling out to eligible accounts."; "select the element, then select Request edits"; "PowerPoint, to continue editing a .pptx file." OpenAI Help Center, ChatGPT Slides (checked 1 November 2026).
• [8] "Documents and presentations have a maximum size of 512 MB per file."; "Presentations PPTX"; spreadsheet and image limits OpenAI Help Center, File uploads (checked 1 November 2026).
• [9] Supported-language list including Arabic, with no right-to-left statement on the page OpenAI Help Center, Supported languages (checked 1 November 2026).
• [10] "The plugin directory is available across ChatGPT plans."; installing or using one "depends on your plan, workspace, role, region" OpenAI Help Center, Plugins in ChatGPT (checked 1 November 2026).
• [11] "the subscription renews at $20 per month unless you cancel" OpenAI Help Center, ChatGPT offer terms (checked 1 November 2026).
• [12] DeepThink and Search on the chat entry; no image generation documented DeepSeek, home (checked 1 November 2026).
• [13] "Chat with leading AI models for free."; iOS, Android, and the Harness desktop app for macOS and Windows DeepSeek, download (checked 1 November 2026).
• [14] "public preview worldwide and open source"; "Organize files, analyze data, draft docs, and create slides"; the named files "Brief.docx", "Data.xlsx", "index.html", "label.ts", "analyze.py", "Report.pdf" and "README.md"; no PowerPoint icon and no .pptx anywhere on the page; "Open source, MIT" DeepSeek, Harness (checked 1 November 2026).
• [15] "Search Per 1,000 successful requests $0.75"; "Fetch Per 1,000 successful requests $0.75"; "These calls are currently free of charge; per-request billing starts at 00:00 Beijing time on November 1, 2026."; "deepseek-flash" and "deepseek-v4-pro" token prices; "Model prices are listed in units of per 1M tokens."; context 1M, maximum output 384K DeepSeek API docs, Models and pricing (checked 1 November 2026).
• [16] "DeepSeek's web and API search capabilities are powered by our in-house DeepSeek search engine"; "supports multilingual search"; entry dated 2026-10-17 DeepSeek API docs, Change log (checked 1 November 2026).
• [17] "The Files API lets you upload images and reference them later by file_id"; "Supported formats: JPEG, PNG, GIF, and WebP" DeepSeek API docs, Files API (checked 1 November 2026).
• [18] "text input, voice input, prompt, uploaded files"; "the right to opt-out of using your Personal Data for training our models"; "by emailing privacy@deepseek.com"; storage in the People's Republic of China DeepSeek, Privacy Policy (last updated 10 February 2026; checked 1 November 2026).
• [19] "Pick one, or copy the same generic MCP row for another server."; the client "starts a configured stdio command or connects to a configured Streamable HTTP URL, discovers MCP tools"; custom plugin markets DeepSeek Harness documentation, MCP and memory, linked from the Harness page as "Developer docs" (checked 1 November 2026; the page carries no visible date).
• [20] Custom connectors on Free, Pro, Max, Ultra, Team and Enterprise plans; "Free users are limited to one custom connector."; "Customize > Connectors"; "+ Add" then "Add custom connector" Claude Help Center, Get started with custom connectors using remote MCP (checked 1 November 2026).
• [21] NOXI listed in ChatGPT's plugin directory: name NOXI, subtitle "AI presentation maker", button "Install plugin", category Creativity ChatGPT plugin directory, NOXI (checked 1 November 2026, signed out).
• [22] One server address for Claude, ChatGPT, Cursor, Claude Code and Codex: "Every assistant connects to the same address. You sign in to NOXI in your browser." Product facts for the reference row: editable .pptx with real text boxes and embedded fonts, imports up to 200 MB, right-to-left Arabic slides with four Arabic fonts, 1,000 credits a month and about 25 for a deck NOXI, Connect NOXI to your AI assistant (checked 1 November 2026).
• [23] Quickstart: "Open Settings > Models, enter a DeepSeek API key, and save it."; the Web UI is launched through the root README, and the same documentation lists a command-line tool and a Python SDK DeepSeek Harness documentation, Quickstart (checked 1 November 2026; the page carries no visible date).

Drafted the text in DeepSeek? Turn it into a designed deck in NOXI and download it free as an editable PowerPoint or a PDF.