--- title: "ChatGPT Features 2026: Projects, Memory, Agent, Sora and More" description: "ChatGPT features 2026: Deep Research, ChatGPT Agent, Custom GPTs, Canvas, Memory, Projects, Tasks. How each works, tier requirements, transparency gaps." url: "https://suprmind.ai/hub/chatgpt/features/" published: "2026-05-07T19:10:17+00:00" modified: "2026-05-07T19:16:12+00:00" author: Radomir Basta type: page schema: WebPage language: en-US site_name: Suprmind --- # ChatGPT Features 2026: Projects, Memory, Agent, Sora and More ChatGPT Features Deep Dive # ChatGPT Features in 2026: What Works, What Was Killed, What to Use It For ChatGPT in May 2026 has the broadest feature set of any consumer AI product. It can read documents, browse the web, control your computer, generate images, transcribe and synthesize speech, run Python in a sandbox, remember things across conversations, group related work into Projects, schedule tasks, and host user-built Custom GPTs. It also recently lost a feature: Sora video generation, OpenAI’s flagship video model, was discontinued on April 26, 2026. This page walks through each feature in 2026, with honest notes on what each one is genuinely good for, the limits that surface in real use, and the tier required to access it. Where a feature has hallucination implications – and most of them do – the relevant data is anchored to [Suprmind’s AI Hallucination Rates and Benchmarks reference](https://suprmind.ai/hub/ai-hallucination-rates-and-benchmarks/). Where a comparison with competing models is sharper than ChatGPT alone, it is flagged. See also: [ChatGPT 2026 overview →](https://suprmind.ai/hub/chatgpt/) ## Projects – Persistent Workspaces Projects launched in November 2025 with Project Memory following in August 2025. Each Project acts as a self-contained workspace with three layers: a system instruction set, uploaded files (5 to 40 depending on tier), and a Project Memory scope that captures facts the model learns within that Project but does not bleed into main chat or other Projects. The architectural decision is that memory is partitioned. Memories created in main chat do not flow into Projects, and Project memories do not leak into other Projects or main chat. For users running multiple consulting clients, multiple research threads, or work and personal contexts in parallel, this isolation is the feature that makes Projects valuable rather than just folders. Tier limits matter. Free gets 5 files per Project. Go and Plus get 25. Pro, Business, and Enterprise get 40. Within those caps, individual files cap at 512MB, with 50MB for spreadsheets and 20MB for images. Up to 10 files can be uploaded per message. The token cap on text and document files is 2 million tokens per file, which means the model can index a substantial book inside a single file slot. What Projects do not have as of May 2026: built-in calendar features or collaboration features. Shared Projects (multi-user) are Business and Enterprise only. For a solo workflow with parallel contexts, Projects close 90% of the gap that custom Apple Notes or Notion-as-RAG-store solutions try to fill. For team work, the collaboration ceiling is real. ## Memory – The Persistent Profile Memory beyond Projects stores facts the model extracts from conversations – preferences, past decisions, personal context – in a profile editable at chatgpt.com/settings/personalization. Users can view, edit, or delete individual memory entries, or disable Memory entirely. Memory has no published expiration. It persists until you manually delete it. Number of stored items, token cost per memory injection, and exact retrieval mechanism are not publicly specified. The privacy posture is straightforward. Memory is opt-out, not opt-in by default. Disabling Memory excludes you from memory-based personalization but does not retroactively delete stored memories – you have to delete those manually. For Business and Enterprise customers, OpenAI’s data training opt-out is separate from Memory and applies regardless of Memory state. Memory is most useful for sustained work where the same context comes up repeatedly. A working professional whose role and project list and writing style preferences live in Memory does not need to re-establish them in every session. The friction reduction is real. The privacy trade is also real, and the manual-delete model does not match how most users expect “off” to work. ## Deep Research – Multi-Step Research Agent Deep Research is a multi-step research agent that issues sequential web queries, reads retrieved pages, synthesizes findings across sources, and produces a structured report with citations. Sessions take 5 to 30 minutes and can read dozens of pages. Unlike single-query web search, Deep Research builds an iterative search-read-synthesize loop where you can review and modify the proposed plan before execution. As of February 2026, Deep Research connects to any MCP (Model Context Protocol) server. This unlocks enterprise data integration without custom API plumbing – you can point Deep Research at your internal documentation, knowledge base, or proprietary datastore and it will treat them as sources alongside the public web. Tier availability: Plus gets 10 queries per month, Pro gets higher limits (exact count not publicly disclosed), Business and Enterprise included. Free does not get Deep Research. The honest caveat: Deep Research synthesizes from sourced web content. It does not independently verify facts. The report contains citations but you must verify claims against the originals. Per the [Suprmind Multi-Model Divergence Index](https://suprmind.ai/hub/multi-model-ai-divergence-index/) (April 2026 Edition, n=1,324 production turns), Research Analysis is the domain where Claude vs ChatGPT is the top combative pair, with 52.2% of contradictions in that domain being critical severity. If your research is consequential, cross-checking with another model is the practical answer. For comparison context: Perplexity’s Deep Research uses live web retrieval with a 32-hour average data freshness lag. ChatGPT’s Deep Research uses ChatGPT’s browsing layer with similar real-time capability. The Columbia Journalism Review citation audit found Perplexity at a 37% citation hallucination rate and ChatGPT at 67% when web search is disabled. With browsing enabled, both improve substantially. The question is not which Deep Research is more powerful. The question is whether you trust the report enough to act on it without manual verification. See also: [ChatGPT Deep Research vs Perplexity →](https://suprmind.ai/hub/chatgpt/vs-other-ai/) ## Canvas – Side-by-Side Editing Canvas is a side-by-side editing mode where the user message and the model output appear as a live collaborative document. You can edit the document directly, ask ChatGPT to revise specific sections, and track changes. It differs from a standard chat thread by preserving output as an editable artifact rather than a conversational reply. Canvas is most useful when iteration matters more than single-pass generation. Long-form drafting, document editing with multiple rounds of revision, code where you want to see the full file in one view, and structured outputs where you need to manipulate sections – all benefit from Canvas over chat threading. Available on Plus and above. The interaction model is intuitive enough that no formal training is needed. The limit is that Canvas is a single-document workspace. For comparing two drafts side by side, you still need two browser tabs. ## ChatGPT Agent – Computer Use ChatGPT Agent is the consumer-facing name for what was originally Operator (launched January 2025 for Pro users in the US, integrated into ChatGPT in July 2025). The agent operates a virtual machine with a visual browser, text browser, terminal, and OpenAI APIs. It can browse websites, click, type, scroll, execute code, download files, and interact with connected third-party services like Gmail and GitHub. For authenticated actions, a special browser view allows secure login without exposing credentials to the model. GPT-5.5’s score on OSWorld-Verified – the standard benchmark for computer-use agents – is 78.7%. The human baseline on the same benchmark is 72.4%. ChatGPT Agent is, by this measurement, performing better than humans on standardized desktop and browser tasks. GPT-5.4 was the first model to clear human baseline at 75%. GPT-5.5 extended the lead. What that means in practice: the agent can complete most multi-step tasks that involve clicking around websites, filling forms, and pulling data into structured outputs. It cannot reliably handle every edge case, every CAPTCHA, every authenticated login, every file format. The 78.7% OSWorld score also implies a 21.3% failure rate on standardized tasks – and your tasks are not standardized. Risk awareness matters. Agentic systems inherit standard agentic risk: irreversible actions (form submission, file deletion, payments), credential exposure risk, prompt injection from web content, unpredictable failure modes. OpenAI documents a “minimal footprint” principle and human confirmation for sensitive operations, but the discipline still falls on you. For a research run that pulls data from 30 websites into a spreadsheet, the agent is excellent. For an agent that sends emails, places orders, or modifies your calendar, the human-in-the-loop discipline is essential. Available on Plus, Pro, and Business at launch in July 2025. Enterprise and Edu followed in subsequent weeks. Session length and action-count limits are not publicly specified. ## Advanced Voice Mode – Spoken Conversation Advanced Voice Mode runs on a specialized audio model (the GPT-4o Audio pipeline) that processes spoken input and produces spoken output without intermediate text transcription. It supports emotional tone in some configurations and video input on Business with the “advanced voice with video” feature. A persistent user complaint: as of late 2025, Advanced Voice Mode users on Reddit reported the feature still felt tied to an older model with shallower depth than text-mode GPT-5.x. No public confirmation of a GPT-5.x audio upgrade has been issued. The quality gap between spoken ChatGPT and text-mode ChatGPT is real and visible in extended use. The API exposes a separate `gpt-realtime-1.5` endpoint for the best voice-in/voice-out experience. Audio in is $32.00 per 1M tokens, audio out is $64.00 per 1M tokens, with text-only paths at $4.00/$16.00. For developers building voice products, the realtime endpoint is where the latest capability lives. Standard ChatGPT users get the consumer Advanced Voice Mode, which trails by some margin. Available on Plus and above. Standard voice (a lower-fidelity voice mode) is Business and above. Advanced voice with video is Business and above. ## Sora Video Generation – What It Was, What Happened, What to Use Now Sora was OpenAI’s flagship video and audio generation model. The original Sora launched as a standalone web app in September 2024. Sora 2 released September 30, 2025 with substantially improved temporal consistency and 1080p output. ChatGPT integration was reported as planned in March 2026 per The Information. The integration never materialized. The Sora web and app experiences were discontinued on April 26, 2026. The Sora API will sunset on September 24, 2026. As of dossier date, Sora is listed as “Limited” on the Business tier feature matrix as a legacy access designation. Treat Sora as deprecated for any new use case. What Sora 2 was capable of at sunset: text-to-video generation up to 25 seconds at 1080p resolution, video-to-video editing with style transfer, character consistency across cuts, scene continuation from a reference image, and on Pro $200 specifically, non-watermarked output. The 1080p limit was the constraining feature for most professional use – serious post-production work needs 4K source. For social-format video, demo reels, marketing teases, and prototypes, Sora 2 was more than sufficient. The discontinuation surprised observers because Sora 2 was less than seven months old at sunset. OpenAI did not publish a formal explanation beyond the help center notice. Speculation in industry coverage points to compute prioritization, the cost-per-second economics of video generation, and a possible pivot toward video-from-Codex agentic flows. None of these are confirmed. What to use instead, as of May 2026: Runway Gen-4 for production-grade video. Pika 2.0 for fast iteration. Google’s Veo for prompt fidelity. Luma Dream Machine for motion quality. Each has trade-offs. None is a drop-in replacement for the ChatGPT-Sora integration that was planned and never shipped. For users who built workflows around Sora, the September 24, 2026 API sunset is the hard deadline to migrate. ## Code Interpreter (Advanced Data Analysis) Code Interpreter (renamed Advanced Data Analysis in late 2024) lets the model write and execute Python in an isolated sandbox. It accepts CSV, Excel, JSON, PDFs, and images, and produces charts, processed files, and computed results. The sandbox has no internet access. Code that calls external APIs must be run locally by the user. Code and output are visible in the conversation – you can audit what the model ran, copy snippets to your own environment, and modify approaches mid-task. Available on Plus and above with no toggle required since 2025. On the API via the `code_interpreter` tool in the Responses API. Sandbox execution time and compute caps are not publicly specified. File upload limits apply: 512MB per file, 50MB for spreadsheets, 20MB for images. The use cases that work best in Code Interpreter are data analysis on uploaded files, statistical work on small to medium datasets, document conversion (PDF to structured data, image OCR to spreadsheet), chart generation from raw data, and one-off scripts that need to run on confidential data without leaving the conversation. Anything requiring external API calls or libraries not pre-installed in the sandbox needs to be run locally. ## Custom GPTs and the GPT Store Custom GPTs are user-built versions of ChatGPT configured for a specific purpose: a system prompt, optional knowledge files (up to 20 files at 512MB each), configured tools (web search, image generation, code interpreter), and optional API actions. The GPT Store launched January 10, 2024 and now hosts hundreds of thousands of user-built GPTs. As of June 2025, builders can select from any available model when creating or running a custom GPT, not just GPT-4o. OpenAI added a “Recommended Model” setting that auto-applies if a user’s tier lacks access to the configured model. A documented friction point: if a custom GPT specifies a model unavailable to the user’s tier, OpenAI silently substitutes an alternative. The user may not be running the model the GPT was built around. This means a Custom GPT designed and tested on GPT-5.5 will behave differently when run by a Free-tier user routed to GPT-4o mini. The substitution is invisible. GPT Store browsing is on Free and above. Creating and publishing requires Plus or above. Workspace-private GPTs (private to a Business or Enterprise workspace) are Business and above. The use case that works best for Custom GPTs is encapsulating a workflow that you run repeatedly with stable inputs. A research assistant for one specific domain, a code reviewer with your team’s style guide built in, a customer-facing support agent with product knowledge files. The use case that does not work well is anything where you need fine-grained control over which model is running – the silent substitution undermines reliability for high-stakes work. ## Tasks – Scheduled Operations Tasks lets users schedule recurring or one-time operations – reminders, recurring research queries, scheduled reports – that ChatGPT executes at a specified time even when the user is not actively in the app. ChatGPT proactively suggests tasks from conversation context, with explicit user approval required before activation. Notifications come via push or email. Available on Plus, Business, and Pro from beta launch in January 2025. Free tier access is not confirmed as of dossier date. Limits on task count and execution windows are not publicly disclosed. The current state of Tasks is workable but limited. It is good for scheduled reminders, recurring weekly research summaries, daily news briefings on tracked topics, and time-shifted execution of one-off jobs. It is not a workflow automation platform – if you need conditional logic, multi-step branching, or tight error handling, you will outgrow Tasks quickly. For sustained automation, ChatGPT Agent or external orchestration tools are the next step up. ## File Uploads and Document Handling ChatGPT accepts PDF, DOCX, XLSX, CSV, TXT, JSON, HTML, images (JPEG, PNG, GIF, WebP), code files, and audio files for transcription. File size cap is 512MB per file, with separate caps of 50MB for spreadsheets and 20MB for images. Text and document files cap at 2 million tokens each. Per-message limit is 10 files. Per-Project limit ranges from 5 (Free) to 40 (Pro and above). Per-3-hour rolling window is 80 files on Plus. Storage limits run to 10GB per user and 100GB per organization on Business and Enterprise. The Business pricing page does not publish the exact storage cap explicitly. Parser fidelity matters more than the size limits. Plain text, structured CSVs, and DOCX parse cleanly. Complex multi-column PDFs with heavy formatting may experience extraction degradation. Tables in PDFs sometimes survive intact and sometimes do not. OpenAI does not publish a parser fidelity metric. There is no visible upload-quota indicator in the UI – file counting and limit resets are opaque to users. The practical advice: for high-stakes document work, run the document through Code Interpreter to extract text rather than relying on the inline file-reading layer. The extra step gives you a verifiable text artifact and surfaces parsing errors before they corrupt downstream output. ## Web Browsing and Search ChatGPT issues search queries through an internal retrieval layer, receives web results, and incorporates them into responses with citations. All GPT-5.x models default to having browsing capability available. Browsing is the single largest hallucination-reduction lever ChatGPT users have access to. Per Suprmind’s [AI Hallucination Rates and Benchmarks reference](https://suprmind.ai/hub/ai-hallucination-rates-and-benchmarks/), GPT-5’s hallucination rate drops from 47% to 9.6% with browsing enabled. That is a 37-point reduction that exceeds the effect of switching from GPT-5 to a different model entirely. The Columbia Journalism Review citation audit measured ChatGPT’s citation hallucination at 67% with browsing off versus dramatic improvement with browsing on. For citation-dependent work, browsing is not optional. Available on Free and above. API web browsing is metered at $10.00 per 1,000 calls. Search content tokens are free. The mechanism is straightforward. The model issues queries, receives results, attaches inline citations to claims linking back to retrieved URLs. Citations appear as footnote references in the UI. When browsing is not active, responses carry no citations – the model generates from training data. Deep Research reports include structured citations with source links. There is no formal distinction in the UI between training-sourced and web-sourced claims. All citations reference web URLs retrieved during the session. Knowledge from training carries no citation, creating an implicit credibility asymmetry: the model speaks confidently about what it knows from training, but only the web-retrieved claims are explicitly attributable. Users should treat unattributed assertions with the same skepticism as any single-source claim. FAQ ## Frequently Asked Questions How does ChatGPT’s memory feature work? + When Memory is enabled, ChatGPT extracts facts from conversations – preferences, past decisions, personal context – and stores them in a persistent profile. The model injects these stored memories into subsequent sessions automatically. Users can view, edit, or delete individual memory entries at chatgpt.com/settings/personalization or disable memory entirely. What is ChatGPT Deep Research and how is it different from regular search? + Deep Research is a multi-step research agent that issues multiple sequential web queries, reads retrieved pages, synthesizes findings across sources, and produces a structured report with citations. Unlike ChatGPT’s single-query web search, Deep Research takes 5 to 30 minutes per session and can read dozens of pages. It is available on Plus tier (10 queries per month) and Pro tier (higher limits). Can ChatGPT control my computer? + Yes, via ChatGPT Agent mode. The agent can control desktop software, operate browsers, fill forms, and execute multi-step workflows. On the OSWorld-Verified benchmark, GPT-5.5 scores 78.7%, above the human baseline of 72.4%. The agent is limited to software and browser control – it cannot access hardware sensors, make phone calls without integration, or access other users’ accounts. What is ChatGPT Canvas? + Canvas is a side-by-side editing mode where the user’s message and the model’s output appear as a live collaborative document. You can edit the document directly, ask ChatGPT to revise specific sections, and track changes. It differs from a standard chat thread by preserving the output as an editable artifact rather than a conversational reply. Does ChatGPT have a Tasks feature? + Yes. Tasks allows users to schedule recurring or one-time operations – reminders, recurring research queries, scheduled reports – that ChatGPT executes at a specified time even when the user is not actively in the app. Available on Plus tier and above. How does ChatGPT’s Projects feature differ from regular conversations? + Projects group related conversations under a shared context: instructions, uploaded files, and Project Memory that persists across all chats within that Project. A Project behaves like a persistent workspace – the model carries context from prior project conversations. Main-chat memories do not bleed into Projects, and vice versa. Is Sora available in ChatGPT? + No, not anymore. The Sora web and app experiences were discontinued on April 26, 2026. The Sora API will discontinue on September 24, 2026. The integration into ChatGPT that was rumored in March 2026 did not materialize before the product was shut down. What file formats does ChatGPT accept for uploads? + PDF, DOCX, XLSX, CSV, TXT, JSON, HTML, images (JPEG, PNG, GIF, WebP), code files, and audio files for transcription. File size limit is 512MB per file. Up to 10 files per message. Up to 80 files per 3-hour window on Plus. What is the difference between Custom GPTs and the GPT Store? + Custom GPTs are user-built ChatGPT configurations with a specific system prompt, optional knowledge files, and configured tools. The GPT Store is the marketplace where users browse, install, and use Custom GPTs built by others. Browsing the Store is on Free tier and above. Creating and publishing requires Plus or above. Why doesn’t Advanced Voice Mode use the latest GPT model? + Per user reports through late 2025, Advanced Voice Mode appears tied to the GPT-4o Audio pipeline rather than a GPT-5.x audio architecture. OpenAI has not publicly confirmed an audio upgrade. The API exposes a separate `gpt-realtime-1.5` endpoint for the latest voice capabilities. Standard ChatGPT users on Plus or above get the consumer Advanced Voice Mode that trails the API endpoint. ## Stop guessing. Start cross-checking. Suprmind runs your prompt across ChatGPT, Claude, Gemini, Grok, and Perplexity in parallel. See where they agree, where they disagree, and which insights only one model surfaced — before you act. [Start Your Free Trial](/signup/spark) [See How It Works](https://suprmind.ai/hub/platform/) --- *Source: [https://suprmind.ai/hub/chatgpt/features/](https://suprmind.ai/hub/chatgpt/features/)* *Generated by FAII AI Tracker v3.3.0*