Context Rankings
Research · Head-to-Head

Claude Projects vs. ChatGPT Projects for Long-Running Work

The two consumer AI apps most people already pay for both offer a Projects feature now. We spent two weeks running the same research, writing, and reference workloads through each to see which one actually keeps its head on straight.

Tested by Priya Venkataraman · August 7, 2026 · 5 rounds
Claude Projects
Anthropic
2rounds
87 / 100 overall
vs
ChatGPT Projects
OpenAI
3rounds
82 / 100 overall
The verdict

If your Project is a document-heavy workspace (a book manuscript, a client dossier, a codebase, a stack of research papers), Claude Projects is the better tool. Claude loads uploaded files fully into a 200K-token window and falls back to retrieval only when it has to, so answers stay grounded in the actual text you gave it. If you want a single app that also generates images, runs Python, browses the web, and shares the workspace with teammates, ChatGPT Projects is the more capable environment, and it's the only one of the two with a real free tier. Most working readers will do better on Claude for reference-heavy work and keep ChatGPT for the everything-else layer. If you can only pay for one and your day is mostly writing, editing, and long-document analysis, pay for Claude.

Both OpenAI and Anthropic have shipped a "Projects" feature over the last two years, and by mid-2026 they've converged on the same rough idea: a named workspace that carries custom instructions, a set of attached files, and a scoped memory across every chat inside it. The question isn't whether either app supports long-running work anymore. It's which one actually holds context the way a working professional needs it to.

We ran both features side by side for two weeks against the same four workloads: a 400-page research corpus for a policy brief, a client account with about 25 mixed documents (contracts, briefs, past deliverables), a small TypeScript codebase (~40 files) used as a reference library, and a weekly editorial project where drafts accumulate over time. Every round below names the concrete procedure we used, then the result. We tested on paid plans (Claude Pro and ChatGPT Plus, both $20/month) so the file limits below reflect what a typical paying reader will hit.

Round by round

Grounding in your uploaded files
WinnerClaude Projects

How we testedWe uploaded the same 12-document research corpus (about 380 pages of PDFs and DOCX) to a project in each app, then asked ten identical questions that required pulling specific claims, numbers, or quotes from named documents. We graded each answer on whether the fact was correct, whether the source document was correctly identified, and whether the model made anything up. A second reviewer spot-checked ten answers per side blind.

Claude answered more of the ten questions with the correct document and the correct quotation, and it hallucinated less on the questions where the answer wasn't in the corpus. The mechanics explain why: for smaller Projects, Claude loads uploaded files fully into a 200K-token window, and only switches to retrieval when the corpus grows larger. ChatGPT Projects lean on retrieval from the start, which is more scalable but leaves the model working from snippets rather than whole documents on medium-sized corpora. On the mid-size research load (big enough to matter, small enough to fit in a big window) full-context loading was the difference.

Memory across conversations in the same project
WinnerClaude Projects

How we testedOver five days we ran a rolling editorial workflow in each project: draft on Monday, revise on Wednesday, cross-reference on Friday, all in separate chats inside the same Project. We then opened a new chat and asked each app to recall specific decisions ("what did we agree the headline framing should be?") and to reuse voice and structure from earlier drafts.

Claude's project memory could search across past conversations in the same project and pull specific decisions back out. ChatGPT's project memory works differently: it creates automatic memory logs of what the model deems important within the project, more like curated sticky notes than a searchable archive. In practice that meant ChatGPT recalled the vibe of earlier chats but sometimes missed exact decisions, while Claude could quote the earlier discussion back. If you regularly ask "what did we decide last week," Claude is the better fit. If you want automatic context that builds up without management, ChatGPT is fine.

Tools inside the project
WinnerChatGPT Projects

How we testedWe assigned each project three tasks that required leaving pure text: generate a chart from a CSV, produce an infographic image tied to the knowledge base, and pull a fresh fact from the web to slot into an existing draft. We scored each attempt on whether it worked end-to-end inside the Project without leaving the app.

ChatGPT Projects can use the same tools as the rest of the app, including Canvas for drafting, image generation, Python in a sandbox for data work, web browsing, and, on paid plans, agent mode and deep research. Claude Projects don't include an image generator at all, which is a real gap for anyone whose workflow includes visuals. Claude has Artifacts and code execution, but for a "one app that does everything inside a workspace" test, ChatGPT is the broader environment.

File limits and free-tier reach
WinnerChatGPT Projects

How we testedWe compared published limits at each tier: files per project, per-file size, supported formats, and whether the feature is available at all on the free plan. We then tried to break each ceiling by uploading a 30MB PDF, a 40-file corpus, and mixed formats.

ChatGPT Projects have a firm cap on how many files a project can hold (5 for Free, 25 for Plus, 40 for Pro, Business, and Enterprise) but the feature is available to free users worldwide and includes project-only memory as a scoping control. Claude Projects allow effectively unlimited files as long as the total content fits within the context window (each file up to 30MB), and paid plans switch to retrieval mode that expands capacity by up to 10x when the window fills. Claude's ceiling is higher for paying users, but there's no free tier to speak of. If you need to hand a teammate a shared workspace without asking them to pay, ChatGPT is the only option here.

Sharing and collaboration
WinnerChatGPT Projects

How we testedWe tried to share a project with a second reviewer on each platform: inviting a collaborator, checking what they could see, and testing whether their chats added to the shared project's context.

ChatGPT rolled out shared projects to all plans (Free, Plus, Pro, and Go) on web, iOS, and Android in October 2025, with role-based access controls and project-only memory that turns on automatically once a project is shared. Claude restricts project sharing to Team and Enterprise plans; Free and Pro users can't share a project at all. For anyone collaborating with a partner, editor, or client on a paid plan short of Team, ChatGPT is the only workable choice.

Both features solve the same problem (“every new AI chat starts from zero”) and both do it well enough that the choice is now about which failure mode you can live with, not whether the feature works.

Where Claude Projects wins

Claude wins on the two things that matter most for reference-heavy work: how it holds the files you give it, and how it remembers what you’ve discussed inside the project. For projects under the 200K-token ceiling, the 200,000 tokens of active context (roughly 500 pages of text) can expand up to 10x through built-in RAG when needed , which means Claude reads the whole document first and only falls back to retrieval on the largest corpora. That’s the opposite of ChatGPT’s default behavior, and in our testing it produced fewer wrong-document citations on medium-size research loads.

Project memory is the other differentiator. Claude Project Memory offers full conversation access, you can explicitly ask Claude to reference past discussions and it will actually search through past project conversations and pull specific information, while ChatGPT Project Memory creates automatic memory logs of snippets it thinks are important but can’t access full conversation history. It’s more like sticky notes than a filing cabinet. For any workflow where “what did we decide last Tuesday” is a real question, Claude behaves more like an assistant with a filing cabinet and ChatGPT more like one with a stack of Post-its.

The catch is access. Project sharing is available exclusively for Team and Enterprise plan users, with granular permissions, either “Can use” or “Can edit”, and Free and Pro users cannot share projects. If your work requires handing a workspace to a collaborator without paying for a team plan, Claude Projects is a non-starter.

Where ChatGPT Projects wins

ChatGPT Projects wins on breadth. It reaches more people (free tier included), it does more things inside the workspace (image generation, Python, browsing, agent mode where entitled), and it’s designed for teams. OpenAI announced on September 3, 2025 that ChatGPT Projects is now available to free users worldwide, marking a significant expansion of functionality previously reserved for paid subscribers. That’s the single biggest change since we last looked at these features head to head.

The file caps are stricter than Claude’s, but published clearly: up to 5 files per project for Free, 25 for Plus, and 40 for Pro/Business/Enterprise, plus project-only memory controls for more tailored context. For most everyday knowledge bases (a few briefs, a style guide, a couple of reference docs) 25 files on Plus is enough. The moment your corpus grows into the dozens of PDFs, Claude’s “upload as many as you like as long as the tokens fit” model becomes the more practical ceiling.

Sharing is where ChatGPT’s lead is decisive. Projects in ChatGPT let you organize work into dedicated spaces where you can gather context, files, and instructions, and you can now invite teammates to collaborate in the same project. Shared projects are available on every tier including Free, and once a project is shared, project-only memory turns on automatically , which sidesteps the most common privacy footgun.

Who should pick which

Pick Claude Projects if the workspace is document-heavy and the answer needs to be grounded in the actual text you uploaded: a book, a research corpus, a codebase you want to query, a client’s paper trail. The 200K-token full-context load and the searchable project memory both pay off there, and Claude’s writing quality inside a well-configured project is still the single most-cited reason working writers prefer it. Pick ChatGPT Projects if you need images alongside the text, if you want to bring a collaborator into the workspace without paying for a team plan, or if the Project is really a container for the whole app (chat, browsing, Python, agents) rather than a reference library. If you subscribe to both, and a lot of working professionals now do, the split is straightforward: research and long-document work in a Claude Project, everything else (team coordination, visuals, data, day-to-day chatting) in a ChatGPT Project.

One thing worth watching: OpenAI has been steadily narrowing the gap on retrieval quality inside Projects, and Anthropic has been steadily expanding what Projects can do (memory, retrieval mode, MCP connectors). We’ll re-run this comparison in a few months. If either side ships a major change to how files are loaded or how memory works, the round tally could move.

Sources