Source: raw/Claude_COWORK_vs_ChatGPT_WORK_-_Which_should_you_use.md (current round), raw/Claude_vs_ChatGPT_-_Which_is_ACTUALLY_better_my_updated_2026_review.md (historical), raw/ChatGPT_5.2_vs_Gemini_3_vs_Claude_Opus_4.5_-_Ultimate_AI_Showdown.md (historical). Auto-caption transcripts; “Co-work” normalized to Cowork, “GPT6 Astra” to GPT-6 Astra, “Elliot” to Eliot.
Creator: Eliot Prince · Platform: YouTube
| Round | Video | URL | Duration | Published | Models tested |
|---|---|---|---|---|---|
| 3 (current) | Claude COWORK vs ChatGPT WORK — Which should you use | https://www.youtube.com/watch?v=cVHqpZpbG-s | 50:24 | 2026-09-15 | Fable 5.1 (medium) vs GPT-6 Astra (medium) |
| 2 (historical) | Claude vs ChatGPT — Which is ACTUALLY better (updated 2026 review) | https://www.youtube.com/watch?v=eeKkFTszvIo | 37:57 | 2026-04-13 | Opus 4.6 vs ChatGPT Pro |
| 1 (historical) | ChatGPT 5.2 vs Gemini 3 vs Claude Opus 4.5 — Ultimate AI Showdown | https://www.youtube.com/watch?v=IWVOb7mg4No | 20:43 | 2025-12-13 | GPT-5.2 Thinking vs Gemini 3 vs Opus 4.5 |
Eliot Prince ran the same kind of side-by-side test three times in ten months, each time on real tasks from his own business (content, marketing assets, a sales page, financial analysis). Round 3, recorded on Claude Cowork and ChatGPT Work the day before Anthropic announced that chat and Cowork were merging, is the one to act on. His conclusion there is that the two are now very similar and that the portable folder of context underneath matters more than the app. Rounds 1 and 2 are kept for the pattern they show: Claude led on analysis, branded documents and writing with story every time, and image generation was Claude’s gap every time. Eliot is a Claude-first creator who says he will stay in Claude, so read his verdicts as one practitioner’s view.
Key Takeaways
- Round 3 verdict (September 2026): close enough that the setup underneath decides it. “They’re very, very similar… the more important thing is what you’re running underneath.” He got “almost identical outputs” because both apps ran off the same folder of context. He is staying in Claude but expects to use ChatGPT more for marketing work because it can generate images.
- The “ChatGPT gives you more usage” claim did not hold in his test. On the 20 of extra credit and another 100 Max plan (5x usage) the same work used 27% of the 5-hour window. The plans differ, so this is not like for like; Eliot’s own reading is that the percentages “worked out probably the same.” The GPT-6 article collects similar reports of Astra draining usage.
- ChatGPT Work puts more modes on a menu; Cowork makes you ask for them. The ChatGPT Work + menu has set a goal, plan mode, sites (build and deploy a website), document, PDF, spreadsheet and presentation modes, browser control and appshots. Eliot says Cowork can do nearly all of this but “you just have to remember to invoke them in your prompting,” and that a goal mode exists only in Claude Code, not Cowork.
- Claude was easier to set up and run. Cleaner UI, skills uploaded from Customize in one step (ChatGPT needed the web settings page or an install prompt), automatic use of his voice-copywriting skill where ChatGPT did not invoke it, and phone notifications when Claude needed a decision so he could walk away.
- Four tests, roughly even. Marketing ideas and LinkedIn copy: close, with ChatGPT’s ideas slightly better and Claude better at using his skills. Carousels: even. One-shot live sales page: ChatGPT, via its sites feature. Bank-loan cash-flow forecast: Claude, faster and with a better written summary and chart.
- Learn the terminology before switching. The same words mean different things in the two apps (table below). The most confusing: a ChatGPT “plugin” is a Claude “connector,” while a Claude “plugin” is a bundle of skills plus connectors.
- Switching is now a one-click import. ChatGPT Settings → Import → “Import from another AI app” pulled in his Claude Cowork settings, instructions, plugins, skills, project folders, sessions and custom MCP servers. ChatGPT reads
AGENTS.mdwhere Claude readsCLAUDE.md. The ChatGPT-to-Claude direction is covered in Moving from ChatGPT to Claude. - Across all three rounds Claude won the analysis-and-deliverable tasks (keyword strategy, branded spreadsheets and decks, scripts, financial forecast) and never had native image generation. ChatGPT’s coding and website output improved from round to round.^[inferred — pattern across the three videos, not stated by Eliot]
Round 3 — Claude Cowork vs ChatGPT Work (2026-09-15)
Setup. Both desktop apps, Claude in light mode and ChatGPT in dark. Claude on Fable 5.1 at medium effort, ChatGPT on GPT-6 Astra at medium. Both pointed at the same “second brain” project folder after the import, with the same skills installed.
Timing note. Eliot shows Claude’s home with a chat / Cowork toggle on the input bar, and a separate entry for code. That is the August 2026 layout; Anthropic announced on 2026-09-16 that chat and Cowork merge into one Claude with no mode to pick, rolling out over the following weeks (see Claude Cowork). Interface remarks below describe the pre-merge app. He also shows a built-in web browser in Claude, which the wiki had recorded as announced but not yet live for other testers.
Terminology map (as Eliot describes it):
| Claude | ChatGPT Work | Note |
|---|---|---|
| Projects (sidebar item that holds a project’s chats, outputs, memory files and instructions) | Project chosen from your folder structure; appears as a running list in the sidebar | Claude gives each project its own page |
| Customize | Plugins | Settings area for extensions |
| Skills | Skills | Same concept; ChatGPT’s entry point is easy to miss in the top left |
| Connectors | Plugins | A ChatGPT plugin is a Claude connector |
| Plugins (skills + connectors bundled) | No direct equivalent named | |
CLAUDE.md | AGENTS.md | Project instruction file read on launch |
| Artifacts list (all outputs, including live artifacts) | Library, buried in settings | Eliot could not find an easy outputs view in ChatGPT |
| Scheduled tasks | Schedules | Work the same way in both |
| Record a skill (+ menu) | Record a skill | Both turn a narrated screen recording into a skill |
| Manually approve / automatically approve / skip all approvals | Ask for approval / approve for you / full access | He runs on automatic approval in Claude |
Features only one side had (per Eliot):
- ChatGPT Work: goal mode, plan mode as a switch, sites (publishes a live website at a shareable ChatGPT site URL), selectable output modes, appshots (attach a screenshot of your most recent app window to a message), a preview and edit pane for markdown output, native image generation.
- Claude Cowork: one-step skill upload, phone pings when a decision is needed, artifacts list, and plugins as installable skill-plus-connector bundles. Claude’s artifacts can be shared by link as a live page with a “made with Claude” bar, but Eliot says they are “not really live deployable websites.”
- Both: a built-in browser panel, Chrome control (Claude in Chrome on Claude’s side), scheduled tasks, record a skill, large connector catalogues.
The four tests:
| Test | Task | Result (Eliot’s call) |
|---|---|---|
| Marketing | Mine a YouTube transcript for ideas with his “search and rescue” skill, then write a LinkedIn post and five hooks | Close. ChatGPT finished first and pulled slightly better ideas; Claude used his voice-copywriting skill, ChatGPT did not. Neither post was one he would publish without rewriting. |
| Visual | Branded Instagram and LinkedIn carousel, told to avoid stock AI design | Even. Claude finished first with a clean, on-brand set; ChatGPT took 7 minutes and added generated images. |
| Web | One-shot a live sales page for his masterclass from his copy | ChatGPT. Its sites feature produced a tidier live page in 13 minutes 28 seconds and found his registration link and testimonials in the shared folder unprompted. Claude asked for the missing links and returned an artifact that “needs a lot of work.” |
| Analysis | Three-year cash-flow forecast for a bank loan from five messy finance files (management accounts, aged debtors and creditors, bank transactions, subscriptions ledger) | Claude. About 17 minutes, a spreadsheet built with formulas, an interactive closing-cash chart with scenarios, and a written summary flagging a likely covenant breach. ChatGPT reached the same breach finding and added a working-capital sheet, but stopped mid-run for a credit top-up and took close to half an hour. |
Round 2 — Claude vs ChatGPT in chat (2026-04-13, historical)
Historical: Opus 4.6 has since been replaced by the Claude 5.5 family and Fable 5.1, and chat has since merged with Cowork. The workflow and the setup points still apply.
- Setup parity. Same memory loaded in both, same project knowledge files, same custom instructions, same skills. ChatGPT project instructions are capped at 8,000 characters, so he had ChatGPT shorten his Claude instructions first.
- Memory style. Claude’s memory read as “one long flowing document”; ChatGPT stores short separate saved memories. He prefers Claude’s because more context fits.
- Workflow tested: deep research on a video topic → three angles with titles and hooks → outline with his structure skill → thumbnail → teaching dashboard built with a skill.
- Results: deep research even (Claude 10 min 35 s and 252 sources; ChatGPT 20 min and 371 searches). Angles and scripting to Claude: ChatGPT drifted off the beginner brief and needed a fresh chat, and its script was “boilerplate.” Thumbnail to ChatGPT only because Claude cannot generate images, though the ChatGPT likeness was unusable and Gemini’s was better. Dashboard to Claude: built, branded and published as a shareable artifact link while ChatGPT stalled. Eliot’s tally: Claude 4, ChatGPT 2.^[ambiguous — his spoken scoreboard does not add up cleanly; see Open Questions]
- Useful small differences. Claude’s “copy to project” puts a deep-research report straight into project knowledge; ChatGPT offers markdown, Word and PDF export instead. ChatGPT cited project files inline in answers.
- Pricing as stated: both start at 8. He advises against free plans for real work.
Round 1 — GPT-5.2 vs Gemini 3 vs Opus 4.5 (2025-12-13, historical)
Historical: all three models are superseded. Kept for what it shows about each platform’s strengths at the time.
- SEO keyword strategy from several thousand rows: Claude won. Quickest, one shot, a branded spreadsheet (via his brand skill) with pillars, intent, page type and priority, openable in Google Drive. GPT-5.2 needed a second pass with a template; Gemini returned thin output and CSV text.
- Client deck from the strategy: ChatGPT’s had the best charts but was off-brand; Claude’s was on-brand and well laid out; Gemini’s was text-heavy and “looks AI generated.”
- Email subject lines: all three similar.
- Website from a screenshot: ChatGPT best but slowest (built working forms); Claude close behind and fast; Gemini a hair behind.
- Image (career map infographic): Gemini’s Nano Banana model best; ChatGPT’s misspelled text; Claude wrote an interactive web page instead, which he could publish as a link.
- Claude context compaction: Opus 4.5 compressed earlier chat history so the conversation could continue, which Eliot called a “massive leap forward.”
Try It
- Decide by task, not brand. If your work is analysis, branded documents and long-form writing, Eliot’s results favour Claude. If you need images or a one-shot live website, ChatGPT Work’s image generation and sites feature led in his tests.
- Make your context portable first. Keep instructions, knowledge files and skills in a local folder with a
CLAUDE.md(and anAGENTS.mdcopy for ChatGPT). This is what let him run the same project in both apps. - Run a same-setup trial before switching. Load identical memory, files, instructions and skills in both, then run two or three of your real recurring tasks. Compare outputs and time, not demo prompts.
- Watch usage on your actual plan. Check the 5-hour and weekly meters in both apps during the trial, since Eliot found ChatGPT’s “more usage” reputation did not hold at medium effort on Astra.
- In Cowork, prompt for the modes ChatGPT puts on a menu. Say “plan this first, don’t do any work yet” for plan mode, name the output format (spreadsheet, deck, document) explicitly, and ask for Claude in Chrome when you want browser control.
- Turn on phone notifications for Claude decisions if you hand off long tasks and walk away.
Open Questions
- Do Eliot’s Round 3 interface notes survive the merge? He recorded on the pre-merge app. Whether merged Claude now exposes plan or goal behaviour without prompting is not covered.
- Is there really no goal mode in Cowork? Eliot says goal mode exists only in Claude Code; not checked against Anthropic documentation.
- Usage comparison on matched plans. His test pitted ChatGPT’s 100 Max tier, so it does not answer cost per task on equal plans.
- Round 2 scoreboard. He lists deep research as even, angles and scripting to Claude, thumbnails to ChatGPT, then says “coding the dashboard, one zero there for ChatGPT” although ChatGPT had stalled, before announcing “four to Claude and two to ChatGPT.” The dashboard point was probably meant for Claude.
- Round 1 image model name. The caption reads “Nano Banana 3”; he calls it “Nano Banana two” in April. Which Gemini image model he used is unclear.
Related
- Moving from ChatGPT to Claude — memory import, project transfer and full export.
- Claude Cowork — the September 2026 merge and the August layout Eliot tested on.
- GPT-6 Astra, Sol and Luna — the ChatGPT models, pricing and usage reports.
- Claude Fable 5.1 — the Claude model in Round 3.
- Artifacts — Claude’s shareable outputs, the counterpart to ChatGPT sites.
- Record a Skill — the feature both apps now share.
- Kahn v. Anthropic — what Max 5x and 20x are marketed as versus delivered.
- generate skill — one way to add image generation to Claude.
- Opus 5.5 vs GPT-6 Astra (AI Master) — a model-level counterpart: six real-task rounds, scored 5–1 to Opus, with Astra ahead on live-price research.