The fastest-growing class of AI tool in 2026 is not a chatbot — it is an AI coworker: a persistent agent that opens your files, runs multi-step tasks, and keeps working after you close the laptop. Anthropic released Claude Cowork in January 2026 as a research preview; OpenAI shipped ChatGPT Work on July 9, 2026 alongside a rebuilt desktop app with Codex baked in. Both move the interface from "type a prompt, read a reply" to "describe an outcome, watch the agent work." The honest verdict: Claude Cowork is the better choice if your work lives in local files and ad-hoc desktop workflows, because its "Record a skill" feature learns by watching your screen. ChatGPT Work is the better choice if you want cloud-hosted subagents, scheduled background tasks, and a path to publishing live websites through ChatGPT Sites. Many small businesses will end up using both.
What is an "AI coworker" (and why prompting alone is no longer the ceiling)
An AI coworker is a persistent agent that takes a goal (not just a single prompt), breaks it into steps, uses tools (browser, file system, code execution, third-party connectors) to complete each one, and reports back. The key shift is that the human no longer writes a perfect prompt for every sub-step — the agent handles that decomposition itself, asking only for an outcome and oversight.
Both products sit on top of agentic infrastructure that the parent model had before:
- Claude Cowork is described by Anthropic as Claude Code's agentic engine packaged for knowledge workers instead of developers, available on Claude Pro ($20/mo), Max, Team, and Enterprise plans through the Claude Desktop app (macOS since Jan 12, 2026; Windows since Feb 10, 2026).
- ChatGPT Work is the consumer/business surface of OpenAI's Codex agent stack, available on every paid plan including Plus ($20/mo) and free for all plans (including Free) on the new desktop app; the cloud-execution and Sites hosting extras require paid plans.
The model you used to prompt is now one layer inside an agent that can act. For small businesses, the practical question is whether your recurring work fits one of these two stacks — and which stack's strength matches your day.
Claude Cowork vs ChatGPT Work: the side-by-side feature map
| Capability | Claude Cowork | ChatGPT Work (incl. Codex) |
|---|---|---|
| Where it lives | Claude Desktop app (macOS, Windows) | ChatGPT desktop app; Work on web + mobile |
| Min plan for the agent | Pro $20/mo | Plus $20/mo; Free tier gets the desktop app's Chat + Work + Codex |
| Trains itself by watching you | Record a skill (screen + voice, since July 21, 2026) | Record-and-replay in Codex (similar launch window) |
| Runs on schedule / trigger | Yes (scheduled Cowork tasks) | Yes (Scheduled Tasks) |
| Spawns parallel subagents | Yes (Codex-style multi-agent) | Yes — Codex subagent framework, GA by default |
| Hosts a live website | Not a primary surface | ChatGPT Sites (public beta) — describe, build, deploy at a URL |
| Explores local files | Yes (sandboxed, mounted folders) | Yes via the desktop app |
| Browser agent | Claude in Chrome | Built-in ChatGPT browser (absorbed the Atlas browser Aug 9, 2026) |
| Connectors ecosystem | Claude connectors built on MCP (anyone can wire in) | Plugins directory (Slack, Drive, Teams, SharePoint, email, calendar) |
| Execution location | Local on your machine (mostly) | Cloud by default; Codex can run locally too |
| Pricing model | Draws from plan usage limits | Token-metered credits; included allowance then consumption |
| Best for | Knowledge workers who live in local files | Teams that want shared cloud agents + publishing |
Sources: Anthropic's Jan 12, 2026 Cowork launch; OpenAI's July 9, 2026 ChatGPT Work launch; OpenAI Help Center's ChatGPT Work and Codex doc; OpenAI pricing page; Anthropic's July 21, 2026 Skills announcement; OpenAI's Codex subagent docs.
What is Claude Record a skill (and why it changes who can automate)?
Record a skill is the feature that closes the gap between "I know how to do this task" and "the AI knows how to do this task." Instead of writing a multi-paragraph prompt describing how to pull a CRM report, you open the recorder, perform the workflow once on screen, talk through why you make each choice, and Claude converts the demonstration into a reusable, rerunnable Skill — no API or scripting required.
Anthropic announced it on July 21, 2026 to Pro, Max, and Team subscribers, accessed from the + menu inside a Cowork workflow. The early-tested behavior matters: real users taught ManyChat Instagram automations (a tool with no official API or MCP support) and the recorder generalized — it picked a different video and a new keyword on replay rather than copying the exact demo. That generality is the whole point; a recording is not training a model, it is producing a task-specific instruction package Claude loads when the matching need arises.
This is a structural unlock for small businesses because most operational knowledge never gets documented. The person who reconciles invoices knows the three edge cases and the one vendor email that always trips the import — and they cannot be talked into writing any of that down. Record-a-skill-style capture turns tacit, "show me once" expertise into a Skill that survives turnover.
For a deeper treatment of the broader shift away from manually written prompts, see our companion piece on how context engineering and skill-by-demonstration replace prompt engineering — it covers the same Anthropic feature but focuses on the longer-term pattern.
What is ChatGPT Sites (and when does a small business actually need it)?
ChatGPT Sites lets ChatGPT Work build and host a live, interactive web app at a shareable URL from a plain-language description — a tracker, an internal dashboard, a review workspace. OpenAI announced it on July 9, 2026; it has been in public beta for paid plans except Free and Go, with Pro, Pro Lite, Enterprise, and Edu getting access first. It is not available in the EEA, Switzerland, or UK at launch, and in Enterprise workspaces public publishing is off by default until an admin enables it.
When does a small business actually need it? When you would otherwise pay a developer to spin up a lightweight internal tool or status page. A field-services team can describe a job-status board, ChatGPT deploys a working app, and the team opens it with the same authentication as their ChatGPT workspace. It is not for regulated data — OpenAI's Sites Terms forbid payment-card data and protected health information (PHI), and the feature is preview-stage.
The closest competitor is Claude Artifacts, which publishes interactive apps at a claude.ai link — available on every Claude plan including Free, and the visitors' Claude subscription is billed for usage rather than yours. If you are an individual or small team without Business/Enterprise, Artifacts is the path you can reach today; if you are on a ChatGPT Business or Enterprise contract, Sites fits company-shaped publishing better. See our best AI website builders comparison for how Sites stacks up against Wix and other no-code paths.
How do you actually "show" the AI your workflow? (Step by step)
The shared mechanic both products converged on — watching you do something once, then replaying it autonomously — is the single biggest workflow change of 2026. Here is the canonical sequence on Claude Cowork (ChatGPT Codex's record-and-replay is intentionally parallel):
- Open Cowork and start a task. In the Claude Desktop app, enter a Cowork workflow.
- Open the + menu and choose "Record a skill."
- Perform the workflow on screen — clicking through the application, typing, switching windows. Claude logs each action as a discrete step.
- Narrate the why out loud. Surface the judgment a written SOP always misses: why this row, what condition triggers the fallback, what "done" looks like.
- Let Claude convert the recording into a Skill file. This is a written instruction package (not a fine-tune) the model will load next time.
- Test on a fresh example. Verify the skill generalized — different data, not the exact input from the demo. If it copied a detail specific to your one example (a vendor name, a date), refine.
- Replay on demand or on a schedule. The skill now triggers automatically when the matching task comes up, and runs unsupervised in the background.
The same recipe shows up in our AI agent workflows that replace your repetitive tasks guide — that piece covers five other recurring workflows that fit the same showing-not-telling pattern.
Which plan do you actually need, and what does it cost?
Claude Cowork is included on the Pro plan ($20/month), plus Max, Team, and Enterprise. Record a skill at launch required Pro, Max, or Team. Anthropic has not published a dedicated cost-per-task page; Cowork draws on your plan's existing usage limits.
ChatGPT Work is on every plan, including Free on the desktop app, with cloud-execution and Sites reserved for paid plans. Token-based consumption pricing went live for Business and Enterprise on July 6, 2026, and ChatGPT for Excel/Sheets metered charging is already in effect. As of the OpenAI rate card: GPT-5.6 Sol is roughly 10 credits per message; GPT-5.6 Sol Pro roughly 50 credits/message. ChatGPT for PowerPoint stays free for Business and Enterprise through August 6, 2026, then joins the same token-metered model. ChatGPT Sites pricing "to be announced" — beta usage is included up to plan-specific limits ChatGPT notifies you about as you approach them.
For small businesses, the cost shape that matters is not the sticker price — it is how fast your allowance burns on a long-running agent. A code-heavy or research-heavy task that the agent runs for an hour can consume a meaningful slice of a monthly plan. Neither vendor publishes a per-task price; both expect you to discover the burn rate empirically. If you have not yet, our Claude Opus 5 enterprise cost strategy guide maps the same per-token math to the effort dial that helps contain it.
When should you choose Claude Cowork over ChatGPT Work?
Pick Claude Cowork if:
- Your recurring work happens in local files — spreadsheets, PDFs, a CRM export dumped to disk, a folder of contracts — and pulling it into cloud connectors would be a project of its own.
- Your expertise is tacit and visual: you know it when you see it, writing it down would lose the nuance, and Record a skill captures the why a written SOP drops.
- You want open connectors — Claude's connectors are built on MCP (Model Context Protocol), so anything not in the directory can be wired in by anyone with a developer.
- You already pay for Claude Pro at $20/mo.
Pick ChatGPT Work if:
- You want cloud-hosted agents that keep running after your laptop is closed, with Scheduled Tasks kicking off automatically on a trigger (Slack mention, email arrives, calendar event).
- You want to publish live internal web tools at a URL through Sites — a dashboard, a tracker, a review workspace — without standing up a deployment pipeline.
- Your team wants shared workspace agents built once and rolled out across the org, which is the structure OpenAI invested in for its April 22, 2026 workspace-agents release.
- Your work already lives in Slack, Google Drive, Microsoft Teams, or SharePoint — OpenAI's plugin directory is broader at the connector level even if it is less open as a protocol.
For most small businesses the answer is "start with the one you already have a paid plan for, instrument a single recurring workflow for two weeks, then decide whether the other stack's missing capability justifies the seat." Our complete Claude setup guide walks through the seven settings that turn Claude from a search engine into an operating system; the same pattern applies to ChatGPT Work.
What this means for you
- Stop optimizing prompts for tasks an agent can own. If a workflow repeats weekly and has ≤10 discrete steps, the highest-leverage move in 2026 is recording it once and letting the coworker run it — not writing a better prompt next week.
- Pick the stack that already has your data. A small business on ChatGPT Plus with Slack + Google Drive integrations should start in Work; the same business with Claude Pro and a folder of local contracts should start in Cowork. The agent's value compounds with what it can already reach.
- Treat record-a-skill as a retention tool. The colleague who closes the books each month knows the corner cases. Capture one Skill per critical recurring task, and that knowledge survives their next job change.
- Reserve Sites/Artifacts for tools you would otherwise never get built. A small business rarely gets engineering time for an internal status board. Now it costs a sentence and a publish step.
- Budget empirically, not by sticker. Both vendors meter agent usage against an included allowance. Run a representative recurring task for two weeks of production use before scaling — the burn rate is the only honest number.
Should you switch off Claude entirely to ChatGPT Work (or vice versa)?
No — most heavy users in 2026 pay for both, because the strengths are complementary rather than overlapping. Claude Cowork is the stronger pick for local-file, show-don't-tell work; ChatGPT Work is the stronger pick for cloud-hosted, scheduled, publishing work. The "one subscription to rule them all" framing is a marketing story; the operating reality is that two ~$20/mo subscriptions cover roughly 80% of the useable AI-coworker surface and the cost is recoverable inside a single automated monthly workflow.
FAQ
Q: What is the difference between Claude Cowork and ChatGPT Work? A: Claude Cowork is Anthropic's agent in the Claude Desktop app that runs multi-step tasks against your local files and connected apps; ChatGPT Work is OpenAI's equivalent in the ChatGPT desktop/web/mobile surface, with stronger cloud-execution defaults, scheduled tasks, and the ability to publish live websites through ChatGPT Sites. Both turn prompting into delegating; Cowork leans local-plus-demonstration, Work leans cloud-plus-publishing.
Q: Do I need a paid plan to use Record a skill on Claude? A: At launch (July 21, 2026), Record a skill required Claude Pro, Max, or Team. Free-tier access was not part of the initial rollout. ChatGPT's record-and-replay equivalent sits inside Codex, which is available on every plan including Free on the new desktop app, but Codex on web/mobile is not selectable.
Q: Can either product publish a website for free?
A: ChatGPT Sites requires a paid plan (Pro, Pro Lite, Enterprise, Edu, then Plus and Business rolling out); it is excluded on Free and Go and is not yet available in the EEA, Switzerland, or UK. Claude Artifacts, by contrast, publishes interactive apps on every Claude plan including Free, at a public claude.ai link. If "publish a free interactive app today" is the goal, Artifacts is the reachable path.
Q: Is ChatGPT Work safe to connect to production business data? A: It depends on your governance posture. OpenAI's enterprise controls let admins restrict browser access, plugin access, local file access, and agent network permissions at the workspace, group, and user level. Two Codex vulnerabilities were disclosed and patched in early 2026 (a GitHub branch-name command injection patched Feb 5, 2026, and a DNS-based side-channel exfiltration patched Feb 20, 2026). Neither was known to be exploited before patching. Treat the published numbers as directional and run a proof-of-concept on a single low-sensitivity workflow before connecting production data.
Q: How many people are actually using these agent features? A: OpenAI's agents crossed 10 million users after the July 9, 2026 ChatGPT Work launch (Bloomberg, July 21, 2026), inheriting more than 5 million weekly users from Codex plus over 1 million non-development users who were already using Codex for work beyond coding. ChatGPT overall reports 900+ million weekly active users and 50 million paying subscribers as of February 2026 (OpenAI VP Nick Turley on X). Anthropic released Cowork on January 12, 2026 to Max subscribers and surveyed 1.2 million anonymized Cowork sessions in May; almost half of usage came from professionals simplifying operations and creating content.
Q: Which should a small business choose — Claude Cowork or ChatGPT Work? A: Start with the stack you already pay for and run one recurring workflow for two weeks of production. Claude Cowork wins if your work is local files and tacit visual processes: Record a skill captures judgment prompts cannot. ChatGPT Work wins if you want hosted cloud agents, scheduled tasks, and live published internal tools. Many small businesses will end up paying for both at $20/month each — a cost usually recoverable by automating a single monthly recurring task.
Discussion
0 comments