--- title: ChatGPT Work hit 20M users but barely anyone outside... description: ChatGPT Work, announced by OpenAI on July 9th, shows a stark internal-external adoption gap: 98% of OpenAI employees use Codex internally, but under 1% of... canonical: https://daily.dev/posts/chatgpt-work-hit-20m-users-but-barely-anyone-outside-openai-actually-uses-it-0vuihq62d twitter:card: summary_large_image twitter:site: @dailydotdev og:type: website og:site_name: daily.dev og:title: ChatGPT Work hit 20M users but barely anyone outside OpenAI actually uses it | daily.dev og:description: ChatGPT Work, announced by OpenAI on July 9th, shows a stark internal-external adoption gap: 98% of OpenAI employees use Codex internally, but under 1% of... og:url: https://daily.dev/posts/chatgpt-work-hit-20m-users-but-barely-anyone-outside-openai-actually-uses-it-0vuihq62d og:image: https://api.daily.dev/og/posts/0VUihq62D.png og:image:alt: ChatGPT Work hit 20M users but barely anyone outside OpenAI actually uses it og:image:width: 1200 og:image:height: 630 og:locale: en --- > ## Documentation Index > Fetch the complete documentation index at: https://daily.dev/llms.txt > Use this file to discover all available pages before exploring further. # ChatGPT Work hit 20M users but barely anyone outside OpenAI actually uses it **[Trends](https://daily.dev/sources/trends)** · 3 min read · 2 upvotes · 0 comments ## Summary ChatGPT Work, announced by OpenAI on July 9th, shows a stark internal-external adoption gap: 98% of OpenAI employees use Codex internally, but under 1% of individual subscribers and only 17% of organizations have touched ChatGPT Work despite 20 million users. The product is actually two things sharing a name — a cloud version with GPT-5.6, headless Chrome, persistent filesystem, Cloudflare Workers deployment, and sub-agent orchestration that can hit the open internet, and a local desktop version that's essentially a re-skinned Codex, unlike regular ChatGPT which can't browse. OpenAI's head of product frames the strategy as deliberate diffusion of Codex-style capability to non-engineers, contrasting with Anthropic's more confirmatory Claude Cowork approach. Critics like Peter Yang argue both are partial solutions compared to Grok Bot's approach for non-technical users. Concerns raised include unclear pricing versus heavy token usage, thin security documentation despite a ## Content OpenAI announced ChatGPT Work on July 9th, and the numbers tell two very different stories depending on who you ask. The headline number is 25 million active users, which OpenAI head of product Thibault Sottiaux announced alongside a usage reset for all paid subscribers. That sounds impressive. The other number, buried in a Wired-style profile of the product, is less flattering: individual subscriber adoption sits under 1%. Organizations are at 17%. Inside OpenAI, it's 98%. That gap is the whole story. Gergely Orosz put it plainly: "The Codex that we see and use is nowhere nearly as powerful as the one they see and use inside. Theirs integrated anywhere & everywhere inside OpenAI, nearly full access." In other words, OpenAI employees are using a version of this product that the rest of us aren't. The external product is a constrained, friction-heavy approximation of what convinced them this was worth building. What does ChatGPT Work actually do? More than most people realize. It's not just a chat interface with a fancier name. It runs a full headless Chrome browser that executes JavaScript against live page DOMs, has a persistent filesystem across sessions, can deploy websites via Cloudflare Workers, orchestrates sub-agents, and schedules automated prompts. The internet access isn't the sandboxed proxy that regular ChatGPT Chat uses. One security researcher flagged the obvious concern: private data access plus untrusted content exposure plus exfiltration channels is a "lethal trifecta" for prompt injection. The UX friction problem is real. Peter Yang's take cuts to it: "If you asked people to explain what the difference between ChatGPT Work and Codex is... I bet most people have no idea." His argument is that Grok Bot's framing (a computer that runs in the cloud) is actually more legible to non-technical users than anything OpenAI has shipped. The Anthropic comparison keeps coming up too. OpenAI's philosophy is a "magic box" that just handles things; Anthropic's Claude CoWork approach is more conversational and confirmatory. There's a real design argument here about whether non-technical users want autonomy or oversight. The adoption numbers suggest OpenAI hasn't answered it yet. GPT-5.6 brings an 80% price cut on the Luna tier and improvements in document processing. The model capability is probably fine. The question is whether the harness around it is good enough for the accountants and doctors OpenAI is pitching, or whether this stays a power-user product that OpenAI employees love and everyone else ignores. ## Questions this post answers ### What is the adoption rate difference between OpenAI employees and external users for ChatGPT Work? Internal OpenAI adoption of ChatGPT Work sits at 98%, while individual paid subscribers adopt it at under 1% and organizations at 17%. OpenAI head of product Thibault Sottiaux announced 25 million active users alongside a usage reset for paid subscribers, but the gap suggests the external product is a constrained version of what employees actually use internally. _daily.dev surfaces takes like this for anyone weighing whether an agentic AI tool is worth adopting._ ### Why is ChatGPT Work considered a security risk for prompt injection? ChatGPT Work runs a full headless Chrome browser executing JavaScript against live page DOMs and uses unsandboxed internet access rather than the proxy regular ChatGPT Chat relies on. Combined with a persistent filesystem and private data access, a security researcher described this combination of private data plus untrusted content plus exfiltration channels as a 'lethal trifecta' for prompt injection attacks. _security-minded developers evaluating agentic tools can track risks like this trifecta on daily.dev._ ### What new features does GPT-5.6 include and what changed in pricing? GPT-5.6 introduces an 80% price cut on the Luna tier along with improvements to document processing. The model's capability improvements come alongside ongoing concerns about whether the surrounding harness and UX are ready for non-technical users like accountants and doctors that OpenAI is targeting with ChatGPT Work. _developers tracking model pricing shifts can follow updates like GPT-5.6's Luna tier cut on daily.dev._ ## Community take How the wider developer community reacted, aggregated from 1 discussion and 139 comments across x (as of 2026-09-13). **TL;DR:** Replies largely agree that internal OpenAI usage looks nothing like the external product, attributing the gap mainly to permissions, integration access, and cost rather than a hidden smarter model, while many separately vent about surprisingly high API bills and confusing product overlap between Chat/Work/Codex. **Sentiment:** 15% positive · 35% mixed · 50% skeptical **The case for** - Some argue the internal/external gap is overstated since anyone can wire the same public models into internal tools, dashboards, and credentials with real permissions. - A few note this heavy internal adoption is described as a grassroots, organic phenomenon rather than mandated top-down usage. **The pushback** - Many argue the real gap isn't the model but permissions and integration: internal Codex has full repo/tool/system access that external users don't get. - Several say the accumulated organizational context and internal tooling access (the 'nervous system') can't be replicated by third-party or public harnesses. - Multiple commenters report shockingly high real-world API/token costs, calling their bills 'eye-watering.' - Some describe the unified Chat/Work/Codex product surface as confusing, unclear when to use which mode. - A few speculate legal/security teams outside AI labs are wary of prompt-injection and data risk, slowing adoption. - One suggests companies won't pipe confidential data into a third-party model/harness the way OpenAI uses its own first-party Codex. **By community** - x (mixed): Replies mix technical explanations for the internal/external capability gap (permissions, integration, accumulated context) with unrelated complaints about API costs and product confusion. **Hottest debate:** Whether the internal/external gap is really about a superior hidden model versus simply about permissions, integration access, and organizational context. **Open questions** - What fraction of Codex's total compute usage is internal OpenAI employees versus external customers? - Is the internal model actually the same model as the public one, or a different/less constrained version? - How much of the perceived internal advantage comes from unlimited token budgets versus deeper system integration? **Highlights** > @norbertomarques it's not that they save it, it's that other companies probably won't pipe their confidential data into a third-party model and harness for OpenAI, Codex is 1st party Companies like Ramp that build their own see same gains: > — [GergelyOrosz on x · 40 points, 4 comments](https://x.com/GergelyOrosz/status/2094546800069775731) > @GergelyOrosz @norbertomarques the underappreciated part of the ramp story is that an in-house agent compounds: every correction, incident, and review comment makes it better at your codebase. a third-party harness starts from zero each time. the moat isn't first-party models, it's accumulated context > — [jbtradin on x](https://x.com/jbtradin/status/2094565473102725484) > @GergelyOrosz the gap is never the model. it's what the harness is allowed to touch. full repo, internal tools, real deploys, and the thing stops feeling like a chat box. everyone else gets the sandboxed version. > — [benatcortexai on x](https://x.com/benatcortexai/status/2094555057601823138) > @GergelyOrosz Our API bill is eye watering > — [thsottiaux on x · 44 points, 27 comments](https://x.com/thsottiaux/status/2094586662798201081) > @MattJamesBoyle @GergelyOrosz I'm finding the unification to be overall confusing. Sometimes I'm in a work chat. Other times, I'm not. Gets even more confusing on desktop. It's not clear to me (outside of working in my repos which is rarer for me now on ChatGPT) when I should be in Chat, or Work, or Codex. > — [migidoes on x · 2 points, 1 comments](https://x.com/migidoes/status/2094670650912960999) **Source threads** - [x](https://x.com/GergelyOrosz/status/2094542273690153418) · 0 points · 139 comments --- Tags: [#ai-agents](https://daily.dev/tags/ai-agents), [#openai](https://daily.dev/tags/openai), [#chatgpt](https://daily.dev/tags/chatgpt), [#prompt-injection](https://daily.dev/tags/prompt-injection), [#openai-codex](https://daily.dev/tags/openai-codex) [View this post on daily.dev](https://daily.dev/posts/chatgpt-work-hit-20m-users-but-barely-anyone-outside-openai-actually-uses-it-0vuihq62d) ```json {"@context":"https://schema.org","@graph":[{"@type":"Organization","@id":"https://daily.dev/#organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180},"sameAs":["https://twitter.com/dailydotdev","https://github.com/dailydotdev","https://www.linkedin.com/company/daily-dev-ltd"]},{"@type":"WebSite","@id":"https://daily.dev/#website","url":"https://daily.dev","name":"daily.dev","publisher":{"@id":"https://daily.dev/#organization"},"potentialAction":{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https://daily.dev/search?q={search_term_string}"},"query-input":"required name=search_term_string"}}]} {"@context":"https://schema.org","@type":"TechArticle","headline":"ChatGPT Work hit 20M users but barely anyone outside OpenAI actually uses it","url":"https://daily.dev/posts/chatgpt-work-hit-20m-users-but-barely-anyone-outside-openai-actually-uses-it-0vuihq62d","mainEntityOfPage":{"@type":"WebPage","@id":"https://daily.dev/posts/chatgpt-work-hit-20m-users-but-barely-anyone-outside-openai-actually-uses-it-0vuihq62d"},"datePublished":"2026-08-31T00:08:21.040Z","dateModified":"2026-09-13T19:32:10.787Z","description":"ChatGPT Work, announced by OpenAI on July 9th, shows a stark internal-external adoption gap: 98% of OpenAI employees use Codex internally, but under 1% of...","image":"https://media.daily.dev/image/upload/f_auto,q_auto/v1/posts/fa90df2c86d67acce675aebc3de0b145?_a=AQAEuop","thumbnailUrl":"https://media.daily.dev/image/upload/f_auto,q_auto/v1/posts/fa90df2c86d67acce675aebc3de0b145?_a=AQAEuop","isAccessibleForFree":true,"articleSection":"Trends","inLanguage":"en","publisher":{"@type":"Organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180}},"author":{"@type":"Organization","name":"Trends","logo":"https://media.daily.dev/image/upload/s--ZfSp3asX--/f_auto,q_auto/v1780996004/logos/trends?_a=BAMAMiWQ0","url":"https://daily.dev/sources/trends"},"commentCount":0,"discussionUrl":"https://daily.dev/posts/chatgpt-work-hit-20m-users-but-barely-anyone-outside-openai-actually-uses-it-0vuihq62d","interactionStatistic":[{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":2},{"@type":"InteractionCounter","interactionType":{"@type":"CommentAction"},"userInteractionCount":0}],"keywords":"ai-agents,openai,chatgpt,prompt-injection,openai-codex","timeRequired":"PT3M"} {"@context":"https://schema.org","@type":"BreadcrumbList","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https://daily.dev"},{"@type":"ListItem","position":2,"name":"Trends","item":"https://daily.dev/sources/trends"},{"@type":"ListItem","position":3,"name":"ChatGPT Work hit 20M users but barely anyone outside OpenAI actually uses it"}]} {"@context":"https://schema.org","@type":"FAQPage","@id":"https://daily.dev/posts/chatgpt-work-hit-20m-users-but-barely-anyone-outside-openai-actually-uses-it-0vuihq62d#faq","mainEntity":[{"@type":"Question","name":"What is the adoption rate difference between OpenAI employees and external users for ChatGPT Work?","acceptedAnswer":{"@type":"Answer","text":"Internal OpenAI adoption of ChatGPT Work sits at 98%, while individual paid subscribers adopt it at under 1% and organizations at 17%. OpenAI head of product Thibault Sottiaux announced 25 million active users alongside a usage reset for paid subscribers, but the gap suggests the external product is a constrained version of what employees actually use internally. daily.dev surfaces takes like this for anyone weighing whether an agentic AI tool is worth adopting."}},{"@type":"Question","name":"Why is ChatGPT Work considered a security risk for prompt injection?","acceptedAnswer":{"@type":"Answer","text":"ChatGPT Work runs a full headless Chrome browser executing JavaScript against live page DOMs and uses unsandboxed internet access rather than the proxy regular ChatGPT Chat relies on. Combined with a persistent filesystem and private data access, a security researcher described this combination of private data plus untrusted content plus exfiltration channels as a 'lethal trifecta' for prompt injection attacks. security-minded developers evaluating agentic tools can track risks like this trifecta on daily.dev."}},{"@type":"Question","name":"What new features does GPT-5.6 include and what changed in pricing?","acceptedAnswer":{"@type":"Answer","text":"GPT-5.6 introduces an 80% price cut on the Luna tier along with improvements to document processing. The model's capability improvements come alongside ongoing concerns about whether the surrounding harness and UX are ready for non-technical users like accountants and doctors that OpenAI is targeting with ChatGPT Work. developers tracking model pricing shifts can follow updates like GPT-5.6's Luna tier cut on daily.dev."}}]} ```