OpenAI ChatGPT User Growth Hits 1.2 Billion, but the Enterprise Race Is Still Open
OpenAI said ChatGPT reached more than 1.2 billion weekly active users, marking another sharp increase in OpenAI ChatGPT user growth. At DevDay 2026, the company also said ChatGPT Work and Codex now serve over 35 million people each week. Those numbers put OpenAI’s reach in a class of its own. They do not, however, settle the contest for enterprise AI.
The disclosure came less than three months after OpenAI said ChatGPT had crossed one billion weekly users. It also follows a year in which Anthropic became a formidable rival among business customers and software developers. OpenAI is therefore presenting two connected stories: consumer distribution is still expanding, and its work products are beginning to convert that reach into sustained professional use.
That second story matters more than the headline number. A vast chatbot audience gives OpenAI an entry point into companies, but enterprise adoption depends on security, reliability, governance, and measurable results. The 35 million figure suggests that the transition from conversation to delegated work is advancing. It does not reveal how deeply those users rely on the products.
What OpenAI Actually Announced at DevDay
The important change is not simply that ChatGPT added users. OpenAI now claims professional agents have reached mass-market scale.
OpenAI disclosed the figures during DevDay 2026 on September 29 in San Francisco. The company’s DevDay program positioned the event around technical builders, new APIs, and products that perform work beyond ordinary chatbot conversations.
The headline figure was more than 1.2 billion weekly active ChatGPT users. That represents an increase of at least 200 million from the one-billion milestone OpenAI discussed during the summer. It also extends the growth reported at DevDay 2025, when the company listed more than 800 million weekly users.
OpenAI separately said more than 35 million people use ChatGPT Work and Codex every week. ChatGPT Work is an agent for completing longer tasks across apps, files, browsers, and desktop environments. Codex began as a coding agent but has expanded into research, data analysis, document production, and workflow automation.
The wording around the 35 million figure deserves care. It appears to describe the combined weekly audience for ChatGPT Work and Codex, not two independent groups of 35 million users. OpenAI has not publicly provided enough detail to calculate overlap between those products.
A person who uses both services may be represented once in a combined measure or counted within separate product totals. The company also has not published a geographic breakdown, retention curve, or split between personal and organizational accounts. Those missing details limit comparisons with competitors that report different adoption metrics.
Still, the new total shows rapid movement from OpenAI’s earlier baseline. In June, the company said Codex alone had more than five million weekly active users. Its Codex adoption data described a sixfold increase following the desktop application’s February launch.
When ChatGPT Work arrived in July, OpenAI again cited more than five million weekly Codex users. The company presented Work as a broader agent that could prepare presentations, analyze budgets, research customers, and operate across workplace software. Codex supplied part of the technical foundation.
The latest 35 million figure therefore covers a wider category than the earlier Codex milestones. It captures OpenAI’s attempt to turn agentic execution into a general work surface. Agentic execution means software pursuing a multistep goal with tools, files, and applications, rather than returning one response.
OpenAI also says its products now reach 2.5 million businesses. That company count includes organizations using different OpenAI offerings and does not establish how many have deployed Work or Codex widely. Even so, it gives the professional-user figure a meaningful organizational backdrop.
The announcement creates the article’s central tension. OpenAI has unmatched consumer distribution and a rapidly growing agent audience. Yet the enterprise market rewards depth, predictable performance, and customer spending more than raw account totals.
Why OpenAI ChatGPT User Growth Creates a Distribution Advantage
OpenAI can introduce work agents to an audience that already understands ChatGPT, reducing one of enterprise software’s hardest adoption barriers.
Most workplace applications must first persuade employees to create accounts, learn an interface, and change established habits. ChatGPT enters that process with widespread recognition and a large base of existing users. Employees often encounter it personally before an employer approves it officially.
This bottom-up familiarity can accelerate organizational trials. A worker who already uses ChatGPT for drafting or research needs less education before testing a delegated workflow. A manager can begin with an existing behavior instead of launching an unfamiliar software category.
OpenAI’s recent product design supports that strategy. ChatGPT Work operates inside the broader ChatGPT environment while extending it into files, apps, and longer assignments. Codex similarly connects conversational instructions with concrete execution, including editing code, running tests, analyzing information, and generating deliverables.
That progression turns distribution into a product funnel. A user starts with a question, moves to document creation, and eventually delegates a multistep process. OpenAI can expose each new capability without asking that user to adopt an entirely separate interface.
The company’s own research indicates that this shift is already occurring among some users. In its agent usage study, OpenAI reported that non-developer Codex adoption was growing faster than developer adoption. It also found that users increasingly assigned longer tasks to the agent.
Those findings remain company-generated and should not be treated as independent proof of productivity. The research estimated human work time through models, while portions of the analysis used sampled users who permitted data use. It nevertheless illustrates how OpenAI defines the opportunity.
The company is no longer measuring success only through questions answered or messages sent. It wants users to assign complete outcomes, such as preparing a report or investigating a dataset. That change increases both the potential value and the consequences of failure.
Distribution helps OpenAI attempt this transition at unusual speed. A new feature can become visible to hundreds of millions of people through an existing product. Even a small conversion rate can produce a large professional audience.
The 35 million weekly-user claim offers evidence of that mechanism. It equals only a fraction of ChatGPT’s total audience, but it would still represent a substantial population using agent-oriented products. The number is especially notable because ChatGPT Work launched only in July.
OpenAI can also use consumer familiarity during enterprise procurement. Employees may request access because they already understand the brand and interface. Technology leaders may see less training friction compared with a product that workers have never used.
However, familiarity is not the same as organizational readiness. Enterprises require access controls, auditability, identity management, data policies, and predictable support. They also need agents to respect permissions across every connected system.
OpenAI has been adding those controls while presenting usage dashboards for ChatGPT Work and Codex. Its enterprise signals say agent activity represented 64 percent of combined ChatGPT and Codex output tokens among enterprise customers in June. That metric suggests deeper workloads, although it does not measure completed business value.
The distinction matters for buyers. Output tokens can show activity, but they cannot prove that a report was accurate or a workflow saved money. Organizations ultimately need evidence tied to cycle times, error rates, completed projects, or customer outcomes.
This is where OpenAI’s distribution advantage meets its first limit. The company can place an agent in front of users quickly. It must still persuade employers that those agents deserve access to sensitive systems and consequential tasks.
For knowledge workers, the shift also increases the importance of organizing trusted context. An agent performs better when it can retrieve current documents, decisions, and project history. A maintained AI knowledge base can help users distinguish durable context from scattered conversational output.
Distribution gives OpenAI the opportunity to define how people delegate work. Trust, context, and completed outcomes will determine how much of that opportunity becomes durable adoption.
Anthropic Is Still the Test of OpenAI Enterprise Adoption
The primary contest is not OpenAI against every AI laboratory. It is OpenAI’s distribution-led strategy against Anthropic’s enterprise-first position.
Anthropic has built its reputation around Claude, Claude Code, and business deployments involving complex knowledge work. Its products have gained traction with developers and organizations that value long-context tasks, coding performance, and a narrower workplace focus.
Earlier in 2026, that position put visible pressure on OpenAI. In April, the Associated Press reported that OpenAI was increasing its focus on professional work while facing stronger competition from Anthropic. OpenAI then had more than 900 million weekly ChatGPT users, but most did not pay for the service.
That contrast exposed a strategic weakness. OpenAI could dominate consumer attention without leading the highest-value enterprise workloads. A large free audience creates reach, but it also brings substantial serving costs and uncertain conversion.
Anthropic approached the market from the opposite direction. It concentrated heavily on developers, APIs, and organizational customers. That made Claude Code an especially important benchmark for Codex, even though the products continue to expand beyond software development.
Third-party spending data has reflected a close and unstable contest. An August analysis based on Ramp customer transactions found Anthropic at 41 percent and OpenAI at 39 percent among measured business customers. The business adoption data also suggested OpenAI had narrowed Anthropic’s earlier lead.
That dataset covers businesses visible through Ramp rather than the entire enterprise market. It measures paid relationships, not user satisfaction or workload depth. Still, it gives an independent signal that OpenAI’s enterprise push is translating into purchases.
The 35 million announcement adds a different type of evidence. It measures weekly people using Work and Codex, according to OpenAI, rather than companies making payments. Together, the figures suggest OpenAI is competing on both organizational acquisition and end-user activation.
Those are related but distinct achievements. A company can buy licenses that employees rarely use. Workers can also use an individual product intensively without an enterprise agreement. Durable enterprise adoption requires purchasing, active use, retention, and expanded deployment.
OpenAI’s consumer reach helps with activation. Anthropic’s enterprise focus may help with concentrated deployments and developer loyalty. Neither advantage guarantees the final outcome.
The products are also converging. Codex is moving beyond coding, while Claude’s workplace capabilities extend beyond software creation. ChatGPT Work gives OpenAI a general agent layer, while Anthropic has continued building tools for professional tasks.
That convergence makes performance comparisons harder. A benchmark score cannot capture every workflow, and buyers often combine several model providers. One organization may use OpenAI for general assistance, Anthropic for coding, and another provider for specialized workloads.
Multi-model adoption weakens the idea that one company will own every enterprise task. It also increases switching pressure. If interfaces and integrations standardize, buyers can redirect workloads toward the provider offering better reliability or economics.
OpenAI’s response is to connect its consumer product, work agent, coding agent, and developer platform. The company wants users to remain inside a common environment as their tasks become more complex. That can improve continuity, but it raises concerns about concentration and dependency.
Anthropic pressures that strategy by offering a credible alternative at the point where AI becomes operational. If Claude remains attractive for difficult professional tasks, OpenAI cannot rely on ChatGPT familiarity alone. It must prove that distribution leads to better deployment outcomes.
This is why OpenAI ChatGPT user growth matters without deciding the race. The 1.2 billion figure makes OpenAI difficult to displace at the consumer layer. The 35 million figure shows a growing bridge toward work. Anthropic remains the test of whether that bridge can carry valuable enterprise workloads.
The 35 Million Figure Changes the Story, but Does Not Settle It
OpenAI’s reversal is that enterprise weakness no longer looks inevitable, although the disclosed numbers still leave crucial questions unanswered.
At the start of 2026, the simplest competitive narrative separated the market into two strengths. OpenAI owned broad consumer attention, while Anthropic increasingly captured serious enterprise and coding demand. The latest data complicates that division.
Thirty-five million weekly Work and Codex users would constitute a significant professional audience under any reasonable comparison. It suggests OpenAI has moved beyond occasional experiments by a small developer community. It also indicates that general knowledge workers are entering agent workflows.
OpenAI has described several such workflows in product materials. These include gathering information across applications, building spreadsheets, creating presentations, researching customers, and producing internal tools. The company says Work can remain with a project for hours and divide it into smaller tasks.
One early example involved analyzing customer touchpoints across a company’s customer management system and email. The agent identified gaps in follow-up and generated an executive dashboard. OpenAI presented the example when it introduced ChatGPT Work.
That example demonstrates the intended mechanism, but it does not prove that typical users achieve the same outcome. Early trials often receive more support and involve carefully selected tasks. Enterprise buyers will need results across ordinary teams, imperfect data, and changing workflows.
Codex offers a clearer adoption timeline. OpenAI said weekly users exceeded three million in early April and five million by June. The broader Work and Codex category reached more than 35 million by late September, according to the DevDay disclosure.
The comparison is not perfectly equivalent because the later figure includes another product. It still shows that OpenAI widened both the audience and the definition of agentic work. The company is no longer presenting Codex as a specialist tool for software engineers.
This broadening creates new opportunities. A marketer can ask an agent to assemble research into a campaign brief. A finance team can investigate a budget variance. A product manager can combine customer feedback, meeting notes, and usage data into a weekly update.
It also creates additional failure modes. A coding agent generally works inside repositories, tests, and review systems with visible outputs. A general work agent may act across email, documents, calendars, browsers, and business applications.
Those environments contain ambiguous instructions and sensitive information. A mistake may not appear as a failed test. It may emerge as an incorrect report, an unauthorized disclosure, or an action performed with outdated context.
The value of a work agent therefore depends on more than model capability. It requires clear permissions, reliable retrieval, reversible actions, human review, and records of what happened. Organizations must decide which tasks agents can complete autonomously and which require approval.
The 35 million figure does not show how many users cross that threshold. Some may use Work for occasional research while others delegate recurring operations. Both count as weekly users, but they create very different economic value.
Frequency also needs context. A user active once during a week meets a common weekly-active definition. That measure does not reveal sessions, completed assignments, time saved, or continued use across several months.
OpenAI has not disclosed how the 35 million users divide between Work and Codex. It also has not provided organizational concentration. A small group of large deployments could produce different competitive implications than broad adoption across many companies.
Even the 2.5 million-business figure needs interpretation. It indicates extensive organizational reach, but a business may use an API, ChatGPT, Codex, Work, or several services. Customer count alone cannot show deployment depth.
None of these gaps invalidates the milestone. They define what the milestone can support. It is credible evidence that OpenAI’s professional products have expanded quickly, not proof that they dominate enterprise AI.
The reversal is therefore measured but meaningful. OpenAI’s consumer base once looked disconnected from its enterprise position. Work and Codex now provide a visible conversion path, while independent spending data shows the company narrowing Anthropic’s lead.
The next question is whether that path produces lasting work habits. Enterprise software becomes defensible when teams build processes, data connections, and governance around it. Weekly activity is the beginning of that process, not its conclusion.
What the User Numbers Do Not Show
OpenAI’s disclosures measure reach, while buyers need evidence about retention, reliability, security, and completed outcomes.
The first uncertainty is measurement. OpenAI has described more than 1.2 billion weekly ChatGPT users and more than 35 million weekly Work and Codex users. It has not released a detailed methodology for either figure.
Public companies often define active users in formal filings and maintain consistent reporting periods. OpenAI is privately held, so its operating metrics receive less standardized disclosure. Readers must treat the figures as company-reported milestones.
The second uncertainty is overlap. Work exists within the ChatGPT environment, and Codex capabilities can support Work tasks. A user may interact with both during the same project, making product-level comparisons difficult.
The third uncertainty is intensity. Weekly activity says little about whether agents complete high-value tasks. An experimental request and a recurring operational workflow can both produce an active user.
The fourth uncertainty is retention. Rapid launches can generate trials that fade when users encounter limits, errors, or unclear value. OpenAI has not shared cohort retention for Work or Codex.
The fifth uncertainty concerns economics. Agentic tasks can require long model runs, repeated tool calls, and substantial computing capacity. A rapidly growing audience is commercially attractive only if OpenAI can serve it sustainably.
OpenAI has also been expanding lower-cost access and advertising-supported availability. Those strategies can increase reach, but they complicate the connection between active users and revenue. A billion-user platform can still face pressure if professional usage remains expensive to deliver.
For enterprises, security presents a more immediate question. An agent that reads documents and operates applications receives broader access than a conventional chatbot. Every connection creates another boundary involving credentials, data exposure, and unintended actions.
Administrators need policy controls that match those risks. They must know which sources an agent accessed, which instructions it followed, and which actions it attempted. They also need mechanisms to pause, reverse, and investigate activity.
Accuracy remains equally important. Long-running agents can compound an early misunderstanding across several steps. A polished final document may conceal weak evidence, incomplete retrieval, or an incorrect assumption.
Human review does not disappear when agents improve. It moves to task definition, permissions, checkpoints, and evaluation. Teams that skip those controls may produce more work without producing better decisions.
There is also a concentration risk. If an organization connects one provider to its files, code, communications, and business systems, switching becomes more difficult. Buyers must balance convenience against dependency on one model family and interface.
OpenAI’s integrated approach makes that tradeoff especially visible. A common environment can reduce friction across ChatGPT, Work, Codex, and APIs. The same integration can centralize sensitive context under one vendor.
Competition provides some protection. Anthropic, Google, Microsoft, and other providers continue to develop agents and enterprise controls. Their presence gives buyers alternatives and discourages any single company from treating distribution as sufficient.
Yet switching is easiest before teams design their workflows around proprietary features. Enterprises should evaluate portability while adoption is still developing. They should also preserve source documents, decision records, and automation logic outside any one conversational history.
For individual knowledge workers, the practical risk is more mundane. An agent can produce abundant output while making source verification harder. Faster creation increases the need for disciplined review and traceable context.
The skeptical conclusion is not that the numbers are meaningless. It is that audience size answers only the first adoption question. OpenAI has shown that people are arriving, but not yet how consistently they remain or succeed.
Three Signals Will Show Whether OpenAI Can Convert Scale Into Work
The next stage will be decided by active enterprise use, competitive response, and evidence that agents complete valuable tasks reliably.
The first signal is a clearer breakdown of ChatGPT Work and Codex adoption. OpenAI should disclose separate user totals, overlap, retention, and organizational deployment. Those figures would reveal whether the 35 million milestone reflects broad experiments or durable habits.
Growth that continues after the launch period would strengthen OpenAI’s claim. Stable cohorts and rising use inside existing companies would matter more than another combined headline. A sharp slowdown would suggest that initial curiosity exceeded recurring value.
The second signal is Anthropic’s response in coding and enterprise workflows. OpenAI has narrowed the competitive story, but Anthropic retains strong developer credibility and organizational traction. New Claude capabilities, customer wins, or improved deployment controls could preserve that advantage.
Independent purchasing data will be especially useful. If OpenAI moves ahead across a broad business sample, its distribution-led strategy will look stronger. If Anthropic maintains leadership despite OpenAI’s audience, consumer reach will appear less transferable than expected.
The third signal is evidence tied to business outcomes. OpenAI and its customers need to report completed workflows, cycle-time reductions, lower error rates, and sustained use. Token output and active-user totals cannot substitute for those measures.
Independent evaluations would make those claims more credible. Buyers should look for studies comparing agent-assisted work with established processes across ordinary teams. Results should include failures, review time, and implementation costs, not only successful demonstrations.
These signals will also clarify what OpenAI ChatGPT user growth means for everyday users. If Work and Codex remain reliable across common tasks, delegation will become a standard layer of knowledge work. If oversight costs stay high, agents may remain valuable but specialized tools.
Developers should watch whether Codex maintains its identity as a serious engineering environment while expanding into general work. Enterprise buyers should examine governance, portability, and adoption beyond pilot groups. Knowledge workers should judge completed outcomes rather than the volume of generated material.
The 1.2 billion milestone establishes OpenAI’s extraordinary reach. The 35 million Work and Codex figure shows that a meaningful share of that audience is testing a more consequential relationship with AI. The unanswered question is whether those users are building dependable workflows or sampling the latest interface.
That question cannot be resolved by another launch event. It will be answered through retention, competitive switching, and the quality of work completed after the novelty fades. Watch what teams keep using, what they trust agents to access, and which provider earns a permanent place in their operating systems.
