AgentsAutonomy & tool use
OpenAI says Codex and ChatGPT Work generate more than half of output tokens in its EU ChatGPT Enterprise accounts
In OpenAI's EU ChatGPT Enterprise sample, Codex and ChatGPT Work now generate more than half of output tokens, and weekly users at small and medium businesses rose over eightfold since May 2026, though official data show only 20% of EU firms used AI in 2025.
From the OpenAI News article of October 14, 2026 and cited sources: OpenAI's own telemetry shows that two agentic products, the Codex coding agent and the ChatGPT Work agent, produce most of the output tokens in its EU ChatGPT Enterprise sample. [1] [2] [6]
SOURCES[1] OpenAI: The EU's next chapter in AI is agentic openai.com[2] Superpowerdaily: OpenAI says agent tools generate over… superpowerdaily.com[6] The Next Web: OpenAI launches ChatGPT Work thenextweb.com
AT Pulse research deskPublished · Updated · 5 min read · AI-assisted analysisAI-assisted, not human-edited. Methodology Corrections Who owns this site
- What happened
- OpenAI's own telemetry shows that two agentic products, the Codex coding agent and the ChatGPT Work agent, produce most of the output tokens in its EU ChatGPT Enterprise sample. [1] [2] [6]
- Original source
- OpenAI News article, company source
- Brief by
- AT Pulse research desk AI-assisted; publisher and contact on the desk page
- When
- October 14, 2026 article published
- Covered
- Oct 14 edition
- Evidence
- 11 sources across 11 sites
- 1OpenAI: The EU's next chapter in AI is agenticopenai.com
- 2Superpowerdaily: OpenAI says agent tools generate over…superpowerdaily.com
- 3WAM: 20% of EU enterprises use AI technologieswam.ae
- 4Business Forum: Romanian companies fall behind on AI usagebusinessforum.ro
- 5Sidley: EU lawmakers agree to delay AI Act obligationsdatamatters.sidley.com
- 6The Next Web: OpenAI launches ChatGPT Workthenextweb.com
Reach: not measured yet
OpenAI's own telemetry shows that two agentic products, the Codex coding agent and the ChatGPT Work agent, produce most of the output tokens in its EU ChatGPT Enterprise sample. ChatGPT Work can run across connected apps for hours. Weekly users of these tools at small and medium businesses grew more than eightfold from a May 2026 baseline. [1] [2] [6]
Usage intensity varies widely. The agentic share of tokens ranges from 22% of sampled accounts in Slovenia to 85% in Lithuania. EU frontier firms use about four times the tokens per user of typical EU firms but about half the output of non-EU frontier firms, and only 10% of global frontier firms are based in the EU. [1] [2] [8]
The figures count product activity, not finished tasks, productivity or national adoption. Eurostat found that 20% of EU firms with 10 or more employees used AI in 2025, with large firms well ahead of small ones. This suggests the main divide is now how deeply firms use AI rather than whether they use it at all. [2] [3] [10]
agent products account for most measured activity among OpenAI's EU enterprise customers.
Read the full assessment
Implication: firms will compete on depth of AI use, and teams deploying agents need permissions, audit trails and cost controls now that Article 50 transparency duties apply.
Executive brief
OpenAI's Economic Research team reports that in its EU ChatGPT Enterprise sample, two agentic products, Codex and ChatGPT Work, now produce more than half of all output tokens. Weekly users of those tools at EU small and medium businesses grew more than eightfold since May 2026 (OpenAI; Superpowerdaily). The analysis measures product activity inside OpenAI's own customer base. It does not measure completed work, productivity or national adoption. No dataset or detailed method has been published, so no one outside OpenAI has checked the figures. Official statistics put overall AI use at only 20% of EU firms in 2025.
What changed and event timeline
Eurostat sets the baseline
20.0% of EU enterprises with 10 or more employees used AI in 2025, up from 13.5% in 2024. Denmark led at 42.0% and Romania was lowest at 5.2% (;).
SMB baseline month
OpenAI measures its eightfold growth in small-business weekly active users from this month ().
AI Act omnibus deal
EU lawmakers provisionally agreed to delay high-risk obligations to December 2027 and August 2028. Parliament voted for the deal on June 16 ().
ChatGPT Work launches
OpenAI released an agent that works across apps and files and can run for hours, running on GPT-5.6 ().
Transparency duties apply
Article 50 transparency obligations for deployers still start on this date despite the omnibus delays ().
OpenAI publishes the EU analysis
It reports that agentic tools produce most enterprise output tokens, that usage varies widely by country, and that leading EU firms still trail the global frontier (;).
Capabilities and access
- ChatGPT Work: an agent that gathers context across connected apps (Slack, Teams, Drive, SharePoint, email, CRMs) and produces documents, spreadsheets, slides and web apps.
- Rollout: Pro, Enterprise and Edu users first, then Plus and Business (The Next Web).
- EU availability: none of the reviewed sources gives EU-specific availability terms.
Read the full section
- ChatGPT Work: an agent that gathers context across connected apps (Slack, Teams, Drive, SharePoint, email, CRMs) and produces documents, spreadsheets, slides and web apps. It runs on GPT-5.6; one listing calls the model "GPT-5.6 Sol" (The Next Web; Neuronfeed).
- Rollout: Pro, Enterprise and Edu users first, then Plus and Business (The Next Web).
- EU availability: none of the reviewed sources gives EU-specific availability terms.
Technical analysis for researchers and developers
- Metric: the share of output tokens generated through Codex and ChatGPT Work.
- Definitions: a small or medium business is an account with 200 or fewer associated users, not one defined by headcount.
- No dataset, company list or detailed method has been released (Blockchain.News).
Read the full section
- Metric: the share of output tokens generated through Codex and ChatGPT Work. Tokens track how much the tools are used, not whether tasks were finished. Long-running agents produce many tokens per task, so a token share likely overstates the agents' share of actual work compared with chat (inference) (Superpowerdaily).
- Definitions: a small or medium business is an account with 200 or fewer associated users, not one defined by headcount. "Frontier firms" are the top 10% of firms by output tokens per active user.
- Reproducibility: none. No dataset, company list or detailed method has been released (Blockchain.News).
Claims and evidence
- Agentic tools exceed 50% of output tokens in EU Enterprise accounts. This comes from OpenAI's own telemetry, and no independent source corroborates it (Superpowerdaily).
- Country shares range from 22% (Slovenia) to 85% (Lithuania). These describe sampled accounts, not national adoption.
- Gaps between firms: EU frontier firms use about 4x the tokens per user of typical EU firms.
Read the full section
- Agentic tools exceed 50% of output tokens in EU Enterprise accounts. This comes from OpenAI's own telemetry, and no independent source corroborates it (Superpowerdaily).
- Country shares range from 22% (Slovenia) to 85% (Lithuania). These describe sampled accounts, not national adoption.
- Gaps between firms: EU frontier firms use about 4x the tokens per user of typical EU firms. Non-EU frontier firms use about 2x the output of EU frontier firms. Only 10% of global frontier firms are EU-based.
- Security claim: OpenAI says ChatGPT Work blocked 100% of attempts to extract protected data during red-teaming. The Next Web notes this figure has not been independently verified (The Next Web).
Context and prior work
- Eurostat 2025: AI use is 55.0% at large firms, 30.4% at medium firms and 17.0% at small firms.
- US comparison: the Census Bureau's business survey found 19.8% of US businesses using AI in the period ending May 3, 2026 (Epignosis).
- Same pattern elsewhere: in the official data, countries with high general AI adoption (Nordic states, Lithuania) also rank high.
Read the full section
- Eurostat 2025: AI use is 55.0% at large firms, 30.4% at medium firms and 17.0% at small firms. The most common use is analysing written language, at 11.8% (Epignosis; WAM).
- US comparison: the Census Bureau's business survey found 19.8% of US businesses using AI in the period ending May 3, 2026 (Epignosis).
- Same pattern elsewhere: in the official data, countries with high general AI adoption (Nordic states, Lithuania) also rank high. That matches Lithuania's top position in OpenAI's sample. This is inference only, because the two datasets use different methods.
Limitations, safety and contested findings
- OpenAI itself says the figures capture product activity, not completed work, productivity or economic gains (Superpowerdaily).
- Comparable vendor deployment stories, such as Siemens on Salesforce Agentforce, are announcements and not audited results (Blockchain.News).
- The regulatory timeline was contested. Later sources describe a provisional agreement that awaited formal Council adoption (Sidley).
Read the full section
- OpenAI itself says the figures capture product activity, not completed work, productivity or economic gains (Superpowerdaily).
- Comparable vendor deployment stories, such as Siemens on Salesforce Agentforce, are announcements and not audited results (Blockchain.News).
- The regulatory timeline was contested. One report said a round of omnibus talks collapsed with the August 2026 deadline left intact (PPC Land). Later sources describe a provisional agreement that awaited formal Council adoption (Sidley).
Business and practitioner implications
- Treat token share as a vendor usage signal, not a benchmark. Measure task completion and error rates on your own workloads.
- That means permission boundaries, audit trails, cost controls and human approval steps (Blockchain.News).
- The main gap is how intensively firms use AI, not whether they use it.
Read the full section
- Treat token share as a vendor usage signal, not a benchmark. Measure task completion and error rates on your own workloads.
- Plan governance for agents now. That means permission boundaries, audit trails, cost controls and human approval steps (Blockchain.News). Article 50 transparency duties already apply, while high-risk deadlines have moved to 2027–28 (Sidley).
- The main gap is how intensively firms use AI, not whether they use it. Firms already using AI can compete on depth of use, such as agentic workflows wired into their CRM and document systems.