Research acceleration: The view inside OpenAI
Two titles, both real. The heading above is how the lab announced this work. The document actually behind it is titled “The Shift to Agentic AI: Evidence from Codex” — everything below is read from that document.
OpenAI's own Codex usage data show agentic-AI adoption growing over fivefold in early 2026, with users delegating longer, more complex, and increasingly parallel work than with conversational AI.
It is a rare inside look at how a leading lab's own agentic-AI usage is evolving, offered as a preview of how delegation, complexity, and workflow reorganization may spread to other workplaces.
Drew Johnston · David Holtz · Christopher Ong · Prasanna Tambe · Alex Martin Richmond · Aaron Chatterji — meet the researchers →
Two readings, equal authority
How to choose: The paper’s words is verbatim — use it when you need to quote, or to judge how they write. Plain language is a paraphrase written for comprehension — use it when you want the idea fast. Neither is a summary of the other; they are two doors into the same room.
“We analyze usage data from OpenAI’s Codex tool to present large-scale evidence of how agentic AI technology, which can take actions on a user’s behalf, changes how people work. We use an automated, privacy-protecting pipeline to contrast usage across three populations: external personal-account users, external organizational-account users, and workers within OpenAI. We find that agentic AI usage is growing rapidly: the number of active users has grown more than fivefold in the first half of 2026, with the most rapid increase occurring outside the initial audience of software developers. Uptake is uneven across contexts: within OpenAI, Codex usage is nearly universal and has largely replaced business usage of ChatGPT. We document a similar shift to agentic tooling outside OpenAI, particularly within organizations, although external adoption remains lower and more uneven. In addition to headline usage figures, we observe measures of sophistication, and find that a growing number of users have used Codex to change their workflows substantially. We find that more than 10% of users manage three or more concurrent Codex agents at some point each week and that 26.6% use skills, which allow users to share instructions for complex workflows. Alongside these changes in usage practices, request complexity has increased: since the start of the year, the share of individual Codex users who submit at least one request for a task estimated to require more than eight hours for an experienced human to complete has increased nearly tenfold. Concurrently, output has grown rapidly—in June 2026, the median OpenAI employee in a legal role generated 13 times more monthly output tokens across Codex and ChatGPT than they did in November 2025, while the median researcher generated more than 50 times as many. We conclude by discussing the implications of these patterns for productivity, job reorganization, and workforce restructuring.”
OpenAI studied how people actually use its Codex agentic coding/work tool versus ChatGPT, comparing everyday personal users, business customers, and OpenAI's own staff. They found people are shifting from chatting with AI toward handing it bigger, more complex chunks of work, and this shift is furthest along inside OpenAI itself, where Codex has nearly replaced ChatGPT for work. The heaviest users are running several AI agents at once and reusing saved instructions, and the tasks they hand off keep getting bigger.
What this paper defines
Every definition below is the paper’s own sentence, with its locator. The plain gloss is a reading aid and is marked as one.
Agentic AI
“Generative AI systems increasingly differ in the extent to which they can act on a user’s behalf, or in other words, the extent to which they are agentic.”Introduction
In plain terms: AI that can take multi-step actions on a person's behalf, not just answer questions.
Individual users / Organizational users
“We refer to users on personal plans, including Free, Go, Plus, and Pro, as Individual users. We refer to users on Business and Enterprise plans as Organizational users.”Introduction, footnote 3
In plain terms: The paper splits external users into personal-plan customers versus business/enterprise-plan customers.
Persona (Developer / General Knowledge Worker / Personal)
“we classify Codex users into three broad personas based on their recent Codex requests: Developers, General Knowledge Workers, and Personal users.”Section 3
In plain terms: Each user is labeled by what kind of work they mostly ask Codex to do: coding, other job tasks, or personal life tasks.
Systematization
“We refer to this shift from ad hoc use to reusable workflow infrastructure as the “systematization” of agentic work.”Section 5.3
In plain terms: Turning one-off AI requests into saved, reusable routines instead of re-explaining the task every time.
Skill
“In OpenAI’s terminology, a skill is the authoring format for a reusable, task-specific workflow: a directory containing a required SKILL.md file, with name and description metadata and optional scripts, references, and assets.”Section 5.3, footnote 13
In plain terms: A saved, reusable set of instructions a user can invoke instead of retyping a workflow each time.
Plugin
“A plugin is an installable distribution unit, identified by a .codex-plugin/plugin.json manifest, that can package skills together with app integrations, Model Context Protocol (MCP) configuration, hooks, and supporting assets.”Section 5.3, footnote 13
In plain terms: A bundle that packages one or more skills together with app connections and other tooling.
Turn concurrency
“We restrict only to turns which run in different threads, and which overlap one another for at least 30 seconds.”Section 5.1, footnote 11
In plain terms: How many separate Codex agent conversations a user has running at the same time.
Task complexity (estimated human completion time)
“We quantify the complexity of users’ use cases using a prompt, which analyzes the text of a subset of queries and estimates the time that it would take for an experienced human worker to complete the task without the assistance of AI.”Section 4.2
In plain terms: An AI-estimated guess at how many hours a skilled human would need to do the same task without AI help.
What they actually did
Each step is a synthesis. Open any step to see the paper’s own sentence it was derived from, with its locator — so nothing here floats free of the source.
- Built an automated, privacy-protecting data pipeline comparing three user populations without researchers reading raw messages.
Trace this step to the paper
“We use an automated, privacy-protecting pipeline to contrast usage across three populations: external personal-account users, external organizational-account users, and workers within OpenAI.”Abstract
- Assigned each Codex user a persona label (Developer, General Knowledge Worker, Personal) based on their most common request type over the prior 30 days.
Trace this step to the paper
“Each request receives a persona label from an automated classifier, and each user is assigned the persona that appears most often in their Codex activity over the prior 30 days, with ties broken by the most recent request.”Section 3
- Validated the persona classifier by comparing a sample of employees' assigned personas against their actual HR job titles.
Trace this step to the paper
“We validated the persona classifier using a small sample of employees. For each employee, we compared the assigned persona against their HR job title to assess alignment.”Section 3, footnote 7
- Classified organizational users' job titles into department, seniority, manager status, and job-title class using an LLM classifier.
Trace this step to the paper
“We use the gpt-5-mini model with minimal reasoning effort to classify Organizational users’ job titles. The classifier uses only the supplied job title, and returns structured outputs for inferred department, seniority, people-manager signal, and cleaned job-title class.”Appendix C
- Classified Codex requests into a two-level task taxonomy based on the user's requested outcome rather than incidental tool use.
Trace this step to the paper
“we classify Codex requests into a fixed two-level task taxonomy. The classifier assigns each request to a primary task category based on the user’s requested outcome, rather than on incidental tool use during execution.”Section 4.1
- Estimated task complexity (human-equivalent completion time) for a random opted-in sample of individual users' prompts over time.
Trace this step to the paper
“We use the classifier to assign estimated task durations to the prompts sent by a 0.1% random sample of Individual accounts who have opted to allow their data to be used for model training.”Section 4.2
- Used gpt-5-mini with a dedicated prompt to estimate the active human work-time a Codex request would have required without AI.
Trace this step to the paper
“We use the gpt-5-mini model and the following prompt to evaluate the complexity of work performed by Codex:”Appendix F
- Benchmarked the complexity classifier against 1,000 Codeforces competitive-programming problems with known difficulty and solve-time metadata.
Trace this step to the paper
“To benchmark the performance of this model, we apply the classifier to 1,000 random Codeforces problems.”Appendix F
- Measured concurrency by counting each user's peak number of overlapping Codex turns in different threads during a given week.
Trace this step to the paper
“For each user, we calculate the number of overlapping turns they have in different threads during the week prior to June 11, 2026.”Section 5.1
- Measured daily agent runtime by summing active latency across a user's turns, excluding long idle gaps likely caused by waiting on the user.
Trace this step to the paper
“We measure the time an agent is active within each turn by summing the latency of all completed request-response pairs. We include gaps of up to 30 minutes between responses… We remove longer gaps, which are often driven by agents who are awaiting user input or paused.”Section 5.2, footnote 12
- Categorized skill/plugin use into five sources (preinstalled, curated, plugin, custom plugin, custom) to measure workflow systematization.
Trace this step to the paper
“In our analysis, we distinguish between five sources of skills:”Section 5.3
- Applied a disclosure-avoidance suppression rule before reporting any raw job-title strings, to protect individual identifiability.
Trace this step to the paper
“To limit disclosure risk, we apply a pre-specified suppression rule to raw job-title strings: after normalizing superficial title variants, we display a title only if it is observed in at least two distinct non-OpenAI organizations or in OpenAI’s internal sample.”Appendix C.2
Exactly what was run, and how
| Model | Developer | Temp | Effort / reasoning | Deployment | Other settings |
|---|---|---|---|---|---|
| gpt-5-mini | OpenAI | not reported | minimal | unstated | — |
| gpt-5-mini | OpenAI | not reported | not reported | unstated | — |
Source for gpt-5-mini settings
“We use the gpt-5-mini model with minimal reasoning effort to classify Organizational users’ job titles.”Appendix C — Job Title Classification
Source for gpt-5-mini settings
“We use the gpt-5-mini model and the following prompt to evaluate the complexity of work performed by Codex:”Appendix F — Query Complexity Prompt
What they reported — and what they left out
The paper names gpt-5-mini (minimal reasoning effort) for the job-title classifier and gpt-5-mini (reasoning effort unstated) for the query-complexity classifier, but never names a model, temperature, or deployment method for the persona classifier (Appendix D) or the task-type classifier (Appendix E), and gives no API/deployment details for any classifier.
The numbers they report
Active agentic-AI usage grew rapidly in the first half of 2026.
more than fivefold growth in weekly active Codex users, Jan 1 - Jun 1 2026
See it in the paper
“the number of active users has grown more than fivefold in the first half of 2026, with the most rapid increase occurring outside the initial audience of software developers.”Abstract
Codex's share of output tokens (versus ChatGPT) is far higher inside OpenAI than among external users.
99.8% (OpenAI workers) vs 63.3% (organizational users) vs 16.5% (individual users) of output tokens, as of 2026-06-11
See it in the paper
“as of June 11, 2026, Codex accounts for 99.8% of output tokens these workers generate across Codex and ChatGPT. Among organizational users, the corresponding share is 63.3%, while among individual users it is 16.5%.”Section 3
Very few individual (personal-plan) users have adopted Codex at all.
<1% of active individual users used Codex in the last 28 days
See it in the paper
“fewer than 1% of active individual users used Codex in the last 28 days.”Section 3
Organizational adoption of Codex is broader than individual adoption but still a minority.
17.3% of organizational users used Codex in the last 28 days
See it in the paper
“In the last 28 days, 17.3% of organizational users used Codex, several times the corresponding share among individual users”Section 3
Among engineers with Organizational accounts, average token share on Codex is much lower than the share among engineers who actually adopt it.
26.8% average share (quintupled since start of year) vs 88.3% share among adopting engineers
See it in the paper
“although Codex accounts for just 26.8% of the output tokens generated by the average engineer with an Organizational account, Codex usage accounts for 88.3% of the total output tokens generated by users of this type.”Section 3
Legal teams show low average Codex usage but higher usage once concentrated among adopters.
1.9% average share vs 17.6% of total tokens (legal function, Organizational accounts)
See it in the paper
“among legal users, Codex accounts for 17.6% of total output tokens, despite accounting for only 1.9% of tokens generated by the average user.”Section 3
The share of individual users sending at least one hour-plus-complexity task roughly doubled in five months.
35.4% (Dec 2025) to 70.2% (May 2026)
See it in the paper
“in December 2025, only 35.4% of active Individual users sent at least one prompt that we estimate would have taken an experienced human at least one hour to complete… in May 2026, we estimate that 70.2% of users sent at least one prompt of similar or greater difficulty.”Section 4.2
The share of users sending at least one 8-hour-plus complexity task grew roughly tenfold.
2.1% to 25.6% since December 2025
See it in the paper
“the share of users who send at least one prompt that would have taken an experienced human 8 hours to complete has risen from 2.1% to 25.6%.”Section 4.2
The most complex requests tend to arrive at the start of a session rather than later in it.
first turn >2x as likely as fourth turn to require >1 hour of human work
See it in the paper
“the first turn of a thread is more than twice as likely as the fourth turn to require action that would take an experienced human more than one hour.”Section 4.2
Concurrent/parallel agent use is common within OpenAI but rare among external users.
67.4% (org) and 63.9% (individual) use no concurrent turns; 10.7% of OpenAI users use only a single workflow, 28.6% ran 5+ concurrent agents
See it in the paper
“Roughly 67.4% of Organizational users and 63.9% of Individual users do not use concurrent turns at all during this period… Only 10.7% of users in this group use a sole workflow at any one time during this period, and nearly 28.6% managed five or more concurrent agents at some point during this period.”Section 5.1
Heavy OpenAI users run agents for far longer per day than the typical employee, and that gap is growing.
median 2.5 hours/day vs 99th-percentile ~71 hours/day; +88% cumulative daily runtime since 2026-04-07 (99th percentile)
See it in the paper
“on June 11, 2026, the median employee had Codex turns running for 2.5 hours… OpenAI employees at the 99th percentile of the distribution have recently run about 71 hours of agent turns within the average day… 99th percentile users within OpenAI increasing their cumulative daily runtime by nearly 88% since April 7, 2026.”Section 5.2
Skill use (reusable saved instructions) is far more common within OpenAI than among external users.
25.7% (individual) / 30.4% (organizational) / 96.2% (OpenAI) invoked a skill in the 7 days ending 2026-06-11
See it in the paper
“In the 7-day window ending on June 11, 2026, 25.7% of active Individual Codex users and 30.4% of active Organizational Codex users invoked at least one skill. Within OpenAI, skill use is nearly universal: 96.2% of active Codex users invoked at least one skill.”Section 5.3
Skill use rose sharply over a few months.
5.4% (2026-03-01) to 26.6% (2026-06-11)
See it in the paper
“the share of active Codex users invoking any skill rose from 5.4% on March 1, 2026 to 26.6% on June 11, 2026.”Section 5.3
Output-token volume rose dramatically across every OpenAI job function, with legal and research growing the most.
at least 10x in every job function (Nov 2025-Jun 2026); legal 13x; research >50x
See it in the paper
“in June 2026, the median OpenAI employee in a legal role generated 13 times more monthly output tokens across Codex and ChatGPT than they did in November 2025, while the median researcher generated more than 50 times as many.”Abstract
The task-complexity classifier's estimates correlate positively with independent, external difficulty measures.
r=0.65 vs fastest-solver completion time; r=0.69 vs Codeforces difficulty score
See it in the paper
“a correlation of 0.65 between the model-predicted time and the average completion time among the five fastest human solvers of the problem, and a correlation of 0.69 between the difficulty score and the model-predicted time.”Appendix F
What they assert, beside what they showed
Left is the claim in the paper’s own words. Right is the data offered for it. Where the two do not fully meet, a gold band names the distance.
Agentic AI adoption is growing quickly and is no longer confined to the original software-developer user base.
“We find that agentic AI usage is growing rapidly: the number of active users has grown more than fivefold in the first half of 2026, with the most rapid increase occurring outside the initial audience of software developers.”
“Figure 2 shows that Codex adoption is not limited to the initial developer base. Developers remain an important share of users, especially among individual and organizational accounts, but growth is faster among Non-developers.”
Section 3Codex use represents delegated production ('doing') rather than the mostly conversational ('asking') pattern seen in prior conversational-AI research.
“Second, Codex use is strongly oriented toward delegated production. Users ask the model to carry out concrete work tasks like debugging, refactoring, validating changes, configuring applications, drafting documents, and analyzing data.”
“This contrasts with conversational AI usage patterns documented in Chatterji et al. (2025). In that context, researchers found that a larger share of work focused on “asking” (which represented nearly half of all prompts), than on “doing”.”
Section 2, footnote 6OpenAI's internal usage pattern, while unrepresentative today, previews what agentic AI use will look like elsewhere once adoption frictions fall.
“OpenAI usage is therefore not representative of the typical organization today. However, it provides a view of what agentic AI use may look like in the future, when adoption frictions are minimal.”
“OpenAI is an unusually favorable environment for agentic AI: workers are highly familiar with frontier models, usage is cheap at the margin, organizational buy-in is high, training and informal knowledge sharing are common, and many workflows are close to the systems being developed.”
IntroductionIntensive users organize Codex use around large, repeatable, parallel workflows rather than one-off requests.
“Fourth, intensive users organize Codex use around large, repeatable, and parallel workflows.”
“Only 10.7% of users in this group use a sole workflow at any one time during this period, and nearly 28.6% managed five or more concurrent agents at some point during this period.”
Section 5.1Rising skill/plugin use reflects genuine 'systematization' of work into reusable routines, reducing the need to repeatedly re-supply context.
“Systematization is an important step toward delegated production: without it, users must repeatedly supply task context, procedural guidance, and instructions, limiting the extent to which work can be handed off.”
“the share of active Codex users invoking any skill rose from 5.4% on March 1, 2026 to 26.6% on June 11, 2026.”
Section 5.3The task-complexity classifier reliably distinguishes simple from complicated tasks.
“we interpret the evidence from this benchmarking exercise as indicative of the fact that the prompt and model we use to predict human completion times are able to distinguish between simple and complicated tasks.”
“We find a strong, positive correlation between the model-estimated difficulty and each of the two difficulty measures derived from Codeforces, with a correlation of 0.65… and a correlation of 0.69 between the difficulty score and the model-predicted time.”
Appendix FNear-term measured effects of agentic AI likely understate its longer-run productivity potential, because organizations have not yet redesigned workflows around it.
“The history of technological adoption suggests that the near-term effects of agentic AI may understate its longer-run productivity potential if firms have not yet discovered, adopted, or scaled the new production processes that the technology makes feasible.”
“We draw here on the canonical example in David (1990), who studies the delayed productivity gains from the transition from steam power to electric power in manufacturing.”
Section 5How they frame it, and what they want next
Their framing
The authors frame agentic AI as categorically different from a more capable chatbot, arguing the key margins are what work gets delegated, how much execution the system performs, and whether users build repeatable/parallel workflows around it. They present OpenAI's own internal usage as an unrepresentative but informative preview of where the frontier is heading rather than as evidence about typical organizations today.
Register: The authors state their own measured statistics in confident, declarative terms ('we find', 'we document'), while consistently and explicitly flagging that OpenAI-internal findings may not generalize, and framing forward-looking productivity claims as suggestive analogy rather than proven fact.
Where they hedge
“OpenAI usage is therefore not representative of the typical organization today.”Introduction
“These measures do not fully capture organizational redesign, but they provide early evidence on whether agentic AI use is moving beyond one-off assistance toward more persistent, repeatable, and parallel workflows.”Section 5
“We believe that this is in part because we are measuring our modeled prediction against the speed of a nonrandomly selected group, which is largely made up of competitive programmers.”Appendix F
“Our evidence from OpenAI is not necessarily representative of the typical organization, but it shows what agentic AI use can look like when adoption frictions are low: use expands across functions, output rises sharply, and work increasingly takes the form of delegated tasks rather than isolated exchanges.”Conclusion
What they say it means
- Traditional usage metrics like active users or message counts may stop being informative measures of AI's impact as agentic use spreads.
the paper’s words
“Standard measures of AI use, such as active users, chats, or message volume, may become less informative as agentic systems diffuse. Future analyses may need to track delegated task complexity, runtime, workflow reuse, concurrency, and production output.”Conclusion
- As routine execution gets offloaded to agents, the relative value of human judgment, oversight, and coordination increases.
the paper’s words
“Agentic systems make it possible to offload larger, more repeatable, and more modular units of work, potentially reducing the importance of some routine execution tasks while increasing the importance of judgment, oversight, coordination, and review.”Conclusion
- Jobs may shift toward directing and reviewing AI agents rather than performing each task step directly.
the paper’s words
“Jobs may increasingly involve directing, monitoring, and integrating the outputs of AI agents rather than executing each component task directly.”Conclusion
- Organizational structures such as team composition, hiring, and career ladders may need to change as parallelized agent labor spreads.
the paper’s words
“These changes could affect team composition, hiring needs, career ladders, and the distribution of work across skill levels.”Conclusion
What they call for next
- Future research should measure new dimensions of AI use — delegated task complexity, agent runtime, workflow reuse, concurrency, and production output — rather than relying only on legacy usage metrics.
the paper’s words
“Future analyses may need to track delegated task complexity, runtime, workflow reuse, concurrency, and production output.”Conclusion
Limitations they state
“OpenAI usage is therefore not representative of the typical organization today.”Introduction
“Tool use is an imperfect proxy for agency: some tool invocations are part of simple conversational interactions, while some agentic workflows involve limited tool use.”Introduction, footnote 1
“These measures do not fully capture organizational redesign, but they provide early evidence on whether agentic AI use is moving beyond one-off assistance toward more persistent, repeatable, and parallel workflows.”Section 5
“We believe that this is in part because we are measuring our modeled prediction against the speed of a nonrandomly selected group, which is largely made up of competitive programmers.”Appendix F
Moves worth stealing
Organizes the entire empirical narrative around four numbered 'stylized facts' announced up front, giving readers a scannable roadmap before the detailed sections.
“We organize the evidence around four stylized facts.”
Anchors a new finding in a canonical academic reference (the electrification productivity-paradox literature) to argue near-term data may understate long-run impact, borrowing established credibility rather than resting only on new data.
“We draw here on the canonical example in David (1990), who studies the delayed productivity gains from the transition from steam power to electric power in manufacturing.”
States its own study's non-representativeness as an explicit caveat rather than letting the reader infer it, while still arguing for the data's forward-looking value.
“OpenAI usage is therefore not representative of the typical organization today. However, it provides a view of what agentic AI use may look like in the future, when adoption frictions are minimal.”
Validates an internal LLM-based measurement tool against an independent external benchmark before relying on it for headline results, building measurement credibility ahead of using the measure.
“To benchmark the performance of this model, we apply the classifier to 1,000 random Codeforces problems.”
Leads with a privacy-protection framing in the abstract itself, preempting concerns about analyzing usage data at this scale.
“We use an automated, privacy-protecting pipeline to contrast usage across three populations.”
Where else this leads
Same territory
- Learning more about Claude's mathematical capabilities Anthropic
agentic ai - Project Pilot: Can AI control a drone? Anthropic
agentic ai - Gram: Assessing sabotage propensities via automated alignment auditing Google DeepMind
agentic ai - Did US Worker Retraining Reduce Participant Automation Exposure? Google DeepMind
labor economics
Published alongside it
The nearest publications in time, across all three labs.
- Formalizing Fermat's Last Theorem Anthropic
2026-09-04 - Designing Proactive Thought Partners for Writing Google DeepMind
2026-09-01 - Visual General Intelligence: A White Paper Google DeepMind
2026-08-26 - How Claude is accelerating protein design and analytical chemistry Anthropic
2026-08-18
What this page was built from
The manifest row for pid P01 gives title 'Research acceleration: The view inside OpenAI', pub_date 2026-09-06, and a source_url for that OpenAI blog post, but texts/P01.txt contains a different, unrelated OpenAI research paper — 'The Shift to Agentic AI: Evidence from Codex' (Johnston, Holtz, Ong, Tambe, Richmond, Chatterji) — which contains zero occurrences of 'research acceleration' and no stated publication date; this extraction is built entirely from that actual text (title cleaned of PDF small-caps letter-spacing artifacts), and pub_date/source_url/pdf_or_paper_url are set to null rather than carrying over the mismatched manifest values. The text itself appears to be the complete working paper including all appendices (A-F) and references.