OpenAI's Codex Agents Now Dominate Its Internal AI Token Use

OpenAI says Codex now generates 99.8% of its internal weekly output tokens, a reversal highlighted by a recent Google News item about enterprise AI spending.

That figure does not show spending, revenue, or productivity. It measures output tokens, the pieces of text generated by a model, inside the company selling the technology. Still, it captures a significant change in how OpenAI employees use AI.

ChatGPT once handled most internal AI work. Codex, an agent that can execute longer tasks and use tools, now dominates activity across engineering, legal, finance, and recruiting.

The important contest is no longer OpenAI against another model provider. It is chat-based assistance against delegated agent work. A chatbot answers a request, while an agent can inspect files, call tools, revise its approach, and operate for minutes or hours.

OpenAI presents that transition as evidence of productive work moving toward agents. The same transition also makes AI consumption harder to predict, attribute, secure, and evaluate.

Enterprise buyers therefore face a sharper question than the Google News headline suggests. They must determine whether growing token use represents valuable delegated work or an expensive loop that produces activity without an equivalent outcome.

What OpenAI's Token Shift Actually Shows

OpenAI's internal data shows a decisive behavioral shift, but it does not establish that every enterprise will achieve the same results.

OpenAI published its analysis on August 12, 2026. The company examined Codex activity among individual customers, organizational users, and its own workforce.

Through August 2025, the average OpenAI employee directed less than 10% of their tokens toward Codex. ChatGPT remained the default interface for internal AI use during that period.

The balance changed as Codex gained longer-running execution, better models, and support for parallel tasks. OpenAI says the average employee now produces more than 85% of output tokens through Codex.

Codex accounts for 99.8% of weekly output tokens across the company, according to OpenAI's agent adoption data. The gap between those two figures matters.

The first number describes the average employee's mix of Codex and ChatGPT activity. The second reflects the total pool of generated tokens, which becomes heavily influenced by intensive agent users.

Agents naturally generate more tokens than conventional chats. They repeatedly read context, plan actions, inspect results, call tools, and revise their work.

One employee can also run several agents at once. That makes output token share a measure of computational activity, not a direct count of workers or completed assignments.

OpenAI's heaviest internal users illustrate the difference. By June 2026, employees at the 99th percentile regularly produced more than 60 hours of Codex agent turns per day.

That number does not mean one person worked for 60 hours. It reflects multiple agents operating in parallel while their human operator directed or reviewed the work.

Longer tasks have also become common among sampled individual users. By May 2026, 80.6% had submitted at least one Codex request that OpenAI estimated represented over 30 minutes of human work.

A further 70.2% submitted a request estimated above one hour. Another 25.6% submitted at least one request estimated above eight hours.

These estimates concern the human effort associated with a task, not necessarily the agent's runtime or the amount of labor eliminated. OpenAI's classification method also cannot prove that every generated result was accepted or useful.

Even with those limitations, the direction is clear. Users are asking AI to handle larger units of work instead of requesting isolated answers.

A chat interaction might produce a summary, draft, or code snippet. An agent can examine a repository, modify several files, run tests, interpret failures, and attempt corrections.

The distinction explains why enterprise AI activity can grow much faster than employee headcount or message volume. Each delegated job contains several model interactions that remain hidden behind one user request.

This is the event's central reversal. The visible prompt is becoming a smaller part of the work, while autonomous execution consumes a larger share of the computation.

For organizations, that changes the meaning of AI adoption. Counting licenses and messages no longer describes how deeply models participate in operational work.

It also changes the risk boundary. A chatbot usually proposes an answer, while an agent can perform actions across connected systems.

That broader authority makes permission design, audit records, review gates, and cost controls part of deployment. They are no longer secondary administrative details.

Why Google News Signals Point Beyond Chatbots

The Google News story matters because OpenAI's data connects rising token demand to agents spreading beyond software engineering.

Codex began as a product associated with coding. Engineers were the earliest employees to move most of their OpenAI usage from ChatGPT to the agent.

The average OpenAI engineer crossed that threshold by December 2025. OpenAI now says the average engineer generates 99% of output tokens through Codex rather than ChatGPT.

The more revealing transition happened elsewhere. Legal, finance, and recruiting moved to majority Codex use around April 2026, after engineers had already established the pattern.

OpenAI reports that the average lawyer or recruiter now generates more than 85% of output tokens through Codex. That does not mean those employees spend most of their day programming.

Instead, Codex has become a general execution environment. Nontechnical users employ it for automation, structured analysis, data transformation, debugging, and internal tooling.

OpenAI's occupational breakdown reinforces that interpretation. More than one-quarter of Codex work generated by employees in business functions fell into engineering or coding categories.

Within finance and business operations, engineering or coding represented 31% of classified Codex output. Knowledge work represented 34%, while financial analysis accounted for 16%.

For product, marketing, and operations employees, knowledge work represented 51% of classified output. Engineering or coding accounted for another 25%.

These categories suggest that agents are lowering the practical barrier between requesting technical work and attempting it directly. They do not show that specialists are no longer required.

A recruiter might ask an agent to transform data from several sources into an internal workflow. A finance employee might have it reconcile files, build a repeatable analysis, or test a small tool.

The employee remains responsible for context, judgment, authorization, and review. However, the execution path can include code without requiring a conventional software project.

Nondeveloper adoption grew quickly across all three populations studied by OpenAI. From August 2025 to early June 2026, weekly nondeveloper users increased 137-fold among individual users.

OpenAI reported a 189-fold increase among organizational users and a 12-fold increase within its own workforce. Those multiples began from different baselines, so they should not be compared as equal adoption levels.

They nevertheless support the same directional finding. Agent use is expanding beyond the technical audience that first adopted Codex.

The pattern also fits OpenAI's broader enterprise reporting. Its enterprise usage study says ChatGPT message volume grew eightfold, while reasoning token consumption per API organization rose approximately 320-fold year over year.

That earlier study covered more than Codex. It included aggregated usage from OpenAI's enterprise products and a survey of 9,000 workers across almost 100 organizations.

OpenAI also reported more than one million business customers and over seven million workplace seats. Nearly 200 organizations had each processed more than one trillion tokens.

Together, these figures describe two adoption curves. Human-facing message volume is rising, but computation inside integrated and multi-step workflows is rising much faster.

That is why chat is becoming an incomplete unit for measuring enterprise AI. One message can initiate a long sequence of inference, retrieval, tool use, validation, and revision.

The shift pressures several groups at once. Finance teams must forecast variable consumption. Security teams must govern system access. Managers must decide which work deserves continued capacity.

Employees also face a new form of operational responsibility. Running many agents can look productive, but parallel activity is not the same as completed, accepted work.

Organizations will need shared evidence about what happened during an agent run. A searchable AI knowledge base can preserve relevant context, but it cannot replace permissions or formal evaluation.

The opportunity is genuine. Agents can help workers cross technical boundaries and complete tasks that previously stalled in another team's queue.

The management challenge is equally genuine. Enterprises must identify where that autonomy produces repeatable value and where it simply multiplies model calls.

Chat Assistance and Agent Execution Follow Different Economics

Agents change enterprise AI from a mostly user-paced service into a variable workload whose consumption depends on execution behavior.

Traditional chat places a natural limit on activity. A person asks a question, waits for an answer, evaluates it, and decides whether to continue.

Agent execution removes some of those pauses. The system can choose the next action, bring previous results back into context, and continue until it reaches a stopping condition.

Each cycle can add input tokens as the agent rereads instructions, files, tool responses, and its accumulated history. Output tokens capture only part of that total workload.

Long tasks also encounter branches. An agent may try one implementation, inspect an error, revise its plan, and run another test.

Those retries can be productive because real work rarely follows a perfect first plan. They can also become waste when the agent lacks the information, permission, or capability needed to finish.

This makes the cheapest model an incomplete answer to cost control. A less capable model might use fewer resources per call but require more attempts and more human correction.

OpenAI makes that argument in its AI investment guidance. It recommends measuring tasks completed, time saved, decisions improved, and workflows ready to scale.

The company says the price per million tokens fell 97% between GPT-4 and GPT-5.4. It also says GPT-5.6 used 54% fewer output tokens on a coding-agent index and finished tasks 57% faster.

Those are OpenAI's reported benchmark results, not a guarantee for each customer workload. A company's own context, tools, evaluation criteria, and failure costs can change the result.

Falling unit costs do not necessarily reduce total spending. Usage can expand faster than efficiency improves, particularly when agents run longer and operate concurrently.

That dynamic resembles cloud computing more than conventional software licensing. Demand depends on runtime behavior, workload design, model selection, context size, and repeated execution.

It also introduces a measurement problem. A high token count can indicate valuable automation, a difficult assignment, unnecessary context, repeated failure, or an agent that never stops efficiently.

Independent research gives buyers a reason to treat consumption cautiously. A 2026 paper examining coding tasks found that agentic workloads used far more tokens than code chat or code reasoning.

The researchers reported consumption differences reaching 1,000 times in their experimental setting. Runs of the same task varied by as much as 30 times.

More token use did not consistently produce greater accuracy. Performance often peaked at an intermediate level and then flattened as consumption increased.

The study also found that models struggled to predict their own token requirements. Reported correlations reached only 0.39, and estimates systematically understated actual use.

These findings come from a particular set of coding tasks and models. They should not be generalized into a universal multiplier for every enterprise agent.

They do demonstrate why a simple budget per prompt becomes unreliable. The same request can produce materially different execution paths and resource demands.

The enterprise response is already visible. OpenAI added a consolidated view of ChatGPT and Codex credit consumption to its Global Admin Console in June 2026.

Administrators can examine usage by user, product, and model. They can also set workspace defaults, group limits, and individual overrides through the company's spend control tools.

Databricks introduced similar controls in its Unity AI Gateway. The system can impose limits, detect runaway consumption, and recommend less expensive models for suitable work.

This emerging control layer reveals where the market is heading. Model access alone is becoming insufficient for enterprise deployment.

Organizations need attribution, policies, evaluation, and intervention around that access. They must connect consumption to a responsible team and a defined business process.

The strongest unit of measurement is therefore not the token. It is the completed outcome under agreed quality, risk, and review requirements.

For a coding agent, that might be an accepted change that passes tests and security review. For finance, it might be a reconciled report with traceable inputs.

For recruiting, it could be a workflow that reduces administrative effort without making unauthorized decisions about candidates. Each outcome requires different evidence.

This economic model favors workflows that repeat often and can be evaluated clearly. It penalizes vague tasks whose success depends on subjective impressions.

The agent shift does not make token measurement useless. It makes token counts one operational signal among several, rather than a proxy for business value.

What the Token Numbers Do Not Prove

OpenAI's dramatic internal adoption figures show demand and intensity, but they leave productivity, quality, and security outcomes unresolved.

The first limitation is selection. OpenAI is the developer of Codex, employs people comfortable with experimental AI, and can give them unusually direct product support.

Its workforce is not a representative sample of banks, hospitals, government agencies, manufacturers, or small businesses. Those organizations operate under different technical and regulatory constraints.

OpenAI also benefits when customers interpret higher usage as evidence of deeper adoption. Its data deserves attention, but readers should distinguish measured behavior from the company's broader prediction.

The second limitation concerns tokens as a success metric. Output volume shows that models generated text or code. It does not reveal how much users retained.

An agent can produce a large patch that a developer rejects. It can generate an analysis that an employee rewrites or create an automation that never reaches production.

OpenAI's long-horizon task estimates carry a similar caveat. A task classified as eight hours of human work does not automatically save eight hours.

The user might still spend time preparing inputs, monitoring the run, checking sources, resolving errors, and integrating the result. Some work may not have happened without the agent.

That expanded output can still be valuable. However, avoided labor is only one possible benefit, and it requires direct measurement.

The third limitation is reliability. Longer tasks create more opportunities for mistakes, stale assumptions, incorrect tool choices, or early errors that affect later steps.

Agents with access to enterprise systems also create a wider security boundary. A flawed answer is harmful, but an unauthorized action can be much harder to reverse.

Permissions should therefore match the task, not the maximum capability of the agent. Read access, write access, external communication, and irreversible actions deserve separate controls.

Security teams also need records that connect a human request to each tool action. Without that chain, incident response and accountability become difficult.

Cost monitoring can create another tension. Companies need enough visibility to identify waste and compromised credentials, but detailed employee surveillance can weaken trust.

An administrator may need to know that a workflow repeatedly calls an expensive model. They do not necessarily need unrestricted access to sensitive content.

Good governance separates operational metadata from business content whenever possible. It also defines who can inspect each layer and under which circumstances.

The fourth limitation involves generalization beyond coding. Codex can assist nondevelopers, but OpenAI's data still shows coding and engineering as major categories.

Structured technical work offers relatively clear validation. Tests can run, files can be compared, and errors can trigger another attempt.

Legal analysis, hiring support, strategy, and financial interpretation often contain more subjective requirements. Errors may survive longer because automated validation is weaker.

That difference makes outcome evaluation more important as agents move into business functions. Enterprises should not assume that adoption speed equals readiness.

Public sentiment has already shifted toward that distinction. An AI spending analysis published in July described growing resistance to "tokenmaxxing," the practice of treating high consumption as an achievement.

Moody's analyst Vincent Gusdorf warned that AI can make it easy to create work an organization does not need. That criticism goes directly to the weakness of token-based status signals.

Heavy consumption can be rational when an agent produces a valuable result. It becomes harder to defend when no owner can connect the activity to an accepted outcome.

The right skeptical position is not that agents are merely wasteful chatbots. OpenAI's adoption data is too substantial to dismiss that way.

The more defensible conclusion is narrower. Agents expand the amount and scope of work AI can attempt, while evidence about realized enterprise value remains uneven.

A company should therefore test OpenAI's thesis inside its own environment. That requires baselines, evaluation sets, review costs, failure rates, and measurable outcomes.

It should also compare an agent workflow against realistic alternatives. Those alternatives include human execution, conventional software, a shorter chat interaction, and a smaller model.

Only then can leaders identify whether rising token use represents leverage or friction. OpenAI's internal ratios cannot answer that question for them.

Three Signals Enterprise Buyers Should Watch Next

The next phase will be decided by outcome reporting, nontechnical retention, and security controls, not by another record token count.

The first signal is whether vendors connect agent consumption to completed business outcomes. Usage dashboards currently emphasize credits, users, models, and trends.

Those views help administrators find unexpected demand. They still stop short of showing whether a workflow produced an accepted result.

Buyers should look for reporting tied to approved code changes, resolved support cases, completed analyses, or other domain-specific outputs. Cost per accepted outcome would make comparisons more meaningful.

If vendors deliver credible outcome attribution, OpenAI's argument becomes stronger. Enterprises could fund high-consumption workflows when the resulting value is visible.

If reporting remains centered on tokens and estimated time, skepticism will grow. Organizations will struggle to separate intensive productive work from intensive failed work.

The second signal is sustained Codex use outside engineering. OpenAI reports striking growth among legal, finance, recruiting, marketing, and operations employees.

The next question is whether those users continue after early experimentation. Retention matters more than a rapid increase from a small starting base.

Enterprises should watch which nontechnical tasks become repeatable workflows. They should also track how often specialists must rescue or rebuild agent-generated work.

Continued use with stable quality would support OpenAI's claim that agents let employees cross functional boundaries. Falling retention would suggest that the strongest value remains concentrated in technical teams.

The third signal is the maturity of controls around agent action. Spend limits are arriving quickly, but cost represents only one part of operational risk.

Organizations also need task-level permissions, approval checkpoints, auditable tool calls, credential isolation, and reliable termination when behavior leaves defined boundaries.

Those safeguards must work across multiple models and tools. Enterprises rarely standardize every workflow on one vendor forever.

Progress here would make broader delegation easier to justify, especially in regulated industries. Weak or fragmented controls would slow adoption regardless of model capability.

These three signals are connected. Outcome reporting explains why an agent deserves resources. Retention shows whether employees find it useful after novelty fades.

Security and governance determine whether the organization can permit that agent to perform consequential work. Missing any one of the three can prevent a successful deployment.

The Google News headline captures a real shift from chat toward agents, but token share is the beginning of the story. It is not the final score.

OpenAI has shown that its employees increasingly delegate long and complex tasks to Codex. It has not shown that 99.8% of output tokens deliver an equivalent share of organizational value.

Enterprise buyers should ask a practical question before celebrating or restricting that consumption: which completed outcomes became possible, and what evidence proves they were worth the risk?

That question offers a better guide than token volume alone. It lets organizations expand workflows that work, constrain those that drift, and judge agent adoption by results rather than activity.