THE SIGNAL IN ONE SENTENCE
OpenAI has published a much fuller look at how coding agents are changing research inside its own laboratory. The headline number is hard to ignore. OpenAI estimates that agents completed 46 percent of delegated research tasks requiring two to four hours of human work without human intervention in October. The comparable estimate was 23 percent in March. For tasks estimated at four to eight human hours, the figure rose from 13 percent to 35 percent. That does not mean an agent now does half of AI research. It means a changing set of delegated tasks, grouped by estimated human duration, received a changing estimated success rate inside one company. OpenAI says the mix of work may have changed as researchers learned what to delegate. The report also finds a very jagged frontier. Agents were estimated to succeed without intervention on 79 percent of tasks involving monitoring training runs for problems and 72 percent of technical-question tasks. The estimate fell to 35 percent for launching training runs, 29 percent for building evaluation, reinforcement-learning and agent environments, and 25 percent for writing inference and serving code. At the broader category level, OpenAI reports estimated success rates of 69 percent for communication tasks, 58 percent for analysis and 43 percent for building things during the August 1 through October 18 measurement window. The weak category can govern the whole system. A laboratory does not ship a pile of completed subtasks. It turns ideas, people, data, compute, software, experiments, review and safety work into knowledge and, eventually, model capabilities. If an agent writes monitoring code quickly but a researcher still has to find the right experiment, correct the data, secure scarce compute, diagnose a bad assumption and decide whether the result is real, the task-level gain is useful. It is not the same as laboratory acceleration. The plain signal is simple: measure the whole lab. OpenAI's report deserves credit for moving beyond a single productivity percentage. It separates inputs, process and outputs. It describes limitations. It reports uneven performance. It includes evidence that people still originate and judge many of the important ideas. The report also shows why a complete measurement system is difficult. Agent use has exploded. OpenAI defines eight hours of agent runtime as one agent workday. By that measure, average use rose from 0.5 agent workdays per researcher per day to 5.7. Researchers on pretraining and posttraining teams averaged roughly eight agent workdays per calendar day, while scaling and infrastructure averaged roughly four. Concurrency helps explain the apparent time travel. By mid-October, more than two-thirds of included researchers averaged a daily peak of at least four agents running at once. More than 10 percent averaged a peak of at least sixteen. The useful unit is not agent hours alone. A machine can spend eight hours producing a result, repeating a bad branch or patiently setting compute on fire. Runtime measures activity. It does not establish value. The spending figures need the same treatment. At public API list prices, OpenAI values the median researcher's daily agent use at about $2,400 in October, up from $1,500 in September. At the 90th percentile, the figure reached about $23,000, up from $11,000. Those are API-equivalent values, not OpenAI's marginal internal cost. The report says internal Codex, ChatGPT and API use together account for only 2 percent of total research compute. It also says a temporary internal harness change sharply reduced recorded token use before being reversed. Price can be a useful common ruler. It should not be confused with a bill. Start the lab scorecard with accepted tasks. For every delegated task, record the task family, estimated human duration, agent runtime, model and harness, tools, context, number of retries, human interventions and final acceptance. Define acceptance before the run. Code should pass the relevant tests and review. Analysis should survive reproduction. Data should meet quality checks. A training launch should use the intended configuration and produce a usable run. Separate first-pass success from success after repair. A task that needed forty minutes of human correction is not zero-intervention success. It may still be a strong economic win, but it belongs in a different bucket. Keep failures. The failed branch is part of the cost. Record discarded code, dead experiments, duplicated work, unusable artifacts, invalid assumptions and compute consumed before somebody noticed. OpenAI's interviews describe agents wandering into unnecessary steps, generating low-quality pull requests and making invalid analytical assumptions. That is not an argument against agents. It is an argument for an honest denominator. Next, measure idea formation. OpenAI asked small groups of key contributors where important ideas for GPT-6 Astra and GPT-6.2 Astra came from. Posttraining and safety-monitoring researchers attributed the named ideas to people, with models helping implement them. Pretraining researchers also attributed the earlier Astra ideas to people. For GPT-6.2 Astra, some pretraining researchers described developing ideas through conversation with Codex, including suggestions that helped an idea work. That is a more interesting picture than human versus machine. Track who proposed the hypothesis, what evidence motivated it, how alternatives were compared, who rejected weak ideas and what role an agent played in sharpening or implementing the surviving one. Count novel ideas that reach a serious test, not the total number of suggestions emitted. Idea volume is cheap. Research taste is expensive. Then measure experiment selection. An agent can make experiments cheaper enough that a team runs more of them. OpenAI reports 1.68 times as many logged experiments per active experimenter during the four weeks ending October 18 as the 2025 baseline. More experiments can mean faster learning. They can also mean more low-value branches. For every experiment, link the question, expected information gain, compute budget, stopping rule, result and decision. Record whether the experiment changed a belief, killed a bad path, improved a model or merely created another chart. Compare the portfolio with a counterfactual: what would the team have run without the agent, and what valuable work was displaced? That counterfactual will never be perfect. Write it down anyway. Without it, increased activity quietly becomes proof of increased progress. Measure researcher attention as a scarce resource. Agents can free a researcher from typing while increasing the number of decisions arriving at the desk. Track planning time, prompt and context preparation, review, debugging, experiment triage, coordination, interruption and recovery. Measure queue depth and wait time for human decisions. A team with sixteen concurrent agents may have enormous execution capacity and one exhausted reviewer standing at the merge gate. Review quality matters too. Sample accepted work for hidden errors. Use independent reproduction where the result is consequential. Rotate reviewers so a single researcher's habits do not become the evaluation standard. Track how often an accepted artifact later causes a failed run, misleading analysis or rollback. Now add the physical bottlenecks. Research moves through data preparation, access approvals, compute scheduling, training, monitoring, evaluation, interpretation and release gates. Measure utilization and wait time at each stage. If agents accelerate code creation but the accelerator cluster, data pipeline or evaluation team becomes the queue, the lab has moved its bottleneck rather than removed it. This is why OpenAI's task breakdown matters more than one average. Monitoring runs appears much easier for agents than launching them. Writing inference and serving code is much harder than answering technical questions. A system-level scorecard should preserve those differences. Publish the bottleneck map by task family, team and month. Do not average the jaggedness into a comforting blob. Safety work belongs inside the productivity accounting. OpenAI says documented researcher workweeks devoted to Safety and Alignment increased 77 percent from January through September. Its method classifies contributions from company-public Slack posts, pull requests and internal documents, then divides a researcher's week across verified projects. OpenAI calls this a proxy and notes that category boundaries may change. The increase is still important. Faster capability work can create more evaluation, monitoring and governance work. A lab should report both sides. Track safety experiments, monitoring coverage, discovered failure modes, mitigations, unresolved risks and the human time required to validate them. Do not treat safety labor as overhead that reduces an otherwise pure productivity number. It is part of producing a system responsibly. Finally, measure outputs. Lines of code and experiments are process measures. OpenAI says so. Its report adds an internal capability index modeled on Epoch's methodology, combining public and internal benchmarks. OpenAI reports a frontier trend of 16.4 index points per year since August 2025, compared with Epoch's reported 14-point trend since late 2024. That comparison does not establish that agents caused the difference. Model capability changes can come from more compute, better data, algorithmic ideas, improved evaluations, staffing, infrastructure and many interactions among them. An internal benchmark mixture can also favor the models built by the organization that designed it. OpenAI notes that its models score relatively better on its internal index than on Epoch's public index. So link outputs to inputs cautiously. Publish capability changes, safety changes, research-cycle time, compute, staff time, agent use and major methodological changes on the same timeline. Mark causal claims separately from correlations. When the harness, model, task mix or benchmark changes, preserve the old series and annotate the break. The best lab ledger has at least nine columns: Task acceptance. Idea quality. Experiment selection. Compute consumed. Failed branches. Human attention. Safety work. Bottleneck delay. Capability output. Add a tenth for uncertainty. OpenAI's report is internal measurement published by the organization building and using the systems. Its task-success estimates, work classifications, surveys, interviews, internal capability index and API-equivalent spending each carry different limitations. The survey of 101 research staff used an ad hoc sample and hypothetical tradeoffs. A separate late-October question drew 28 responses. The delegated task mix changed. Historical trends may not predict future progress. None of that makes the report useless. It makes it a serious first draft of the measurement problem. The wrong response is to flatten the report into a slogan that agents now do half of research. The better response is to adopt the structure, preserve the caveats and demand comparable reporting across laboratories. How many tasks were accepted without repair? Which task families remain weak? How much compute reached failed branches? Who generated and rejected the important ideas? Where did the bottleneck move? How much human time went into verification and safety? Did the resulting systems become more capable, more reliable and more governable? Coding agents are clearly becoming meaningful research infrastructure inside OpenAI. Nearly half of a particular short-task bucket is a milestone. The laboratory is the measurement.
01
WHAT ACTUALLY CHANGED
OpenAI published an expanded internal report on coding-agent use, research process and capability outputs.
Estimated zero-intervention success for delegated two-to-four-hour tasks rose from 23 percent in March to 46 percent in October.
Estimated success for four-to-eight-hour tasks rose from 13 percent to 35 percent over the same period.
Task performance remained uneven, from 79 percent for monitoring training runs to 25 percent for writing inference and serving code.
Average agent runtime rose from 0.5 to 5.7 eight-hour agent workdays per researcher per calendar day.
More than two-thirds of included researchers reached a daily peak of at least four concurrent agents by mid-October.
OpenAI reported 1.68 times as many logged experiments per active experimenter as the 2025 baseline during the four weeks ending October 18.
Documented Safety and Alignment workweeks increased 77 percent from January through September.
OpenAI introduced an internal capability index alongside code and experiment activity measures.
The report says changing delegation choices, internal methods and historical limits prevent simple causal or predictive conclusions.
02
WHY THIS MATTERS
A completed coding task is only one step in a research system that also depends on ideas, data, compute, experiments, review and safety work.
Uneven task performance can move bottlenecks rather than eliminate them.
Concurrent agents can increase execution capacity while overwhelming human review and decision gates.
More experiments do not automatically mean more knowledge or better capability.
API list-price equivalents do not represent OpenAI internal marginal costs.
Activity measures cannot by themselves establish that agents caused faster capability progress.
Safety and verification work must be counted as part of responsible research output.
Comparable public measurement could improve governance of increasingly capable research agents.
03
WHERE IT COULD HELP
- Define acceptance criteria before delegating each research task.
- Separate first-pass success, repaired success, rejection and unsafe output.
- Record model, harness, tools, context, runtime, retries and human interventions.
- Keep failed branches and compute consumed in the cost ledger.
- Track who proposed, refined, tested and rejected each research idea.
- Measure experiment information gain and decisions changed, not just experiment count.
- Write down the counterfactual portfolio the team would have run without agents.
- Measure planning, review, debugging, triage and recovery time for researchers.
- Sample accepted work for hidden errors and independently reproduce consequential results.
- Map wait time and utilization across data, compute, training, evaluation and release gates.
- Report task-family results instead of one blended productivity percentage.
- Track safety findings, mitigations, unresolved risks and validation work beside capability progress.
- Annotate model, harness, benchmark and methodology changes on every time series.
- Separate observed correlations from causal claims and forecasts.
KEEP A HAND ON THE WHEEL
The report describes OpenAI's internal work and is not an independent audit. Delegated task types may change as researchers learn where agents work well, so success-rate trends do not isolate model capability improvement. Task duration is an estimated human equivalent, not measured saved labor. Zero-intervention task completion does not equal end-to-end research completion. API list-price equivalents are comparison values, not OpenAI's internal marginal cost or cash spending. The research-area analysis uses classified Slack posts, pull requests and internal documents as a proxy for attention. OpenAI says category boundaries may change. Survey results are self-reported, hypothetical and based on ad hoc sampling. Interviews cover ten researchers. More code, experiments and agent runtime are activity measures. They do not prove better research or faster capability progress. OpenAI's internal capability index mixes public and internal benchmarks and scores OpenAI models relatively better than the public Epoch index. Historical results do not establish recursive self-improvement and may not predict future progress.
04
TERMS WORTH KEEPING
OPEN GLOSSARY CARD
Calibration
Measuring and adjusting a system so its controls and readouts match how the hardware actually behaves.
OPEN GLOSSARY CARD
Benchmark
A fixed test used to compare how systems perform on the same tasks.
OPEN GLOSSARY CARD
Audit trail
A durable record of actions, changes, identities, and times that lets someone reconstruct what happened.
SOURCES AND VERIFICATION STATUS
This article was written from the materials below. Product claims and dates were checked against those sources on November 19, 2026.
THE PUBLICATION ENGINE
WANT A SIGNAL OF YOUR OWN?
We build source-grounded publications, private briefings, and editorial systems for organizations with something useful to say.
WORK WITH US