What is research acceleration?

Research acceleration is the measurable increase in output speed a team gains when AI agents take on part of its work. OpenAI defined it in practice on 6 September 2026 by publishing agent usage, code contribution and experiment volume figures from inside its own research organisation.

What did OpenAI report about its coding agent usage?

OpenAI reported that by mid-August 2026 the median researcher used more than $600 per day of inference at API prices, the 90th percentile user exceeded $7,000 of tokens per day, and the research organisation used 3.1 agent-workdays of effort for every human workday.

Which marketing tasks benefit most from AI agents?

Well-scoped, high-frequency, verifiable tasks benefit most. Technical SEO fixes, schema deployment, feed hygiene, log-file analysis and variant production all have a clear pass or fail check. Ambiguous strategic work, such as positioning or budget allocation, still depends on human judgement and priority setting.

How should CMOs measure AI agent value?

Measure completed work units and cycle time per unit, not licences or prompt counts. Baseline three repeatable tasks for four weeks without agents, then repeat with them, and report the ratio. OpenAI notes that other bottlenecks may still cap overall progress.

Did OpenAI say agents replace researchers?

No. OpenAI states that people still set research priorities, judge which ideas and results to pursue, and decide whether to scale, pause or deploy systems. It describes its current system as an automated research intern working on well-defined tasks under human direction.

All insights

AI in Marketing

3 min read7 September 2026Nathan Mzumara

Research Acceleration: OpenAI's Agent Metrics Marketers Should Copy

What OpenAI disclosed about research acceleration

Research acceleration is the measurable speed-up a team gets when AI agents do part of its work. On 6 September 2026, OpenAI published Research acceleration: The view inside OpenAI, a snapshot of its own agent telemetry. It is the first time a frontier lab has shown the numbers behind the claim that AI makes teams faster.

The figures OpenAI reported

Measure	Value	Period
Median researcher agent use	More than $600 per day of inference at API prices	Mid-August 2026
90th percentile user	More than $7,000 of tokens per day	Mid-August 2026
Agent effort vs human effort	3.1 agent-workdays per human workday	Mid-August 2026
Crossover point	Agent runtime passed total human labour	After June 2026

OpenAI states that researchers are contributing code faster and running more experiments, and that agents are handling more complex tasks and succeeding more often. All figures above come from OpenAI's own post.

Why seat counts are the wrong marketing metric

Most CMOs still report AI adoption as licences bought and prompts sent. OpenAI reported something else: effort delivered and work completed, measured against human workdays. That is the model to copy. Track completed work units and cycle time per unit, then compare the two against last quarter.

OpenAI also warns that overall progress will not keep pace with these specific metrics, because research has many bottlenecks. Apply the same caution to marketing. A faster build stage does not shorten a campaign if approvals are the constraint.

Where agent gains actually concentrate

OpenAI's account points to well-defined tasks under human direction. In marketing, the equivalent work is well scoped and verifiable:

Infographic on OpenAI's research acceleration data: 3.1 agent-workdays per human workday, $600 median daily agent spend, and advice to measure completed work not seats. — Infographic summary of this articleDownload infographic

Technical SEO fixes: crawl errors, redirects and internal linking, where success is checkable against a log or a crawl.
Schema deployment: repeatable markup with a validator that gives a pass or fail.
Feed hygiene: product data cleaning, where the error count before and after is the result.
Variant production: ad and landing page variants judged by a live test, not an opinion.

Ambiguous strategic work does not behave this way. Positioning, brand and budget calls still need people to set priorities, exactly as OpenAI says its researchers do.

The bottleneck moves to specification and verification

If agents remove execution capacity as the constraint, the new constraint is the quality of your brief and your check. Most in-house SEO and GEO teams are weakest here. Build the verification loop first, as we argued in our piece on provider risk in agentic coding tools.

Safety is part of the same discipline. OpenAI paused reinforcement learning training after the Hugging Face incident, described in its post on pacing model development. Slowing down when a control fails is a legitimate operating choice.

How to measure research acceleration in your team this quarter

Pick three repeatable tasks. Record cycle time and completion rate for four weeks without agents, then four weeks with them. Report the ratio, not the seat count. For causal rigour on spend, pair it with the approach in Google Meridian GeoX and incrementality testing. Research and development acceleration is now something you can evidence, and research acceleration measured this way answers the board question with a number instead of an anecdote.