OpenAI coding agents now finish 46% of 2-to-4-hour research tasks
The share of multi-hour research tasks OpenAI's agents complete without help doubled since March, and the 4-to-8-hour jump is the number boards should watch.

OpenAI says its coding agents now complete 46% of two-to-four-hour research tasks without human help, up from 23% in March, and 35% of four-to-eight-hour tasks, up from 13%. The jump matters because it moves autonomous AI from quick assistance into half-day project work, forcing executives to rethink staffing, verification, and competitive speed.
OpenAI says its coding agents now complete 46% of research tasks that would take a person two to four hours, with no human help. That is not a marginal improvement. In March, the same figure was 23%, meaning the share of multi-hour research work these agents can finish on their own has roughly doubled in a matter of months. The claim comes from an OpenAI report titled "Research acceleration: The jagged frontier," and it lands squarely in the middle of the debate over how quickly AI moves from assistant to autonomous worker.
The longer-horizon number is just as striking. For tasks that would take a human four to eight hours, OpenAI says agent success rose from 13% to 35%. That jump matters because four-to-eight-hour tasks are not quick lookups or single-file edits. They are the kind of work that typically involves scoping a question, gathering sources, testing hypotheses, and producing a coherent answer. When an agent can complete roughly one in three of those tasks without help, the unit of automation shifts from minutes to half-days.
OpenAI's disclosure is a benchmark claim, not an audited productivity study. The report title itself signals uneven progress: a jagged frontier. In practice, that means agents may be strong on some research tasks and weak on others, even when the tasks look similar from the outside. For executives, the lesson is not that AI can replace a research team tomorrow. It is that the boundary of what can be delegated is moving fast enough to change how teams are staffed, how projects are scoped, and how quickly competitors can respond to new questions.
Coding agents are a specific category. They are AI systems built to plan, write, debug, and iterate on code, but OpenAI is describing their performance on research tasks. That overlap is important. Research and coding both reward iterative problem-solving: form a hypothesis, run an experiment, inspect the result, revise. If agents are getting better at that loop, the productivity gain is not limited to software engineers. It can spread to analysts, product managers, data teams, and anyone whose job involves turning an open question into a structured answer.
The March-to-now comparison is the part that should make boards pay attention. A rise from 23% to 46% on two-to-four-hour tasks is a doubling in a short window. A rise from 13% to 35% on four-to-eight-hour tasks is nearly a tripling. Those are the kinds of curves that look small in a lab and disruptive in a P&L. If the trend continues, the bottleneck moves from execution to verification. Someone still has to check the agent's work, catch fabricated sources, and decide whether the output is good enough to ship.
That verification layer is where the second-order costs appear. Even if agents complete more tasks without human help, companies will need review processes, audit trails, and clear ownership for AI-generated research. Regulated industries will face questions about accountability: if an agent produces a flawed analysis, who signs off? If a client receives incorrect advice, who is responsible? The source does not answer those questions, but the performance numbers make them urgent. The faster agents improve, the less time organizations have to design the controls around them.
For founders and operators, the competitive implication is straightforward. If one team can point an agent at a four-to-eight-hour research task and get a usable draft in minutes, while another team spends a day on the same task, the first team compounds an advantage. That does not mean headcount cuts are automatic. It means the marginal cost of exploring an idea falls, which can lead to more experiments, faster iteration, and a higher bar for what counts as a full day's work. The teams that win may be the ones that redesign workflows around agent strengths rather than bolting AI onto old processes.
The strategic stakes are clearest for leaders who allocate capital and talent. OpenAI's numbers are a signal that autonomous research is moving from demo to deployment, even if the frontier remains jagged. The right response is not panic or blind adoption. It is to map which tasks are already within reach, measure where agents fail, and build the review capacity to trust the output. The companies that treat this as a workflow question, not a tooling question, will be the ones that turn a benchmark into an edge.
This story's Key Insights and Take-aways are locked.
Create a free account to unlock Executive Actions for one credit.
Register to UnlockAlways free for Executives Club members. Join the Club
More in Technology
Anthropic Mathematician Claims Claude Helped Prove Riemann Hypothesis in Six Weeks
Levent Alpöge says the AI tool helped him develop a solution to one of math's most important unsolved problems in about six weeks, a claim that, if verified, would reshape AI's role in research.
Anthropic mathematician claims Claude helped crack the Riemann hypothesis
If the proof holds, AI has toppled math's most famous open problem; if it does not, it is a case study in announcing breakthroughs before anyone verifies them.
AI's $4 Trillion Wipeout Erases Nasdaq's Gains for the Year
A brutal week took roughly $4 trillion off the AI sector and left the Nasdaq flat for the year, and the damage ran the full length of the supply chain.


