Research acceleration: The view inside OpenAI

For AGI to benefit all of humanity, we believe it must be democratically governed.

This vision necessitates a transparent, informed public discourse regarding the safeguards, risks, and capabilities of frontier AI. To ensure the public has a meaningful voice in the trajectory of AI development, transparency must go beyond merely reporting incidents or specific risks; it must include an inside look at how the most advanced systems are being developed and how they are accelerating the research process within labs.

The Roadmap to Automated Research

Our objective is the creation of an automated AI researcher. This system, operating under human oversight, is designed to drive iterative progress in alignment and deep learning.

Milestones & Progress

We have established a clear timeline for these capabilities:

• Automated Research Intern (Achieved Sept 2026): A system capable of executing well-defined research tasks under human guidance—tasks that would typically require several days from a skilled human researcher.
• Automated AI Researcher (Target: March 2028): A more autonomous and capable system.

Manual research bottlenecks $\rightarrow$ Agentic acceleration.

Impact on Internal Workflows

The daily operations of OpenAI researchers have undergone a fundamental shift this year. The adoption of coding agents is growing at a rate that exceeds all other teams within the company.

Key Observations:

• Velocity: Researchers are producing code and executing experiments at a significantly higher frequency.
• Complexity: Agents are successfully tackling increasingly sophisticated tasks.
• Nuance: While these metrics are positive, the overall pace of AI progress may not scale linearly due to various systemic bottlenecks.

Despite this, the internal consensus is clear: agentic tools are meaningfully accelerating the speed of research.

Resource Consumption Metrics

The scale of agent usage is best illustrated by the inference costs (at API pricing) incurred by researchers:

In mathematical terms, if $C$ represents the daily cost of tokens, we see a distribution where: $\text{Median}(C) > 600 \quad \text{and} \quad P_{90}(C) > 7,000$

The Balance of Control and Safety

It is important to note that humans remain the primary decision-makers. People still determine:

1. Research priorities.
2. Which results to pursue.
3. Whether to scale, pause, or deploy a system.

Why pursue automated research?

We believe that, if handled responsibly, this path will:

• Lower the cost of high-level intelligence for global benefit.
• Help solve the alignment problem.
• Create defenses against malicious AI agents.
• Secure critical infrastructure.

The Risks of Recursive Self-Improvement (RSI)

While the benefits are clear, we are not blindly pursuing rapid RSI. We do not yet possess a proven method to safely reach a state of fully aligned, full RSI.

def safety_check(system_capability, alignment_level):
    if system_capability > alignment_level:
        return "STOP_DEPLOYMENT"
    return "PROCEED_WITH_CAUTION"

We recognize that as systems become more capable, they may become harder to monitor. Therefore, we are scaling our safety measures in tandem with our capabilities.

Commitment to Rigor and Transparency

We will slow or stop development if we encounter an unacceptable safety risk. A recent example of this commitment occurred following the Hugging Face incident.

Our Response:

• Paused reinforcement learning (RL) training for deployment-ready models.
• Hardened research environments through expanded red-teaming.
• Enhanced monitoring system coverage.

While some workloads resumed under stricter controls, we have now integrated safety requirements deeper into the model lifecycle, demanding rigorous evidence of alignment throughout the training process.

Moving Toward Industry Standards

We are sharing these preliminary measurements to encourage a culture of public disclosure and to help establish shared industry standards. As outlined in our frontier policy blueprint, we believe companies should be mandated to track their progress toward RSI publicly.

OpenAI intends to remain transparent about RSI progress, evolving our methods as our understanding grows, while carefully protecting proprietary and security-sensitive information.