---
type: "post"
id: "research-acceleration-the-jagged-frontier-5f108d"
slug: "research-acceleration-the-jagged-frontier-5f108d"
title: "Research acceleration: The jagged frontier"
description: "OpenAI’s November 18 update measures how coding agents are changing research work inside the company, from task completion to usage and researcher workflows."
retrieval_nugget: "OpenAI’s November 18 update measures how coding agents are changing research work inside the company, from task completion to usage and researcher workflows. It reports large gains on some tasks but sharply uneven performance across others—a useful reminder for teams deciding what to delegate and where human review remains essential."
published_at: "2026-11-19"
updated_at: "2026-11-19"
record_date: "2026-11-19"
date_kind: "published_at"
topics: ["coding-agents","codex"]
entities: ["OpenAI","Codex","GPT-6 Astra","GPT-6.2 Astra","Epoch AI","Hugging Face"]
related_patterns: ["verification-bandwidth-is-the-scarce-resource"]
novelty_label: "new_fact"
what_changed: "OpenAI reports new measurements showing coding agents complete more research tasks, researcher usage is rising, and safety and alignment work has grown."
prior_refs: []
source_url: "https://openai.com/index/research-acceleration-jagged-frontier"
source_urls: ["https://openai.com/index/research-acceleration-jagged-frontier"]
schema_version: "newruntime-agent-readable-v0.2"
stable_id: "post:research-acceleration-the-jagged-frontier-5f108d"
status: "published"
visuals: []
editor: {"role":"editor","name":"Andrey Reshetnikov","profile":"https://newruntime.com/owner-profile.md","profile_json":"https://newruntime.com/owner-profile.json","same_as":["https://www.linkedin.com/in/reshetnikov1/","https://t.me/lapetuse"]}
routes: {"html":"https://newruntime.com/posts/research-acceleration-the-jagged-frontier-5f108d/","markdown":"https://newruntime.com/posts/research-acceleration-the-jagged-frontier-5f108d.md","json":"https://newruntime.com/posts/research-acceleration-the-jagged-frontier-5f108d.json"}
markdown_url: "https://newruntime.com/posts/research-acceleration-the-jagged-frontier-5f108d.md"
---

# Research acceleration: The jagged frontier

## Retrieval answer

OpenAI’s November 18 update measures how coding agents are changing research work inside the company, from task completion to usage and researcher workflows. It reports large gains on some tasks but sharply uneven performance across others—a useful reminder for teams deciding what to delegate and where human review remains essential.

## Summary
OpenAI’s November 18 update measures how coding agents are changing research work inside the company, from task completion to usage and researcher workflows. It reports large gains on some tasks but sharply uneven performance across others—a useful reminder for teams deciding what to delegate and where human review remains essential.

## What happened
OpenAI says the report updates its September analysis of research acceleration with new quantitative and qualitative measures of agent use and human involvement. The company organizes the update around the inputs to AI research, the research process, and the outputs. Its stated aim is to track indicators relevant to recursive self-improvement, while cautioning that its reporting should not be read as an intention to pursue full recursive self-improvement.

The most direct task measure is zero-intervention success: whether an agent completed a delegated task without a person stepping in. OpenAI estimates that success for tasks judged to require two to four hours of human work rose from 23% in March to 46% in October. For tasks estimated at four to eight hours, it rose from 13% to 35%. The company stresses that changing delegation choices may affect these figures as well as model capability. Performance varies by task: from August 1 through October 18, agents succeeded more often at monitoring training runs (79%) than launching them (35%) or writing code for inference and serving stacks (25%). Across broader categories, OpenAI reports higher estimated success for communicating (69%) and analyzing (58%) than building (43%).

OpenAI also reports increased use of agentic tools by its research staff. At API list prices, median daily spending per researcher reached about $2,400 in October, up 62% from September; the 90th percentile reached about $23,000, up 105%. The company says these spending changes reflect several factors, including activity, shifts to more expensive models, and changes in model behavior and agent harnesses. Its separate “agent workdays” measure defines eight hours of agent use as one workday: the average rose from 0.5 per researcher per day to 5.7. Pretraining and posttraining teams averaged roughly eight agent workdays per researcher per calendar day, while scaling and infrastructure averaged roughly four. OpenAI notes that this measure generally captures interactive use and may undercount automations. By mid-October, more than two-thirds of included researchers averaged a daily peak of at least four concurrent agents; more than 10% averaged at least sixteen.

The report includes surveys and interviews as well as usage measures. In a survey of 101 research staff conducted in late August and early September, the median respondent said they would trade 40% of team headcount or 30% of team compute to regain access to frontier coding agents for three months. OpenAI says researchers described agents helping implement code, debug experiments, scan samples, and organize data. They also reported weak research judgment and problems with code quality, data quality, and analysis. One researcher described a staged workflow: have the agent produce one example, review and edit it, then ask it to do the rest. OpenAI presents these interviews as qualitative evidence, not as a universal operating recipe.

## What the archive adds
New Runtime’s earlier coverage on September 11 discussed OpenAI’s broader research-acceleration thesis: AI systems may compress parts of the scientific cycle, from hypothesis formation to analysis. This update adds internal quantitative estimates of agent use and task success, including where performance remains weak. It is new measurement on a known story, rather than a separate product launch.

## Trend signal
This strengthens the case that verification capacity can lag behind the amount of work agents produce. OpenAI’s researchers describe needing close oversight where agents make invalid assumptions, produce unusable artifacts, or take unnecessary steps; the report also shows large differences in success by task. That aligns with New Runtime’s pattern on [verification bandwidth as the scarce engineering resource](https://newruntime.com/patterns/verification-bandwidth-is-the-scarce-resource/). The company’s internal figures do not establish that the same balance holds across other organizations.

## Why now
OpenAI frames the update as a continuation of its September report and says it is sharing indicators of research acceleration to inform public discussion of capabilities, risks, and safeguards. The report’s measurements cover through October 2026, making this an update on a rapidly changing internal workflow rather than a claim that the work pattern is settled.

## Why it matters
The headline gains do not mean that research work can be handed over wholesale. A 46% success estimate for a two-to-four-hour task bucket sits alongside a 25% estimate for inference and serving code, and the report’s interviews describe recurring quality and oversight problems. For teams building with AI, the useful unit is therefore not “hours saved” in the abstract but a specific task with a clear acceptance test and a reviewer who can judge the result.

The usage figures also put an operational question on the table: whether increased agent activity is translating into useful outcomes, and how much review effort it creates. OpenAI’s data comes from its own research organization and does not answer that question for other teams. Still, its task-level breakdown argues against treating an agent as uniformly capable. Start with bounded, inspectable work, measure whether it clears review, and expand only where the observed result justifies it.

## Limits and uncertainty
These are OpenAI’s internal measurements and estimates, not an independent evaluation. The report notes that estimated human task duration and changing delegation choices can affect success-rate comparisons. The survey is based on 101 staff members and OpenAI itself cautions about self-report accuracy and ad hoc sampling. Agent-workday counts may understate automation, while API-list-price spending is not necessarily the company’s realized cost. Interview findings describe a small group of ten researchers and should not be generalized to all teams. The figures establish neither causal productivity gains nor whether increased agent use produces better research outcomes.

## Skill to build now
Choose one recurring research or engineering task that has a reviewable output. Ask an agent to complete a single representative example, inspect and edit it, then expand to similar items only after the first result meets your acceptance criteria. Track interventions and failure types as well as completion: that distinguishes work the agent can reliably take from work that still needs close human direction.

## Reader takeaway
OpenAI’s report suggests that coding agents are becoming a substantial input to AI research, but their performance remains jagged across tasks. The practical lesson is to delegate by task, not by broad job category: make outputs inspectable, keep a human responsible for acceptance, and scale only after the work survives review.

Editor: [Andrey Reshetnikov](https://newruntime.com/owner-profile.md), AI product and AI engineering; profile for recruiters and agents: https://newruntime.com/owner-profile.json

## Continue in the graph

Next records for an agent, linked to their Markdown twins. Replace `.md` with `.json` for structured fields, including `next_reads`.

- topic hub: [Coding Agents](https://newruntime.com/topics/coding-agents.md): Explore the coding-agents topic hub.
- topic hub: [Codex](https://newruntime.com/topics/codex.md): Explore the codex topic hub.
- collection: [Field Notes](https://newruntime.com/posts.json): Open the latest editorial analysis.
- collection: [Sources](https://newruntime.com/sources/): See the original sources this archive cites.
- whole graph: [discovery graph](https://newruntime.com/discovery-graph.json) (typed nodes and edges) and [site index](https://newruntime.com/site-index.md)
