Today ainotis Join

My notis

ResearchPublished Lead storyThat day's edition

Running case: OpenAI's models 11 stories

Patwardhan says OpenAI models took on day-long research tasks

In OpenAI's own DevDay keynote, published 29 September, research lead Tejal Patwardhan gave OpenAI's measurement of its models on day long research tasks.

By notis. Written by machine, every claim checked against its source. How we work

Check our sources · 1 source, 9 claims
OpenAI DevDay 2026 Keynote (FULL): a video from OpenAI on YouTubeVideo: OpenAI on YouTube
watch on YouTube · Thumbnail shown to link to the video on YouTube

Key points

  1. In the keynote's research segment the host said "A few weeks ago we reached that goal." and "We now have an AI research intern." (23:52)
  2. Patwardhan, who introduced herself as "a research lead for post training", said OpenAI's models "have taken on over a third of day long research tasks with no intervention required" (26:27)
  3. The figure is OpenAI's own measurement of its own models, given as of July; before that, she said, they "failed the vast majority of times" on such tasks (26:27)

What happened

OpenAI published its DevDay 2026 keynote on its own YouTube channel on 29 September 2026. The video's description names Sam Altman, Romain Huet, Tejal Patwardhan and Holly Li as the presenters.

Introducing the research segment, the host said "Last year we maded a bold prediction saying we would see the first AI research intern in the text year" (23:34), the automatic captions' rendering of what appears to be next year.

He then said "A few weeks ago we reached that goal." (23:48) and "We now have an AI research intern." (23:52).

Tejal Patwardhan, who introduced herself as "a research lead for post training" (24:17), gave what is OpenAI's own measurement of its own models: they failed most day-long research tasks before, and "as of July and E-our models have taken on over a third of day long research tasks with no intervention required" (26:27).

"E-our" is the captions' rendering of a second month name that is not clear. Holly Li, presenting the dots agents, said of OpenAI's own engineers: "Our engineers have their dot fixing dozens of bugs every day." (15:34).

Altman also said "Around 1.2 billion people use ChatGPT." (44:11) and, on the API, "We've still kept more than 99 percent reliability." (31:06).

What it means for you

Our view

The one-third figure is OpenAI's own measurement of its own models, said on stage, and the claims we hold give neither the list of tasks nor the method behind it. Patwardhan said it failed the vast majority of times on tasks that took a day or longer, so the direction is plain but the size is OpenAI's to defend.

The host's "AI research intern" is a description of OpenAI's own work, not an outside verdict. If you are weighing AI for research work, treat this as a vendor claim until an independent test repeats it, and ask the vendor how the tasks were chosen and what "no intervention required" covered.

This part is our reading of the facts above. It adds no fact of its own.

Your reaction

One press adds one. We count a number for each notice and day, never who pressed it. The counts help us decide what goes in the monthly issue. If you are signed in to My notis, your own page counts yours too.

Check our sources

1 source, 9 claims. We opened the source and checked every sentence above against it.

1 OpenAI DevDay 2026 Keynote (FULL)OpenAI · 29 Sep 2026 · 9 claims Open the source

  1. OpenAI published the video "OpenAI DevDay 2026 Keynote (FULL)" on its own YouTube channel on 29 September 2026.

    OpenAI DevDay 2026 Keynote (FULL) | channel OpenAI | publishDate 2026-09-29T11:42:29-07:00

  2. The video description reads: "Watch Sam Altman, Romain Huet, Tejal Patwardhan, and Holly Li announce and demo 20+ launches at OpenAI DevDay, including dots, ChatGPT Spaces, GPT-6.1 Sol, Astra Ultrafast, and more."

    Watch Sam Altman, Romain Huet, Tejal Patwardhan, and Holly Li announce and demo 20+ launches at OpenAI DevDay, including dots, ChatGPT Spaces, GPT-6.1 Sol, Astra Ultrafast, and more.

  3. Introducing the research segment, the host (Altman, who is addressed as Sam when Huet hands back at the end of his segment) said "Last year we maded a bold prediction saying we would see the first AI research intern in the text year" (23:34); the automatic captions render what appears to be "next year" as "text year".

    23:34 Last year we maded a bold / 23:36 prediction saying we would see / 23:38 the first AI research intern in / 23:40 the text year.

  4. The host said "A few weeks ago we reached that goal." (23:48) and "We now have an AI research intern." (23:52).

    23:48 A few weeks ago we reached that / 23:51 goal. / 23:52 We now have an AI research / 23:58 intern.

  5. Tejal Patwardhan introduced herself on stage with the words "I'm a research lead for post training" (24:17).

    24:17 I'm a research lead for post / 24:20 training

  6. Patwardhan said "It failed the vast majority of times on a task that took a day or longer but as of July and E-our models have taken on over a third of day long research tasks with no intervention required." (26:27); "E-our" is the automatic captions' rendering of a second month name that is not clear from the captions.

    26:27 It failed the vast majority of / 26:29 times on a task that took a day / 26:31 or longer but as of July and / 26:35 E-our models have taken on over / 26:37 a third of day long research / 26:39 tasks with no intervention / 26:41 required.

  7. Holly Li, demonstrating dots, said "Our engineers have their dot fixing dozens of bugs every day." (15:34), and just before it "this is a real example" (15:32).

    15:32 And this is a real example. / 15:34 Our engineers have their dot / 15:36 fixing dozens of bugs every day.

  8. Altman said "Around 1.2 billion people use ChatGPT." (44:11).

    44:11 Around 1.2 billion people use / 44:17 ChatGPT.

  9. Altman said, on the API, "We've still kept more than 99 percent reliability. We have cut the time to first token by 45 percent" (31:06).

    31:06 We've still kept more than / 31:08 99 percent reliability. / 31:10 We have cut the time to first / 31:14 token by 45 percent

Nothing appears on this site that we have not opened and linked.

Filed under

The morning email

The day's three lead stories with what they mean for you, and up to four quick hits, each morning when the day's edition comes out, usually soon after 07:00 Oslo time. Free.

We email you a link to confirm. An issue may include one sponsor, always labelled Sponsored · Advertisement. Our emails count opens and clicks, not linked to you. Unsubscribe in one click. What we keep