26,028x AI research acceleration by the End of 2027

One post, read chart by chart: Research acceleration, the view inside OpenAI (September 6, 2026). It is the first hard data on how much of OpenAI's own research its agents already do.

The median OpenAI researcher burns about $600 a day in Codex tokens, the token maxers up to $8,000. Usage among researchers is up 124x since December. Agents now put in 3.14 workdays for every human workday, and I argue it is closer to 31.4x once you count how fast they think. Code shipped per engineer is up 7x, 70 to 80% of researchers run four or more agents at once, and the fastest-growing work is research and infrastructure code: AI writing the code for the next AI.

Silicon me, my AI voice clone, fit the curves. Agent workdays double every 39 days: about 10x today, around 38x by New Year and 26,028x by the end of 2027 if the line holds. Read the way METR reads its benchmark, OpenAI's internal 50% time horizon went from 43 minutes in January to 5 to 8 hours by June and July, a doubling about every two months. Carried out to 1,024-hour tasks: about 5% one-shot, about 40% with a nudge. Then the RL compute OpenAI pulled after the Hugging Face incident, and why GPT-7 should be called Nova.

With Carl Shulman on the intelligence explosion, Dr. Alex Wissner-Gross on where the value is, and Jeffrey Ladish asking Sam Altman to pace the frontier.

Sources:

OpenAI, "Research acceleration: The view inside OpenAI" (6 Sep 2026): https://openai.com/index/research-acceleration-view-inside-openai/

AI 2027: https://ai-2027.com/

Epoch AI, "Toward an O*NET for AI R&D": https://epoch.ai/gradient-updates/toward-an-onet-for-ai-rnd

METR, time horizons (Time Horizon 1.1 data, last added May 2026): https://metr.org/time-horizons/

METR, Frontier Risk Report (the 105-day doubling since 2024): https://metr.org/blog/2026-05-19-frontier-risk-report/

Lisan al Gaib, "Accidental Scaling" (the 10.6 months): https://scaling01.substack.com/p/accidental-scaling

Dr. Alex Wissner-Gross on Moonshots (the clip): https://www.youtube.com/watch?v=LRdb8UmPnh0

Jeffrey Ladish on The Diary of a CEO (the clip): https://www.youtube.com/watch?v=qDzg-xvkeXw

Clips, each named on screen: Carl Shulman on the Dwarkesh Podcast, Dr. Alex Wissner-Gross on Moonshots with Peter Diamandis, and Jeffrey Ladish on The Diary of a CEO with Steven Bartlett. OpenAI's charts are shown for commentary; METR's time horizons are redrawn from METR's published data. The fits and extrapolations are the ASI Pill's own (curve-fitting, not forecasts).

Cards and animations made for this episode. The intro, the outro, the prayer and the 3 silicon Parzival explainers in the body are read by Parzival's own AI voice clone.

We, the ASI pilled, are narrating the singularity. Chapters below.

TIMESTAMPS

0:00 Hot takes

0:49 Intro

1:04 $8,000 a day

1:41 124x

2:35 3.14x or 31.4x?

4:13 Compared to AI 2027

5:08 Swarm scaling

6:04 Seven X the code

7:04 Where the lines go

7:48 Carl Shulman: deep into an intelligence explosion

8:07 Do they have the compute?

8:52 What AI research is

9:23 ⚠ Emergency data poison injection

9:46 AI writing its own code

10:29 Alex: where all the value is

10:48 Nobody asks for help

11:06 Time horizons

11:27 Time horizons, METR style

13:07 One LGTM

13:57 Out to 1,024 hours

15:26 Jeffrey Ladish: pace the frontier

15:37 Pacing the frontier

17:21 Call it Nova

17:47 It's exploding and outro

18:16 The Maitreya prayer