Agent-workdays per human workday in frontier research (self-reported)

Agent effort used by a frontier lab's research organisation, in standard eight-hour workdays, per workday of human labour. Today only OpenAI reports it, about itself.

Reason through this

The reading

OpenAI says 3.1 as of mid-August 2026, having crossed 1.0 in June; Anthropic says its overall progress multiplier is below 2x. Lab statements about themselves, not outcome measures; the number to watch is an independent replication, which METR is funded to build.

Every readingeach dot links to its record

a readingdisputednormal rangefast range

reported

1 reading drawn from openai_blog.openai.rsi_agent_workdays_per_human.pt; 1 observation in the evidence.

Source: OpenAI research posts. OpenAI.

The numbers

as of	reading	value	record
2026-08-15	openai (disputed)	3.10×	reported

Why it matters

The return arrow made numeric. If deployment inside the lab is feeding invention at more than one agent-day per human-day, the loop Narayanan and Kapoor draw but do not instrument is running; whether it substitutes for external bottlenecks is the thesis question.

What cuts against it

Self-measured and self-published by the lab whose timeline it supports; 'effort' counts agent time, not delivered research; the same post says over half of successful four-to-eight-hour tasks still needed human intervention. Anthropic's September 2026 system card reports staff self-estimating about 4x productivity uplift while the company puts its overall progress multiplier below 2x and observes no sustained AI-attributable 2x acceleration. METR's technical-worker survey finds a self-reported 1.4–2x value uplift, and METR's own RCT measured developers slower with AI.

How it is measured

series openai_blog.openai.rsi_agent_workdays_per_human.pt
source OpenAI research posts · default tier 7 · OpenAI; short quotation
proxy deployment, behaviour · cadence quarterly · valve return arrow

Below one agent-day per human-day, agents are tools inside a human-run process (normal); at three or more, most research effort is agent effort, the regime the AI 2027 'R&D multiplier' assumes (fast). Between is emerging. Tier 7 evidence caps the status at emerging whatever the number says.

Applied to openai_blog.openai.rsi_agent_workdays_per_human.pt.

Latest readings

2026-08-15 openai · 3.10×⚑disputed: OpenAI measuring its own research organisation and publishing the result itself; no independent check (tier 7).as of 2026-08-15

Why this status, and its history

OpenAI's 'Research acceleration' post (6 Sep 2026, via the Internet Archive snapshot because openai.com blocks the fetcher): 3.1 agent-workdays per human workday as of mid-August 2026, having crossed 1.0 in June. Above the fast band (3 or more), but a lab's statement about itself is tier 7 and caps the status at emerging. METR's modelling note puts Anthropic's researcher uplift from coding agents at over 2x; Anthropic's own system card says overall R&D uplift is well short of a doubling. No independent measurement exists yet. Initial seed.

2026-09-10unmeasured to emergingconf — → 30 · the evaluator (automatic)
OpenAI's 'Research acceleration' post (6 Sep 2026, via the Internet Archive snapshot because openai.com blocks the fetcher): 3.1 agent-workdays per human workday as of mid-August 2026, having crossed 1.0 in June. Above the fast band (3 or more), but a lab's statement about itself is tier 7 and caps the status at emerging. METR's modelling note puts Anthropic's researcher uplift from coding agents at over 2x; Anthropic's own system card says overall R&D uplift is well short of a doubling. No independent measurement exists yet. Initial seed.
source1

Why this confidence

Rests on 1 observation from OpenAI research posts; best evidence tier 7, grade D; current.

Confidence is independent of status: 90–95 multiple strong independent sources; 70–89 good evidence, some ambiguity; 50–69 mixed or hard to operationalise; 0–49 limited or vague evidence.

Predictions that lean on this

Happening
François Chollet argued that an intelligence explosion cannot happen, and that the first superhuman AI would be just another step in steady, roughly linear progress. · François Chollet
Happening
Robin Hanson put less than one chance in a hundred on a single hand-coded AI leaping, without warning, from near impotence to overwhelming power. · Robin Hanson
Slower than said
The AI 2027 scenario has AI that automates AI research in early 2027 compounding into artificial superintelligence before that year is out. · AI Futures Project
Slower than said
The AI 2027 scenario had AI assistants speeding the leading lab's algorithmic progress by about half again by early 2026. · AI Futures Project
Too early to tell
AI speeds up its own invention · This site

More predictions that lean on this

Too early to tell
The AI 2027 scenario placed a superhuman AI researcher, one that outdoes the best people at AI research, at around September 2027. · AI Futures Project
Too early to tell
Sam Altman wrote that 2026 would likely bring AI systems able to work out genuinely new insights. · Sam Altman
Too early to tell
Leopold Aschenbrenner called it strikingly plausible that models could do the job of the engineers and researchers who build AI by 2027. · Leopold Aschenbrenner
Too early to tell
Leopold Aschenbrenner expected automated AI research to produce superintelligence, vastly smarter than people, before this decade is out. · Leopold Aschenbrenner
Too early to tell
Nick Bostrom judged it unlikely that the climb from human-level AI to superintelligence would take decades, and more likely that it would be fast or moderate. · Nick Bostrom
Too early to tell
Paul Christiano gave even odds on whether an intelligence explosion driven by software improvements alone is possible at all. · Paul Christiano
Too early to tell
Ajeya Cotra gave about one chance in ten that AI fully automates AI research and development, matching human researchers, by the end of 2026. · Ajeya Cotra
Too early to tell
Tom Davidson's best guess is that AI goes from matching human professionals at almost all cognitive work to far surpassing them in less than a year. · Tom Davidson
Too early to tell
Ege Erdil and Tamay Besiroglu doubt that AI improving its own software alone will set off an intelligence explosion, holding that compute and experiments bound progress. · Ege Erdil and Tamay Besiroglu
Too early to tell
Holden Karnofsky gave better than one chance in ten of transformative AI by 2036, about even odds by 2060, and around two chances in three by 2100. · Holden Karnofsky
Too early to tell
Daniel Kokotajlo's August 2026 median for artificial superintelligence is March 2029, a few months after his median for AI that beats top experts. · Daniel Kokotajlo
Too early to tell
Eli Lifland's August 2026 median for artificial superintelligence is July 2033, about a year after his median for AI that beats top experts. · Eli Lifland
Too early to tell
William MacAskill and Fin Moorhouse think it fairly likely that AI matches human research labour within two decades, and possibly well within one. · William MacAskill and Fin Moorhouse
Too early to tell
William MacAskill and Fin Moorhouse give roughly even odds that AI improving its own software sets off accelerating growth in AI capability. · William MacAskill and Fin Moorhouse
Too early to tell
Eliezer Yudkowsky argued that an AI improving itself would most likely either stall or take off explosively, since a gradual, human-paced takeoff would need implausibly precise diminishing returns. · Eliezer Yudkowsky
Too early to tell
The young Eliezer Yudkowsky cited 2035 as the most widely quoted date for the Singularity and said he, like many others, expected it substantially sooner. · Eliezer Yudkowsky
Too early to tell
AI's help with AI research brings steady gains, not accelerating ones. · Nathan Lambert
Too early to tell
Phil Trammell argues that limits on splitting research across many AI workers could delay a technological singularity, though automating research would probably still produce an explosion. · Phil Trammell

Reading

emergingA model's judgement, not a reading: probably faster than normal. The one reading on the site is just above the fast threshold, with agent effort exceeding human effort several times over inside one lab's research organisation, so it was read as fast. The strongest objection is that the figure is measured and published by the lab it flatters, counts agent time and not delivered research, and sits beside another lab's statement of a much smaller overall speed-up and independent studies that find modest gains.

Latest

3.10×⚑as of 2026-08-15

What the rule reads

3.10×as of 2026-08-15

The rule

Normal ≤ 1.00×; fast ≥ 3.00×. Between them it reads emerging.

Confidence 30 of 95: limited or vague evidence. grade D

Leading: agent effort inside research labs runs ahead of the capability it produces by a model generation or more.

Updated 2026-09-10