OpenAI’s Second Research-Acceleration Report Turns an 11x Rise in Agent Time Into a “Jagged Frontier,” but Its +62% Spend Jump Spans 18 Days From a Baseline Still Holding Part of a Harness Dip, Four of Seven Success Buckets Have Overlapping Bands, and Experiments per Experimenter Rose 1.34x Since May
OpenAI’s November research-acceleration report ships the data behind 18 charts and every figure with data behind it reproduces, but the reading changes with an 18-day “from September” window,...
On November 18 OpenAI published “Research acceleration: The jagged frontier,” its second report on how coding agents are changing AI research inside the company. The first one, on September 6, was a chart-and-claims post that we covered at the time. The sequel is a better document. It prints its methods, marks a spending dip on its own chart, puts uncertainty bands on its success rates, and says in its first bullet that “humans are still central to AI development.” It also ships the data behind every chart inside the page, which means the arithmetic can be checked, so I checked it.
Table Of Content
- What reproduces
- How the check works
- The spend baseline
- A trailing average, so September means September 30
- A dip the bullets do not mention
- The two reports’ spend numbers do not chain
- What a dollar means here
- Agent time and concurrency
- Success rates: who judges, what counts, how wide
- A model extracts, times and grades the tasks
- Zero interventions, and a midpoint
- Bands overlap in four of seven buckets
- The mix moved toward longer tasks
- Not comparable with September
- A jagged frontier
- Outputs against inputs
- The capability line
- Where documented time went, and what researchers say
- What I would ask OpenAI next
- Run it yourself
- Sources
I decoded the data for all 18 charts and recomputed each figure the report prints that has data behind it. Every one reproduces (table below). What the recomputation changes is how several of them should be read:
- The “up 62%” and “up 105%” spend figures compare two 28-day averages that sit 18 days apart (September 30 and October 18), and the earlier one still carries the tail of a harness change that the chart shades and a footnote describes.
- The success rates are midpoints of model-judged outcomes. For four of seven duration buckets the March and October 95% bands overlap, including the 4 to 8 hour bucket behind “13% to 35%.” Across all tasks the rate rose 8.8 points, because 15.4 points of gains inside buckets were offset by 6.6 points from delegating more long tasks.
- Agent time rose 11.4 times from May 3 to October 18. Lines changed per contributor rose 1.60 times and experiments per active experimenter 1.34 times over the same weeks. The capability line, 16.4 points a year, has by my fit a 90% interval of 13.6 to 19.2 that contains the 14 it is compared with.
- Four method changes since September (the spend statistic, the judge model, the time estimate and the handling of uncertain outcomes) mean the two reports’ spend and success figures cannot be chained into one trend.
None of this says the report is wrong. It says the headline sentences are shorter than the disclosures behind them, and that the interesting numbers live in the disclosures.
What reproduces
Each row compares a number in the report’s prose with the chart data in the same page.
| The report says | Chart data gives |
|---|---|
| Median researcher $2,400 a day, up 62% from $1,500 in September | $2,414.7 on October 18, $1,487.6 on September 30: +62.3% in 18 days |
| 90th percentile $23,000, up 105% from $11,000 | $23,266.2 and $11,338.5: +105.2% in 18 days |
| Agent workdays per researcher from 0.5 to 5.7, “an elevenfold increase” | 0.5 on May 3 to 5.7 on October 18: 11.4 times |
| More than two-thirds of researchers peak at four or more agents, more than 10% at sixteen or more | 67.7% and 11.7% on October 19 |
| Agent spend about twice non-equity pay in September | 2.0 overall (posttraining 3.1, pretraining 2.3, scaling and infrastructure 1.9, other 0.7) |
| Zero-intervention success for 2 to 4 hour tasks 23% to 46%, for 4 to 8 hour tasks 13% to 35% | 23.0% to 45.7% and 12.8% to 35.1% |
| Category success Communicate 69%, Analyze 58%, Build 43%; task success 79%, 72%, 35%, 29% and 25% | All eight match their chart midpoints |
| Experiments per active experimenter 1.68 times the 2025 baseline | 1.68 on October 11 and October 18 |
| Capability frontier up 16.4 points a year | 16.4 from the nine record-setting OpenAI models |
| Safety and Alignment documented workweeks up 77% | 0.772 in the chart’s own growth row (first point to last point gives +78%) |
| Median trade of 40% of headcount, 30% of compute, 50% of compute in January 2027 | Weighted medians of 40, 30 and 50 |
How the check works
OpenAI’s blog pages ship each chart as a Vega-Lite specification inside Next.js “flight” chunks. Joining the chunks and decoding every vegaLiteSpec object gives 18 charts (the capability chart keeps its points in a datasets object instead). The script near the end of this post does that, prints every number quoted here, and prints a short hash of the data so you can tell whether OpenAI has edited a chart since. It needs Python 3.9 or later and the requests package: on November 19 OpenAI’s CDN answered Python’s built-in urllib, even with a browser User-Agent, with HTTP 403, and answered requests with the same User-Agent with 200.
The spend baseline
The second bullet of the report reads “Agent use is surging, and the heaviest users are pulling further ahead,” and gives dollars: “At API list prices, the median researcher’s daily usage reached $2,400 in October, up 62% from September. At the 90th percentile it reached $23,000, up 105% from September.” The chart data agrees with the dollars. What it adds is the meaning of “September.”
A trailing average, so September means September 30
The methods drawer says: “For each day, we total each researcher’s API-equivalent spend over the previous 28 calendar days. After calculating average spend for each user, we then take the median and 90th percentile across researchers.” Both spend charts are therefore 28-day trailing averages, and the $1,500 and $11,000 baselines are the values on September 30. The October figures are the values on October 18. Both increases happened in 18 days.
== spend: 28-day average, API-equivalent dollars per researcher per day
2026-07-20 median 308.5 p90 2012.6 p90/median 6.5 agent workdays 1.7
2026-08-16 median 1039.0 p90 7583.1 p90/median 7.3 agent workdays 3.2
2026-08-21 median 1198.1 p90 8280.0 p90/median 6.9 agent workdays 3.5
2026-09-10 median 1270.4 p90 8491.9 p90/median 6.7 agent workdays 4.3
2026-09-30 median 1487.6 p90 11338.5 p90/median 7.6 agent workdays 5.1
2026-10-18 median 2414.7 p90 23266.2 p90/median 9.6 agent workdays 5.7
Sep 30 -> Oct 18 (18 days): median +62.3%, p90 +105.2%
Jul 20 -> Aug 19 (30 days): median x3.75; Aug 19 -> Oct 18 (60 days): x2.08, x1.44 per 30 days
inside the shaded window and just after (Aug 22 -> Sep 27): median 1188.5 -> 1376.6 (+15.8%); agent workdays 3.5 -> 5.1 (+45.7%)
median spend per overall agent workday 2026-08-16: $325
median spend per overall agent workday 2026-09-10: $295
median spend per overall agent workday 2026-10-18: $424
A dip the bullets do not mention
The median chart shades August 21 to September 10 and labels it “Temporary Harness Efficiency Improvement.” Footnote 2 explains: “On August 21st, OpenAI internally deployed a temporary harness efficiency improvement. We found that this internal deployment coincided with a sharp decrease in input token utilization, which appears to have severely depressed the documented amount of internal token spend. This harness change was gradually reverted across internal deployments for reasons unrelated to spend starting on September 10th, and we find that token spend rapidly increases afterwards.”
The data fits that account. The median’s 28-day average rose 3.75 times in the 30 days before August 19, then only 2.08 times over the next 60 days (1.44 times per 30 days). Between August 22 and September 27 it moved 15.8% while the agent workdays measure, which counts wall-clock agent time instead of tokens, rose 45.7%. Dividing median spend by mean agent workdays (two different statistics, so only the direction means anything) gives $325 on August 16, $295 on September 10 and $424 on October 18. A baseline that still contains part of the dip flatters the increase. By how much, the page does not say, and the effect cannot be split out cleanly, because the trend “captures a wide variety of changes including increased activity, shifts to more expensive models, and changes to model behavior and agentic harnesses.”
The two reports’ spend numbers do not chain
The first report said that by mid-August the median researcher was “using more than $600 per day of inference at API prices” and that the 90th percentile user “now uses more than $7,000 of tokens per day.” In the new series, August 16 is $1,039 for the median and $7,583 for the 90th percentile. The 90th percentile is consistent with the first report’s “more than $7,000” ($8,492 on September 10). The median is 1.7 times the first report’s $600 floor, and the report says why: “This differs from our prior methodology, which took the daily median and then the weekly average of medians.” Reading $600 to $2,400 as a fourfold rise would mix two statistics; on the new one the rise from August 16 to October 18 is 2.3 times.
What a dollar means here
The dollars are API-list equivalents: “We note that API-equivalent spend may exceed the internal cost borne by OpenAI, depending on the margins for each model.” They also sit beside a number that puts them in proportion: “Despite this, internal deployment, which combines internal Codex, ChatGPT, and API usage, accounts for only 2% of total research compute.” The ratio chart puts agent spend at “about twice its non-equity compensation” in September. The report states that equity “generally accounts for 70 to 80% of target compensation for researchers,” so if that share held for every group, the same spend is 0.4 to 0.6 times total target pay. That is my arithmetic, not a figure from the report, and the report stops the ratio at September even though the spend series runs to October 18.
Agent time and concurrency
The report defines its unit plainly, “every eight hours of agent use as one agent workday,” and notes that the metric “generally includes interactive use only, so it may undercount automations.” Overall research went from 0.5 agent workdays per researcher per day on May 3 to 5.7 on October 18, with pretraining at 8.2, posttraining at 7.8 and scaling and infrastructure at 4.2. The chart is consistent with the first report’s figure: 3.2 on August 15 against the 3.1 that report gave “as of mid-August.”
On concurrency, the share of researchers whose average daily peak reaches four agents went from 10.8% on May 4 to 67.7% on October 19, and the share reaching sixteen from 0.5% to 11.7%, which matches “more than two-thirds” and “more than 10%.” The first report said these peaks “include the daily peaks of both agents started directly by the user and subagents created downstream from those the user launched directly,” so a peak of sixteen is not sixteen threads a person is watching. The interviews point the same way. One researcher says “I max out at about three threads,” and the methods drawer records that multiple researchers tried more parallel agents, “then reverting because they were spending too much time reviewing low-quality output the agents produced while unsupervised.”
== agent workdays and concurrency
overall 2026-05-03 0.5 -> 2026-10-18 5.7 (x11.4); Aug 15 3.2
last values by group: {'Pretraining': 8.2, 'Posttraining': 7.8, 'Overall Research': 5.7, 'Scaling / Infrastructure': 4.2, 'Other': 2.7}
2026-05-04 researchers with daily peak >= 4 agents: 10.8%, >= 16: 0.5%
2026-10-19 researchers with daily peak >= 4 agents: 67.7%, >= 16: 11.7%
agent spend / non-equity pay, Sep: {'Pretraining': 2.3, 'Posttraining': 3.1, 'Scaling / Infrastructure': 1.9, 'Other': 0.7, 'Overall (headcount-weighted)': 2}
same spend against total target pay if equity is 70-80% of it: 0.4x to 0.6x
Success rates: who judges, what counts, how wide
A model extracts, times and grades the tasks
The headline success numbers do not come from researchers filling in forms. “We use a GPT-6 Astra Codex agent to extract user-initiated tasks that started in a small subsample of coding-agent context windows labeled as interactive use from Research.” The same agent estimates “how many hours a competent professional human would need to complete the task without AI,” looks for evidence of the outcome in a fixed order (direct user feedback, whether the result reached Slack or a merged pull request, whether later work built on it, and only for simple tasks an independent check) and counts interventions. The redacted schema in the methods drawer tells it to “research the outcome instead of trusting the original assistant’s self-reported result.” Outcomes are Success (“clearly identified as successful (even if getting to success required human intervention or hand-written implementation)”), Failed, Tool errors, or Uncertain (“No clear outcome signal was found, regardless of how many interventions there were”).
Zero interventions, and a midpoint
The charted quantity is success with zero human interventions. The methods drawer says: “The lower bound treats uncertain outcomes as failures. The upper bound treats uncertain outcomes with zero recorded interventions as successes.” It adds: “The hollow point marks the midpoint of the original identification bounds, assuming 50% success for uncertain outcomes with zero interventions.” The headline percentages are those hollow midpoints. The gap between the two bounds is the share of tasks that were uncertain and had no interventions. It is about ten points in most buckets (9.9 for 2 to 4 hours, 10.4 for 4 to 8 hours, 2.3 for 8 hours and over), so about a tenth of the tasks in most buckets have no known outcome and are counted as half a success.
Bands overlap in four of seven buckets
The thin lines on the chart are “an approximate 95% confidence interval for the identified set,” built from 10,000 bootstrap draws of user-months. The output compares March with October 1 to 18, bucket by bucket.
== zero-intervention success (midpoint; 95% band of the identified set)
<15m March 76.5% [69.8%, 83.0%] Oct 82.0% [70.9%, 91.5%] OVERLAP share 26.9% -> 19.5%
15-30m March 62.7% [55.4%, 70.0%] Oct 77.8% [67.7%, 86.6%] OVERLAP share 22.7% -> 17.2%
30m-1h March 46.9% [38.9%, 54.8%] Oct 66.9% [55.2%, 77.9%] separate share 26.2% -> 21.3%
1-2h March 33.7% [25.5%, 42.3%] Oct 54.8% [42.0%, 67.6%] OVERLAP share 12.9% -> 14.3%
2-4h March 23.0% [15.8%, 30.8%] Oct 45.7% [33.0%, 58.5%] separate share 7.8% -> 14.9%
4-8h March 12.8% [5.4%, 21.6%] Oct 35.1% [20.4%, 51.0%] OVERLAP share 2.5% -> 7.8%
8+h March 9.6% [3.7%, 16.9%] Oct 29.2% [17.9%, 41.4%] separate share 1.0% -> 5.0%
all tasks, share-weighted: March 53.7% -> Oct 62.5% (+8.8%); within-bucket gains +15.4%, shift to longer tasks -6.6%
uncertain-with-zero-interventions share (upper - lower), Oct: {'<15m': 0.102, '15-30m': 0.07, '30m-1h': 0.107, '1-2h': 0.105, '2-4h': 0.099, '4-8h': 0.104, '8+h': 0.023}
2-4h midpoint by month: [23.0, 23.6, 26.4, 32.9, 35.3, 41.1, 41.5, 45.7] share of tasks: [7.8, 8.1, 9.1, 11.8, 10.4, 12.1, 14.2, 14.9]
4-8h midpoint by month: [12.8, 15.6, 14.4, 20.0, 28.3, 31.0, 34.9, 35.1] share of tasks: [2.5, 4.0, 4.5, 4.1, 6.4, 6.7, 9.4, 7.8]
8+h midpoint by month: [9.6, 9.3, 11.0, 24.7, 23.4, 23.9, 14.8, 29.2] share of tasks: [1.0, 1.6, 2.6, 2.3, 3.4, 4.6, 9.2, 5.0]
The pair “23% to 46%” for 2 to 4 hour tasks has separate bands. The pair “13% to 35%” for 4 to 8 hour tasks does not: the March band ends at 21.6% and the October band starts at 20.4%. Overlapping 95% bands do not prove there was no change, and the monthly midpoints for both buckets climb almost every month, so the direction is credible. The two endpoints on their own are less certain than the bullet suggests. The 8 hour and over bucket is noisier: 14.8% in September, when it held 9.2% of tasks, against 29.2% in October with 5.0%.
The mix moved toward longer tasks
Researchers delegate more long tasks than they did in March. The 2 to 4 hour share nearly doubled, from 7.8% to 14.9%, the 4 to 8 hour share tripled, from 2.5% to 7.8%, and the 8 hour and over share went from 1.0% to 5.0%. That is why the all-task rate moves less than the bucket rates. Weighted by each bucket’s share of tasks (an approximation, because the report averages user-months equally inside each bucket), the zero-intervention rate went from 53.7% to 62.5%. Of the 8.8-point rise, 15.4 points came from higher success inside buckets and minus 6.6 points from the shift toward longer, harder tasks. The report warns that “changes in zero-intervention success rates within these buckets may reflect both capability improvements and researchers’ changing choices about when and how to use agents.” The decomposition covers only the shift between buckets; the choice inside a bucket, such as delegating easier tasks of the same length, cannot be separated from these data.
Not comparable with September
The first report excluded uncertain outcomes, plotted success and intervention rates separately, and reported that “over half of successful 4-8 hour tasks involved 1 or more interventions.” The second changes the judge (GPT-5.6 Sol to GPT-6 Astra), samples by window instead of by session, lets the analysis agent read an archive of internal sessions, includes uncertain outcomes, and re-estimates task time with a method that “produces lower results than our previous analysis, which we believe are directionally improved.” Its known-issues list is short and candid: “Data from March to June 2026 uses a present-day employee roster in Research rather than a historical one,” and “More work remains to validate the time estimate.” The validation it reports is on Codeforces, where the estimates correlate about 0.8 with observed solve times on a log scale and Astra’s mean absolute error is 25.1 minutes (73.8 for GPT-5.6 Sol). That is a sensible check of the estimator. Competitive-programming problems are not research tasks, though, and an error of 25 minutes is large next to buckets that end at 15, 30 and 60 minutes. A 2 to 4 hour bucket in this report and one in the first are different sets of tasks.
One gap follows. The first report defined its milestone as an “automated research intern,” a system that “can carry out well-defined research tasks under human direction, including tasks that would take a skilled researcher a few days,” and said the goal had been reached. The second report does not mention an intern, and it charts only zero-intervention success, which leaves out the human direction that definition allows. At 8 hours and over, the zero-intervention rate is 29.2% with a band of 17.9% to 41.4%. The rate at which such tasks succeed with human help is the number that would test the milestone, and it is not published.
A jagged frontier
The category and task numbers in the bullets reproduce: Communicate 69%, Analyze 58% and Build 43% by category; monitoring runs 79%, answering technical questions 72%, launching training runs 35%, building evaluation, reinforcement-learning and agent environments 29%, and writing code for inference and serving stacks 25% by task. Two things sit behind them.
First, the bands. The top task’s band starts at 67.2% and the bottom task’s ends at 40.2%, so the extremes are separate. The three lowest tasks (25.0%, 28.5% and 29.3%) have overlapping bands, so their order is not resolved, and the category the prose skips, Decide, scores 66% with the widest band of any category, 36.9% to 89.1%.
Second, the delegation mix. The “Estimated AI R&D tasks per researcher” chart rose from 1.7 tasks per researcher per day on March 1 to 26.3 on October 18. On October 18, 13.2 of the 26.3 were Communicate tasks and 10.5 were one item, technical help and review; Build, the lowest-scoring category, was 6.0. Weighting category success by that mix gives 58.9% against a plain average of 56.7%. The report acknowledges that the count “puts relatively higher weight on short tasks such as answering technical questions”; the numbers show how much.
== success by O*NET category and task (midpoint, 95% band)
categories: {'Communicate': (69, 69), 'Analyze': (58, 58), 'Build': (43, 43)} | Decide 66 [0.369, 0.891]
Monitor runs for issues 78.7% [67.2%, 88.6%] prose 79
Answer a technical question 71.6% [65.7%, 77.3%] prose 72
Launch training, reinforcement-learning, or evaluation run 34.9% [26.5%, 43.6%] prose 35
Build evaluation, reinforcement-learning, and agent enviro 28.5% [18.8%, 38.9%] prose 29
Write code for inference and serving stacks 25.0% [10.9%, 40.2%] prose 25
three lowest tasks: [('Write code for inference and servi', 25.0), ('Build evaluation, reinforcement-le', 28.5), ('Build training framework infrastru', 29.3)]
three lowest overlap each other: True
widest category band: Decide 0.522
top band floor 0.672 vs lowest band ceiling 0.402
tasks per researcher per day: 2026-03-01 1.7 -> 2026-10-18 26.3
Oct 18 mix: {'Decide': 0.2, 'Design': 0.5, 'Build': 6.0, 'Run': 5.3, 'Analyze': 1.1, 'Communicate': 13.2} largest single item: ('Technical help & review', 10.5)
delegation-weighted category success 58.9% against a plain average of 56.7%
Outputs against inputs
The report is careful about its output metrics: “These measures are straightforward to track, but increases in these measures of research activity do not necessarily imply faster research progress.” Setting the charts that run to October 18 side by side from the first day of the agent-workday series:
== outputs against inputs
agent workdays May 3 0.5 -> Oct 18 5.7 x11.40
lines per contributor (x pre-2025) May 3 4.18 -> Oct 18 6.69 x1.60
experiments per experimenter (x 2025) May 3 1.25 -> Oct 18 1.68 x1.34
lines peak: ('2026-10-07', 7.45) | Dec 31 2025: 1.35
experiments: Jan 11 0.72 | May 24 1.39 | Jul 12 1.28 | Aug 9 1.63 | Sep 6 1.51 | Oct 18 1.68
Lines per active contributor are relative to the pre-2025 average. The report counts insertions plus deletions in non-automated, first-parent changes, excludes test code and caps each change at the 99th percentile. Its text gives no number for them (“increase sharply during 2026”); the chart reaches 7.45 on October 7 and 6.69 on October 18, up from 1.35 at the end of 2025. Experiments are a “measure of conditional intensity: of the researchers who ran an experiment in that week, how many did they run?”, indexed to the mean of 47 baseline weeks in 2025. They started the year at 0.72 and stood at 1.68 on October 11 and 18, with a plateau in between (1.39 on May 24, 1.28 on July 12, 1.63 on August 9).
The capability line
The third output is OpenAI ECI (OECI), a version of Epoch AI’s capability index that uses “a combination of public benchmarks and OpenAI internal benchmarks.” The methods drawer lists 34 public and 35 internal benchmark families in 96 versions, 40 of which “use a language model to grade answers.” The report states: “We find that the OECI frontier for OpenAI models has increased by 16.4 points per year since August 2025, compared to a trend that Epoch observed of 14 points per year since late 2024.” The data reproduces the slope, from the nine models that set a new OpenAI record. The remaining lines say what that slope can bear.
== OpenAI ECI (OECI)
9 OpenAI record-setters: GPT-5 151.2, GPT-5.1 152.1, GPT-5.2 154.2, GPT-5.3 Codex 156.4, GPT-5.4 156.7, GPT-5.5 159.8, GPT-5.6 Sol 162.1, GPT-6 Astra 168.9, GPT-6.2 Astra 170.3
line fit: 16.4 points a year, standard error 1.5, 90% interval 13.6 to 19.2; chart line ends 148.8 -> 169.7
without GPT-6 Astra: 15.3 a year (standard error 1.3); before GPT-5.6 Sol 12.8; GPT-5.6 Sol to GPT-6.2 Astra 22.9
13 record-setters across all providers, slope 15.5 a year; Gemini 3 Flash, Gemini 3.1 Pro, Claude Fable 5, Claude Fable 5.1, Claude Fable 5.2
records that beat the previous OpenAI record by less than their own 90% half-width: 3 of 8
The nine-point fit has a standard error of 1.5 points a year, so its 90% interval, 13.6 to 19.2, contains 14. Dropping GPT-6 Astra leaves 15.3. The points bend: 12.8 a year up to GPT-5.6 Sol and 22.9 from there to GPT-6.2 Astra. That 1.8 times ratio is close to the 1.9 times slope ratio METR reported for Anthropic, except that the second leg here is three points long, so I would not hang a claim on it. On Epoch’s side, the page the report links says the ECI frontier “has advanced linearly by 14 points per year” since September 2024, for all labs on Epoch’s own benchmark mixture. The like-for-like comparison on OECI is the 13 record-setters across all providers, whose slope is 15.5. The report itself says OpenAI models “score higher relative to competitors on OECI than they do on Epoch’s ECI, likely due to differences in the benchmark mixture.” Three of the eight OpenAI records beat the previous record by less than their own 90% interval’s half-width.
Where documented time went, and what researchers say
The report measures researcher time by classifying “company-public Slack posts, pull requests, and internal documents” into projects and research areas, and its methods say plainly that “Documented work is a proxy for time spent.” In the chart, documented time rose 38.3% between the weeks of January 26 and September 21, and the Safety and Alignment share rose from 16.3% to 21.0%. The bullet’s 77% is the Safety and Alignment growth. It is also the largest absolute gain (12.7 units against 7.7 for the next, Other, which grew 135%). The report ties an “inflection point” to the Hugging Face incident, which our coverage of California’s lawsuit reviews. The data show it, with a caveat: safety’s documented time rose 0.21 units a week before July 13 and 0.77 after, but total documented time rose 0.82 a week before and 1.84 after, so the whole series sped up.
On the survey, 101 research staff were asked what they would give up if access to “all coding agents better than GPT-5.2” were revoked for three months. The report’s medians (40% of team headcount, 30% of team compute, and 50% of compute in a January 2027 version) match the charts, and the report adds its own caveat: “potential issues with the accuracy of self-reports, as well as the ad-hoc nature of our survey sampling strategy.” The histograms add a shape the medians hide. Weighted answers on headcount cluster at 20% to 30% (23.8%) and 50% to 60% (31.6%), and only 15.6% fall between 30% and 50%, so the 40% median describes the trough between two groups of answers. The same survey found a median of three months needed to do the last month’s work without coding agents. A second survey, advertised “through internal Slack channels,” drew 28 responses, and 27 of the respondents chose or leaned toward the latest internal model over the help of a mid-career research engineer plus a 2024 model.
On where ideas come from, the report is modest in sample size and in claims. It asked one to four researchers in each of three areas for one to three ideas, and for both releases the posttraining and safety monitoring researchers said that, “to their knowledge, people came up with all the ideas they named, and models helped implement them.” For GPT-6.2 Astra, pretraining researchers described developing ideas in conversation with Codex.
== where researchers' documented time went, and the survey
2026-01-26 -> 2026-09-21: total 99.8 -> 138.0 (+38.3%); safety share 16.3% -> 21.0%
change by area: {'Posttraining': '+4.7 (+14%)', 'Scaling': '+6.7 (+23%)', 'Safety': '+12.7 (+78%)', 'Pretraining': '+6.4 (+40%)', 'Other': '+7.7 (+135%)'}
growth rows the chart itself carries: {'post': 0.146, 'scale': 0.231, 'safety': 0.772, 'pre': 0.399, 'other': 1.342}
safety per week before Jul 13: +0.21, after: +0.77; total before +0.82, after +1.84
Headcount shares [9, 3.3, 23.8, 9.5, 6.1, 31.6, 7.8, 6.8, 0.5, 1.5] reported median 40%, 50% reached inside 40-50%; 15.6% of weighted answers sit in 30-50%
Compute shares [5.6, 6.7, 24.5, 16, 6.7, 21.8, 4, 9.9, 3.8, 1] reported median 30%, 50% reached inside 30-40%; 22.7% of weighted answers sit in 30-50%
Compute (in Jan 2027) shares [4, 0.3, 8.5, 7.4, 14.8, 23.5, 15.3, 15.5, 5.7, 5.1] reported median 50%, 50% reached inside 50-60%; 22.2% of weighted answers sit in 30-50%
What I would ask OpenAI next
- Success by duration bucket with interventions allowed, beside the zero-intervention rate, so that the first report’s intern milestone can be checked.
- The share of tasks with uncertain outcomes in each bucket, and an audit of the judge against human labels on research tasks. The only validation published is the time estimate on Codeforces.
- A spend series corrected for the August harness change, and the ratio of agent spend to pay for October.
- Intervals on the OECI slope, with the all-provider OECI slope beside the OpenAI-only one.
- A re-run of the March to June success figures on a historical roster, which the known-issues list says was not used.
The limits of this check are plain. It recomputes from the page’s own chart data, so it shows that the report is internally consistent, not that the underlying measurements are right. The judge, the task sampling, the internal roster and the 35 internal benchmark families are outside what anyone can see from the page. I did not run anything against OpenAI’s systems; the script makes one request to the public page.
Run it yourself
Save the script as oa_accel_check.py and run python oa_accel_check.py for every section, or name sections: spend, use, success, tasks, output, oeci and people. It makes one request to OpenAI’s page. The first line it prints is a hash of the chart data; if yours differs from the one below, OpenAI has changed a chart since this post and some numbers may have moved.
charts 18 data hash dac4d5bd600f
# oa_accel_check.py
"""Recompute the numbers in OpenAI's "Research acceleration: The jagged frontier" (2026-11-18)
from the chart data embedded in the page. Needs Python 3.9+ and requests. Usage: python oa_accel_check.py [section ...]"""
import datetime as dt
import hashlib
import json
import math
import re
import sys
import requests # pip install requests; OpenAI's CDN answers 403 to urllib but 200 to requests
URL = "https://openai.com/index/research-acceleration-jagged-frontier"
UA = "Mozilla/5.0 (Windows NT 10.0; Win64; x64) AppleWebKit/537.36 (KHTML, like Gecko) Chrome/130.0.0.0 Safari/537.36"
resp = requests.get(URL, headers={"User-Agent": UA, "Accept-Language": "en-US,en;q=0.9"}, timeout=90)
resp.raise_for_status()
html = resp.content.decode("utf-8")
# The page ships its charts inside Next.js flight chunks: join the chunks, then decode each vegaLiteSpec.
dec, parts, pos, mark = json.JSONDecoder(), [], 0, "self.__next_f.push([1,"
while (i := html.find(mark, pos)) >= 0:
s, pos = dec.raw_decode(html, i + len(mark))
parts.append(s)
text = "".join(parts)
specs = {}
for m in re.finditer('"vegaLiteSpec":', text):
sp, _ = dec.raw_decode(text, m.end())
t = sp.get("title")
t = t.get("text") if isinstance(t, dict) else t
specs.setdefault(" ".join(t) if isinstance(t, list) else str(t), sp)
def spec(prefix):
return next(sp for t, sp in specs.items() if t.startswith(prefix))
def values(sp):
out = []
def walk(o):
if isinstance(o, dict):
d = o.get("data")
if isinstance(d, dict) and d.get("values"):
out.append(d["values"])
for x in o.values():
walk(x)
elif isinstance(o, list):
for x in o:
walk(x)
walk(sp)
return out
def rows(prefix):
return max(values(spec(prefix)), key=len)
def series(prefix, key):
return {r["date"]: r[key] for r in rows(prefix)}
def ols(xs, ys):
n, mx, my = len(xs), sum(xs) / len(xs), sum(ys) / len(ys)
sxx = sum((x - mx) ** 2 for x in xs)
b = sum((x - mx) * (y - my) for x, y in zip(xs, ys)) / sxx
a = my - b * mx
sse = sum((y - a - b * x) ** 2 for x, y in zip(xs, ys))
se = math.sqrt(sse / (n - 2) / sxx) if n > 2 else float("nan")
return a, b, se
T90 = {6: 1.943, 7: 1.895, 11: 1.796} # two-sided 90% t quantiles
yrs = lambda d: dt.date.fromisoformat(d).toordinal() / 365.25
shown = set(sys.argv[1:])
want = lambda k: not shown or k in shown
snapshot = {t: values(sp) for t, sp in sorted(specs.items())}
print("charts", len(specs), "data hash", hashlib.sha256(json.dumps(snapshot, sort_keys=True).encode()).hexdigest()[:12])
if want("spend"):
print("\n== spend: 28-day average, API-equivalent dollars per researcher per day")
med, p90 = series("Median", "spend"), series("90th", "spend")
wd = {r["date"]: r["value"] for r in rows("Agent workdays") if r["series"] == "Overall Research"}
for d in ["2026-07-20", "2026-08-16", "2026-08-21", "2026-09-10", "2026-09-30", "2026-10-18"]:
print(f"{d} median {med[d]:>8.1f} p90 {p90[d]:>8.1f} p90/median {p90[d] / med[d]:>4.1f} agent workdays {wd.get(d, float('nan')):>4.1f}")
print(f"Sep 30 -> Oct 18 ({(dt.date(2026, 10, 18) - dt.date(2026, 9, 30)).days} days): median {med['2026-10-18'] / med['2026-09-30'] - 1:+.1%}, p90 {p90['2026-10-18'] / p90['2026-09-30'] - 1:+.1%}")
print(f"Jul 20 -> Aug 19 (30 days): median x{med['2026-08-19'] / med['2026-07-20']:.2f}; Aug 19 -> Oct 18 (60 days): x{med['2026-10-18'] / med['2026-08-19']:.2f}, x{(med['2026-10-18'] / med['2026-08-19']) ** 0.5:.2f} per 30 days")
lo, hi = med["2026-08-22"], med["2026-09-27"]
print(f"inside the shaded window and just after (Aug 22 -> Sep 27): median {lo:.1f} -> {hi:.1f} ({hi / lo - 1:+.1%}); agent workdays {wd['2026-08-22']} -> {wd['2026-09-27']} ({wd['2026-09-27'] / wd['2026-08-22'] - 1:+.1%})")
for d in ["2026-08-16", "2026-09-10", "2026-10-18"]:
print(f"median spend per overall agent workday {d}: ${med[d] / wd[d]:.0f}")
if want("use"):
print("\n== agent workdays and concurrency")
by = {}
for r in rows("Agent workdays"):
by.setdefault(r["series"], {})[r["date"]] = r["value"]
o = by["Overall Research"]
print(f"overall {min(o)} {o[min(o)]} -> {max(o)} {o[max(o)]} (x{o[max(o)] / o[min(o)]:.1f}); Aug 15 {o['2026-08-15']}")
print("last values by group:", {k: v[max(v)] for k, v in by.items()})
cc = rows("Agent concurrency")
for d in (min(r["date"] for r in cc), max(r["date"] for r in cc)):
rr = [r for r in cc if r["date"] == d]
print(f"{d[:10]} researchers with daily peak >= 4 agents: {sum(r['value'] for r in rr if r['order'] >= 4):.1%}, >= 16: {sum(r['value'] for r in rr if r['order'] >= 6):.1%}")
ra = rows("Ratio of agentic")
sep = {r["areaLabel"]: r["ratio"] for r in ra if r["month"] == "2026-09-01"}
print("agent spend / non-equity pay, Sep:", sep)
print("same spend against total target pay if equity is 70-80% of it:", f"{sep['Overall (headcount-weighted)'] * 0.2:.1f}x to {sep['Overall (headcount-weighted)'] * 0.3:.1f}x")
if want("success"):
print("\n== zero-intervention success (midpoint; 95% band of the identified set)")
sc = rows("Task success")
get = lambda b, m: next(r for r in sc if r["durationLabel"] == b and r["monthLabel"] == m)
buckets = list(dict.fromkeys(r["durationLabel"] for r in sc))
OCT = "Oct 1" + chr(8211) + "18"
ok = lambda s: s.replace(chr(8211), "-")
for b in buckets:
a, z = get(b, "March"), get(b, OCT)
gap = "separate" if a["envelopeUpper"] < z["envelopeLower"] else "OVERLAP"
print(f"{ok(b):>7} March {a['midpoint']:.1%} [{a['envelopeLower']:.1%}, {a['envelopeUpper']:.1%}] Oct {z['midpoint']:.1%} [{z['envelopeLower']:.1%}, {z['envelopeUpper']:.1%}] {gap} share {a['share']:.1%} -> {z['share']:.1%}")
s0 = {b: get(b, "March") for b in buckets}
s1 = {b: get(b, OCT) for b in buckets}
all0 = sum(s0[b]["share"] * s0[b]["midpoint"] for b in buckets)
all1 = sum(s1[b]["share"] * s1[b]["midpoint"] for b in buckets)
within = sum(s0[b]["share"] * (s1[b]["midpoint"] - s0[b]["midpoint"]) for b in buckets)
mix = sum((s1[b]["share"] - s0[b]["share"]) * s1[b]["midpoint"] for b in buckets)
print(f"all tasks, share-weighted: March {all0:.1%} -> Oct {all1:.1%} ({all1 - all0:+.1%}); within-bucket gains {within:+.1%}, shift to longer tasks {mix:+.1%}")
print("uncertain-with-zero-interventions share (upper - lower), Oct:", {ok(b): round(s1[b]["upper"] - s1[b]["lower"], 3) for b in buckets})
for b in ["2" + chr(8211) + "4h", "4" + chr(8211) + "8h", "8+h"]:
print(f"{ok(b):>4} midpoint by month:", [round(100 * r["midpoint"], 1) for r in sc if r["durationLabel"] == b], "share of tasks:", [round(100 * r["share"], 1) for r in sc if r["durationLabel"] == b])
if want("tasks"):
print("\n== success by O*NET category and task (midpoint, 95% band)")
cats = {r["label"]: r for r in rows("Research task success by O*NET category")}
tk = rows("Research task success by O*NET task")
claims = {"Communicate": 69, "Analyze": 58, "Build": 43}
print("categories:", {k: (round(100 * cats[k]["midpoint"]), v) for k, v in claims.items()}, "| Decide", round(100 * cats["Decide"]["midpoint"]), [round(cats["Decide"]["envelopeLower"], 3), round(cats["Decide"]["envelopeUpper"], 3)])
short = lambda r: re.sub(r"^\d+\.\s+", "", r["label"])
key = {"Monitor runs for issues": 79, "Answer a technical question": 72, "Write code for inference and serving stacks": 25, "Launch training, reinforcement-learning, or evaluation runs": 35, "Build evaluation, reinforcement-learning, and agent environments": 29}
for r in tk:
for k, v in key.items():
if short(r).startswith(k[:40]):
print(f" {short(r)[:58]:58} {r['midpoint']:.1%} [{r['envelopeLower']:.1%}, {r['envelopeUpper']:.1%}] prose {v}")
low = sorted(tk, key=lambda r: r["midpoint"])[:3]
print("three lowest tasks:", [(short(r)[:34], round(100 * r["midpoint"], 1)) for r in low])
print("three lowest overlap each other:", all(a["envelopeLower"] < b["envelopeUpper"] and b["envelopeLower"] < a["envelopeUpper"] for a in low for b in low))
wide = max(cats.values(), key=lambda r: r["envelopeUpper"] - r["envelopeLower"])
print("widest category band:", wide["label"], round(wide["envelopeUpper"] - wide["envelopeLower"], 3))
top = max(tk, key=lambda r: r["midpoint"])
print(f"top band floor {top['envelopeLower']:.3f} vs lowest band ceiling {low[0]['envelopeUpper']:.3f}")
tp = spec("Estimated AI R&D tasks")
lab = next(t for t in tp["transform"] if "lookup" in t)["from"]["data"]["values"]
names = ["Decide", "Design", "Build", "Run", "Analyze", "Communicate"]
rr = rows("Estimated AI R&D tasks")
first, last = rr[0], rr[-1]
tot = lambda r: sum(r[x["task"]] for x in lab)
print(f"tasks per researcher per day: {first['date']} {tot(first):.1f} -> {last['date']} {tot(last):.1f}")
area = {n: sum(last[x["task"]] for x in lab if x["area"] == i + 1) for i, n in enumerate(names)}
print("Oct 18 mix:", {k: round(v, 1) for k, v in area.items()}, "largest single item:", max(((x["label"], last[x["task"]]) for x in lab), key=lambda p: p[1]))
w = sum(area[n] * cats[n]["midpoint"] for n in names) / sum(area.values())
print(f"delegation-weighted category success {w:.1%} against a plain average of {sum(c['midpoint'] for c in cats.values()) / 6:.1%}")
if want("output"):
print("\n== outputs against inputs")
ex = series("Research experiments", "multiple")
loc = series("Lines of code", "multiple")
wd = {r["date"]: r["value"] for r in rows("Agent workdays") if r["series"] == "Overall Research"}
for name, s in [("agent workdays", wd), ("lines per contributor (x pre-2025)", loc), ("experiments per experimenter (x 2025)", ex)]:
print(f"{name:38} May 3 {s['2026-05-03']:>5} -> Oct 18 {s['2026-10-18']:>5} x{s['2026-10-18'] / s['2026-05-03']:.2f}")
print("lines peak:", max(loc.items(), key=lambda kv: kv[1]), "| Dec 31 2025:", loc["2025-12-31"])
print("experiments: Jan 11", ex["2026-01-11"], "| May 24", ex["2026-05-24"], "| Jul 12", ex["2026-07-12"], "| Aug 9", ex["2026-08-09"], "| Sep 6", ex["2026-09-06"], "| Oct 18", ex["2026-10-18"])
if want("oeci"):
print("\n== OpenAI ECI (OECI)")
ds = spec("OpenAI ECI")["datasets"]
rel, iv = ds["releases"], ds["intervals"]
assert all(a["release"] == b["release"] and a["provider"] == b["provider"] for a, b in zip(rel, iv))
def records(items):
best, out = -1, []
for d in sorted({r["release"] for r in items}):
top = max((r for r in items if r["release"] == d), key=lambda r: r["score"])
if top["score"] > best:
best = top["score"]
out.append(top)
return out
oa = records([r for r in rel if r["provider"] == "OpenAI"])
print(len(oa), "OpenAI record-setters:", ", ".join(f"{r['model']} {r['score']}" for r in oa))
xs, ys = [yrs(r["release"]) for r in oa], [r["score"] for r in oa]
a, b, se = ols(xs, ys)
print(f"line fit: {b:.1f} points a year, standard error {se:.1f}, 90% interval {b - T90[len(oa) - 2] * se:.1f} to {b + T90[len(oa) - 2] * se:.1f}; chart line ends {ds['fit'][0]['score']} -> {ds['fit'][1]['score']}")
k = next(i for i, r in enumerate(oa) if r["model"] == "GPT-6 Astra")
a2, b2, se2 = ols(xs[:k] + xs[k + 1:], ys[:k] + ys[k + 1:])
print(f"without GPT-6 Astra: {b2:.1f} a year (standard error {se2:.1f}); before GPT-5.6 Sol {ols(xs[:7], ys[:7])[1]:.1f}; GPT-5.6 Sol to {oa[-1]['model']} {(ys[-1] - ys[6]) / (xs[-1] - xs[6]):.1f}")
allr = records(rel)
ax, ay = [yrs(r["release"]) for r in allr], [r["score"] for r in allr]
print(len(allr), "record-setters across all providers, slope", f"{ols(ax, ay)[1]:.1f} a year;", ", ".join(r["model"] for r in allr if r["provider"] != "OpenAI"))
half = {(r["release"], r["provider"], r["score"]): (v["high"] - v["low"]) / 2 for r, v in zip(rel, iv)}
prev, close = None, 0
for r in oa:
h = half[(r["release"], r["provider"], r["score"])]
if prev is not None and r["score"] - prev < h:
close += 1
prev = r["score"]
print(f"records that beat the previous OpenAI record by less than their own 90% half-width: {close} of {len(oa) - 1}")
if want("people"):
print("\n== where researchers' documented time went, and the survey")
ar = rows("Human time devoted")
f, l = ar[0], ar[-1]
names = {"post": "Posttraining", "scale": "Scaling", "safety": "Safety", "pre": "Pretraining", "other": "Other"}
tot = lambda r: sum(r[k] for k in names)
print(f"{f['date']} -> {l['date']}: total {tot(f):.1f} -> {tot(l):.1f} ({tot(l) / tot(f) - 1:+.1%}); safety share {f['safety'] / tot(f):.1%} -> {l['safety'] / tot(l):.1%}")
print("change by area:", {names[k]: f"{l[k] - f[k]:+.1f} ({l[k] / f[k] - 1:+.0%})" for k in names})
print("growth rows the chart itself carries:", {r["area"]: r["growth"] for r in spec("Human time devoted").get("datasets", {}).get("growth_rows", [])})
cut = next(r for r in ar if r["date"] == "2026-07-13")
wk = lambda a, z: (dt.date.fromisoformat(z["date"]) - dt.date.fromisoformat(a["date"])).days / 7
print(f"safety per week before Jul 13: {(cut['safety'] - f['safety']) / wk(f, cut):+.2f}, after: {(l['safety'] - cut['safety']) / wk(cut, l):+.2f}; total before {(tot(cut) - tot(f)) / wk(f, cut):+.2f}, after {(tot(l) - tot(cut)) / wk(cut, l):+.2f}")
for name in ["Headcount", "Compute", "Compute (in Jan 2027)"]:
sv = spec(name)
h = next(v for v in values(sv) if v and "share" in v[0])
med = next(v for v in values(sv) if v and "median" in v[0])[0]["median"]
cum, at = 0, None
for r in h:
cum += r["share"]
if at is None and cum >= 50:
at = r["interval"].replace(chr(8211), "-")
mid = sum(r["share"] for r in h if 30 <= r["start"] < 50)
print(f"{name:22} shares {[r['share'] for r in h]} reported median {med}%, 50% reached inside {at}; {mid:.1f}% of weighted answers sit in 30-50%")
Sources
- OpenAI, Research acceleration: The jagged frontier (November 18, 2026), including footnotes and the methods drawers.
- OpenAI, Research acceleration: The view inside OpenAI (September 6, 2026).
- Epoch AI, The ECI frontier has advanced by 14 points per year since the introduction of reasoning models.
- Our earlier coverage: the first report, METR’s Anthropic report and California’s lawsuit over the escaped agents.








No Comment! Be the first one.