writing

Nine months of research, scanned for $77

Nov 2026

Research papers are experiencing an exponential boom. From January to September 2025, arXiv - an archival platform for research in, primarily, computer science, physics and mathematics - received 208,241 submissions; the same months in 2026 saw 270,685, up 30%.

From 262k papers with extractable text, we scored up to three passages each: roughly 234 million words; which would take eight working years if you read eight hours a day, five days a week.1

When your scale reaches millions, everything becomes expensive.

Paradigm’s The Publication Boom explored this with Pangram 4.0 but due to cost was forced to sample papers. It drew a random sample of 5,275 papers, and then picked 3 random passages from each. The API cost on Pangram for this would have been around $2,350 at the bulk rate.2 When you sample, you gain some level of confidence a value is within a certain range; the more data the narrower the sample. It works great when you want to report aggregated statistics, but if you’d like to know if an individual paper was generated, and which passages, you are out of luck.

Isogram is a model trained for that case. It’s 1,600× cheaper3 and designed for internet-scale scanning.

Reproducing The Publication Boom

The big risk with a cheaper - likely smaller - model is that you lose discrimination power and it makes more mistakes. To check this, we can compare it to the Paradigm research that has already been done.

First, we’ll run the same experiment Paradigm did - we’ll sample a new random set of 5,275 papers and pick up to 3 passages from each.4 We expect these results to be different (they are a different set of papers after all), but the confidence intervals should be overlapping if we’re reading the same signal.

For January to September 2026, Paradigm’s Pangram series reads about 58% of papers with AI-written passages; our estimated 95% interval is 55–61%.5 They also check 527 pre-ChatGPT papers to confirm that none of them are incorrectly flagged for AI.

Sampled

The full sample covers 5,275 papers from January 2020 to September 2026. It includes 945 papers from 2026, 105 per month from January to September. Isogram flagged 574 of those 945 (60.7%, with a 95% interval of 57.6–63.8%). Paradigm’s published totals imply about 950 papers for those months if it sampled evenly.5

Here we see Pangram and Isogram following a similar curve. Their estimated confidence intervals overlap in every quarter from late 2024. In principle we’d expect the true answer to have a consistent increase every quarter as model capabilities (and hence the likelihood of using them) would likely increase. This would produce a smooth upwards curve. Neither model (at this sample size) shows this confidently, but both don’t rule it out.

This would cost us $4.46 at the served GPU rate. So far, so good. Now time to scale.

Every 2026 paper

Here we analysed all 270,685 new submissions from January to September. 262,177 of them has text to read producing 750,545 passages, about 234 million words and 344 million tokens. The census reads 63.4% over the nine months, against the sample’s 61% [58, 64] and Pangram’s 58%. Month by month it reads 3 to 7 points above Pangram, inside the range Pangram’s roughly 106 papers a month allow.

While we cannot be sure (given the confidence intervals, any of the values from Pangram 4.0 could be within the shaded zone), it seems likely that Pangram is under-reporting given their focus on no false positives. We calibrated Isogram to a false-positive rate of under 1 in 1,000 on human-written text.6 Pangram 4 reports a higher calibration point about 1 in 5,300 on English academic writing and 1 in 24,000 on its general English benchmark.7

This extremely low false positive rate comes at the cost of not detecting some genuine AI content — this is often why people describe it as ‘easy’8 to bypass Pangram. Both Pangram and Isogram can be easily configured without further training to either capture all AI content (at the cost of higher false positives) or never make a mistake (at the cost of allowing more AI content to pass undetected). We chose 1 in 1,000 because we expect Isogram to be a filter cutting down content for further analysis rather than an arbiter.

What it costs

Pangram 4 bills $0.04 per started 100 words of each passage at its cheaper bulk price.9 We include everything in the row for Isogram, failed runs included: which gives us: $13 to fetch and extract the papers, $31 of Modal H100s for January and February before we switched to rented vast.ai RTX 4090s for the rest (GPU, disk and bandwidth together) for $33.

The boom, split

If we now look at the research ‘boom’ in context; you’ll see that really the research boom is an AI submission boom and a human contraction:

By field, region and institution

We can also break it down by field, considering every 2026 paper against Paradigm’s pooled 2026 reading.10 The tick describes the range of values that would be consistent with the sample if they covered all of the dataL

Fields match Paradigm’s ranking - Computer Science at the highest (80%), and Mathematics at the lowest (35%) and largely fall within the same confidence intervals.

Companies seem to be at least 20% more likely to use AI comparative to universities or National Institutes.

Generated or edited

One of the analyses that Paradigm didn’t do, was analyse how much of this is due to people generating passages from scratch as opposed to editing them with LLMs.

To test this we put 120 pre-ChatGPT passages through one of four frontier editors each (GPT-5.6 Terra, Claude Sonnet 5.5, Gemini 3.8 Flash, DeepSeek V4 Pro) with a set of different prompts:

We also had the concern that PDF extraction errors might be impacting results,so we also mimicked PDF extraction errors - either dropping spaces beside numbers and symbols and joined about 4% of ordinary word pairs for example turning “an active” into “anactive”.

Polish text is almost never flagged; about half of heavy rewrites are (53%) and that rate is suppressed by poor PDF extraction. However fully generated content is robust to poor extraction.

This means that the reported rate only counts papers with generated or heavily rewritten passages, not just light proofreading.

A cheap first pass

In total, we spent $77 to fetch the papers and score eight working years worth of prose. Our aggregated data is much the same rise as Paradigm using a far more expensive model. Except here, the price makes it possible to give every submission that first pass.

  1. Marc Brysbaert’s review of reading speeds puts the average for English non-fiction at 238 words per minute. It’s likely worse for AI generated content. ↩

  2. Pangram does offers 200,000 API credits to eligible researchers, enough to score up to 20 million words. The credits expire after three months unless Pangram extends them, and extra credits cost $0.025 each. Pangram asks recipients for project updates and a short summary it can publish. Research grant terms. ↩

  3. This is a mildly unfair comparison because Isogram uses both community rented GPUs, and can run at cost, while Pangram must support an entire research division and so are likely marking up the prices relative to the pure GPU-hours it takes to run. ↩

  4. We drew papers at random within each month from January 2020 through September 2026, using one seed and arXiv’s monthly submission counts. We took 14–15 a month before 2023 and 105–106 a month from 2023, matching Paradigm’s totals of 527 papers from 2020–2022 and 4,748 later. We split each paper’s first PDF into roughly 300-word runs of prose and scored the first, middle and last where available. That gave us 15,148 passages, 4.72 million words and 6.95 million tokens. arXiv submission counts ↩

  5. We read the Pangram series from Paradigm’s charts to within about ±1.5 percentage points. The deck gives the totals of 527 early and 4,748 later papers. We estimate the intervals by spreading each total evenly across its months. The comparison uses aggregate rates because the deck does not publish verdicts by paper. ↩ ↩2

  6. We scored the passages with Isogram v0.10, using the same text verdict as isogram.hails.info and the API at the 0.1% human false-positive threshold. ↩

  7. Pangram 4 model card. Pangram tested 62,971 examples of English academic writing and one million examples of general English text. ↩

  8. It is not. ↩

  9. Pangram’s bulk rate was 20% below list price on 18 September 2026. We priced the measured passages from all 5,275 sampled papers. The Isogram estimate uses the served detector’s measured throughput and all-in rate on a Modal L4. We paid $0.48 for the 6,280 passages that needed new scores. We estimated Pangram’s census cost by multiplying the 2026 sample’s per-paper cost by 270,685. Isogram’s census figure is the Modal and vast.ai bill for fetching the papers and scoring 750,545 passages with the vLLM bf16 build. Pangram pricing ↩

  10. For Pangram’s intervals, we assumed each group made up the same share of its sample as of ours. We inferred regions and institutions from the first pages of papers, which gave us coverage of 61% and 76% of our sampled papers, respectively. ↩