The million-token context is a billboard. long-context evals updated september 19 show a 30 to 60…
the million-token context is a billboard. long-context evals updated september 19 show a 30 to 60 point accuracy gap between advertised and effective windows past 200k tokens on ruler and nolima.
the model can see it, it just can't use it.
Context
The RULER paper (arXiv 2404.06654) benchmarks 17 models from 4k to 128k tokens and reports that almost all models drop sharply as length grows despite near-perfect needle-in-a-haystack results, with only about half of claim-length models keeping satisfactory performance. The NoLiMa paper (arXiv 2502.05167, v3 of 9 July 2025, ICML) tests 13 models claiming 128K or more and lists claimed against effective length, for example GPT-4o at 128K claimed and 8K effective, with lengths up to 32K in the table read. A secondary page of 17 September 2026 says independent RULER measurements show effective context typically at 50 to 65 percent of the advertised number, and a secondary post of 19 September says there was no standardized cross-provider effective-context benchmark at every depth.
A 30 to 60 point accuracy gap past 200k tokens on RULER and NoLiMa was not found: the papers test to 128K and 32K in the text read, and the secondary pages give an effective-length ratio and not an accuracy gap in points. Updated September 19 is unidentified; the only dated item matching that day is a pricing post, not an evaluation update. The papers show effective lengths below claimed windows for 2024 and 2025 models, and the note blends two different benchmarks under one gap number. The model can see it, it just can't use it is the author's take.
Watch next
- The actual source of the 19 September update, with its model list and depth.
Sources
Provenance
The note above is reproduced unedited from the original post, first published on Threads on 21 September 2026 at 11:56 IST. Sources are the papers and datasets the note draws on.
View the original post ↗Embed this note
More notes
The air is now being asked to keep its own ledger
the air is now being asked to keep its own ledger: ecmwf’s aifs compo becomes the first ai model to forecast atmospheric composition globally every three hours, cleanair simulates 365 days of pm2.5 over china in ten seconds, and a unified framework maps six pollutants at one kilometer across the whole country. the air now files its own composition report.
read the note →The current is now being asked to draw its own map
the current is now being asked to draw its own map: china’s langya 2.0 predicts six ocean phenomena including internal waves and mesoscale eddies, a deep net called wenhai resolves eddies globally with air sea flux formulas built in, and scripps infers surface currents from the way temperature patterns deform in satellite images. the ocean now files its own circulation report.
read the note →The soil is now being asked to report its own carbon
the soil is now being asked to report its own carbon: a nix color sensor paired with generative data augmentation predicts soil organic carbon without a lab, random forest drives 74 percent of soil health mapping studies, and sentinel 2 tracks five year carbon change across france and italy from 922 samples. the dirt now files its own carbon account.
read the note →