The cheapest model is a trap about a third of the time. a stanford-led study found cheaper ai…
the cheapest model is a trap about a third of the time. a stanford-led study found cheaper ai models cost more in 32% of comparisons, once quality, retries and reasoning tokens are priced in, per oct 5.
sticker price is the least reliable part of an llm bill.
Context
A study led by Lingjiao Chen of Stanford, 'The Price Reversal Phenomenon', tested 8 frontier reasoning models across 12 tasks and found the model with the lower listed price cost more in 106 of 336 head-to-head comparisons, about 32%.
Coverage cites a case where Gemini 3 Flash was listed 80% cheaper than GPT-5.4 but cost 38% more in practice. The study attributes this to differences in how many reasoning tokens models use.
The paper is on arXiv from March 2026, with a v2 dated May 28. News coverage ran October 5, so 'per Oct 5' is the press date.
The tests price reasoning tokens. Retries and quality adjustments are not part of the 32% in the sources read, so that part of the note is unsupported here, not refuted.
'Sticker price is the least reliable part of an LLM bill' is the author's opinion.
Related work
- arXiv: The Price Reversal Phenomenon ↗The paper with the 8 models and 12 tasks.
- The Clarity: cheaper-listed models cost more in 32% of tests ↗Reports 106 of 336 comparisons.
- Implicator coverage ↗Gives the Gemini 3 Flash and GPT-5.4 example.
Watch next
- Read the paper's method to see exactly which costs are counted.
Sources
- arXiv 2603.23971 v2, May 28, 2026arxiv.org
- The Clarity, Oct 5, 2026theclarity.today
- Implicator, Oct 5, 2026implicator.ai
Provenance
The note above is reproduced unedited from the original post, first published on Threads on 9 October 2026 at 19:09 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 →