Cheaper reasoning models cost more in practice

// 62d agoRESEARCH PAPER

Cheaper reasoning models cost more in practice

A new paper from Stanford, UC Berkeley, and CMU shows that listed API prices for reasoning models are misleading. Uneven consumption of thinking tokens means theoretically cheaper models often result in higher real-world inference costs.

// ANALYSIS

Relying on sticker price for reasoning models is a trap that could blow up your inference budget.

–Evaluated eight frontier models across nine tasks, revealing significant cost mismatches
–Thinking tokens aren't consumed equally, leading to hidden cost overruns
–"Cost reversals" happen frequently enough to change which model is actually cheapest
–Developers need to actively monitor real usage rather than assuming a cheaper API tier saves money

// TAGS

reasoningllmpricinginferenceresearchthe-price-reversal-phenomenon

DISCOVERED

62d ago

2026-03-26

PUBLISHED

62d ago

2026-03-26

RELEVANCE

9/ 10

AUTHOR

Discover AI

// KEEP READING

More AI developer news from the feed

EXPLORE FULL FEED

NEWS1h ago

Dev lets Claude trade BTC overnight, nets $95 profit

A developer gave Claude a $20 budget to autonomously script and execute Bitcoin trades overnight, waking up to a functional trading bot and a $95 profit across five trades.

OPEN SOURCE2h ago

Plannotator 0.19.24 adds Amp support and configurable storage

Plannotator 0.19.24 is a substantial release that expands the tool beyond Claude Code with native Amp support, adds a `PLANNOTATOR_DATA_DIR` override so users can move the default `~/.plannotator` data directory, introduces Auto Mode in the permission selector for newer Claude Code versions, and fixes a Pi approval crash after plan acceptance. The update folds multiple stacked PRs into one release and pushes the project further toward a multi-agent review layer rather than a single-agent hook utility.

NEWS2h ago

Aaronson says AI turns mathematicians into curators

Scott Aaronson says recent AI results in mathematics, including a GPT-5.5 Pro solution to Erdős’s Unit Distance Problem, suggest humans may increasingly focus on choosing questions and interpreting model outputs. He extends the argument to AI-written fiction and the Vatican’s AI encyclical as signs of a broader cultural shift.