Cost per task for a given LLM capability falls from $1.22 to $0.022, a 56x drop

The cost per task for a model with a fixed intelligence score has fallen from $1.22 in February to $0.022 today, a 56x reduction. This decline means that running the same model ten thousand times now costs a little over a hundred dollars. Previously such a run would have required several thousand dollars, putting large-scale analysis out of reach for many projects. As a result, workflows that need to examine every item in a large collection become affordable.

Engineers often choose a model by balancing its capability against the price of each call. When the price floor drops, the same capability can be obtained for far less, shifting the trade-off toward using cheaper models more frequently. In agentic coding, for example, the improved quality of work used to justify higher cost, but now a less expensive model can be called many times to achieve similar productivity. The change makes it viable to replace occasional sampling with exhaustive checks.

The lower cost does not eliminate scenarios where the highest capability models are still needed. Tasks that demand the very top of the intelligence index, such as solving novel mathematical proofs, may still require the most expensive frontier models. Additionally, factors like latency, model size, or specific strengths not captured by the intelligence index can limit the usefulness of cheap models. Thus the cost floor improves volume work but does not remove the ceiling’s role for specialized, high-stakes problems.

The analysis relies on data from Artificial Analysis and the author’s own observations. No other feeds provided alternative viewpoints, so the interpretation rests on a single source. Readers should treat the projected 100x drop timeline as an extrapolation rather than a guarantee.