Tune thinking effort instead of swapping models

This rundown takes apart the announced metrics of a frontier lab's new top-tier model item by item, then tests them against four outputs made from the same prompt. The change singled out as most important is not the scores but the range of thinking effort: from low to max, results and costs spread out widely, and the interpretation offered is that this makes possible a design in which, instead of swapping models, you use only as much of the same model as each task needs. Per 1 million tokens, input pricing falls from $5 to $4 and output from $25 to $20, cache reuse comes to less than half the regular input price, and the newly added fast mode is summed up as a trade-off of 2.5 times the speed for twice the price. From this comes the implication that caching the long documents sent along with every call is not just a cost saving but a question of architectural design. It does, however, concede outright that long-form Korean output has awkward sentences mixed in and is not at a level that can be handed off without review.

This is a note: my own write-up after reading several sources. It has no source link; events checked against a source are kept as records.