The Core Techniques Behind Token Savings

Prompt caching allows developers to store and reuse previously processed context, so repeated or similar requests draw on cached material instead of paying full token charges again. Prompt compression, by contrast, trims input text before it reaches a model, directly reducing the number of tokens a request consumes. Both are established, officially supported paths to lower token spending, and the review confirmed their standing as central technologies in this space. Representative compression efforts include Microsoft's open-source LLMLingua project, whose materials were examined during the review.

An Equally Official Route: Model Selection

The claim is not exhaustive, however. The review found that model selection — directing workloads to smaller or more cost-efficient models — operates as an equivalent, officially sanctioned means of reducing token costs. Caching and compression are therefore core, but they do not hold a monopoly: formal alternatives stand alongside them as legitimate ways to contain spending.

Fewer Providers Behind Compression

A further asymmetry concerns supply. Caching is available from a broader range of providers, while compression traces back to a more limited group of suppliers. For teams weighing the two techniques, this means compression, though a core technology, may require third-party tooling or in-house implementation rather than arriving as a first-party option from every provider.

Verdict

On the strength of the evidence, the claim that prompt caching and compression are the key technologies for token cost reduction stands as Mostly True — accurate on its central point, but incomplete without noting that model selection coexists as an equally official lever and that compression reaches users through a narrower set of providers than caching.