Fewer tokens.Measured results.
We measure Aether Compress not only by how much it compresses, but also by how much necessary information it preserves while compressing. Our results are evaluated with large-scale local tests and independent model validation.
Two different layers of testing.
Local verification tests for information loss at scale. External model validation is the model generated from the optimized context. It measures whether the answer carries the same information as the original context.
Ortalama40.65% less.
The measured value indicates that the optimized input is equal to the original input. It shows the average decrease in the amount of tokens.
When evaluating Aether Compress, only token reduction is not measured. The primary success criterion is that the optimized context preserves the necessary information.
In the external validation layer the same question is valid for both the original and evaluated on the optimized context and It is measured whether the answers retain the necessary information.
A single content typeWe do not measure according to.
The test set has different density, structure and context properties. It consists of scenarios with
Same measurement, repeatable process.
Benchmark process original input, optimized input and measure the verification result separately from each other.
The test input is taken as an unchanged reference.
The same input is processed with Aether Compress.
Token reduction and information preservation are measured separately.
Selected tests are also checked via the external model.
Measurement limits
The reduction rate varies depending on the structure of the input and the amount of unnecessary information it contains.
This value is the average processing time measured in the validation environment. It may vary depending on hardware and workload.
This value represents the highest throttling observed in the validation set and is not the expected rate for each request.
The 100% value in external validation is the result of 500 tested scenarios and does not imply a guarantee for unlimited all possible inputs.
More important than the benchmark,your input.
Try Aether Compress for free and see how many tokens you can reduce in your own real context.