Transcription
If you use Codex heavily, a lot of your spend goes to one thing: re-sending context the model has already seen. Because it’s how it works with coding agents. We wanted to measure how much waste that creates, and whether Edgee can remove it.
But first, what is Edgee? Edgee is an AI Gateway for Coding Agents. It's really simple to use, you only need two command lines. First, to download Edgee CLI. With curl or Brew… And the second one to execute your favorite coding agent, for instance: edgee launch codex, or edgee launch claude… so your favorite coding agent will be launched through Edgee.
So, for the test, we ran the same benchmark twice on the same repository, with the same model and the same instructions: one session with plain Codex, and one session with Codex routed through Edgee's compression gateway. Then we compared. Here you’re seeing the two sessions run side by side. Both are going through the exact same workflow: first Codex enters plan mode, then it executes the plan, and we repeat that for each instruction. We launched 6 instructions in total. If you want to know more about how the tests are working, feel free to look at the source code, the repo is accessible.
Let's speed up the video now. We kept the test simple on purpose. Same codebase, same task sequence, same model. The only difference is that the session on the right is using Edgee compression and the other one is not. What Edgee changes is not the task itself. It changes how much redundant context gets sent back to the model on every turn. Edgee compresses tool results. For example, if codex asks your computer to execute the cargo build or git log command, the model won't need as much verbosity as these commands naturally generate... so Edgee intercepts these results and makes them lighter.
We cache the compression results in our distributed KV store… and then, it's pretty magical, at the end of the day, it will consume fewer tokens. And that effect compounds across the session. As the conversation grows, the compressed session stays leaner, which is exactly what we want to measure here. And the results are Codex with Edgee used 50% fewer input tokens, and it improved cache hit rate from 76% to 85%, and reduced total session cost by 35.6%.
So the takeaway is simple: Codex + Edgee is way more efficient, more frugal, and significantly cheaper than Codex alone. Stop Paying Codex to Re-Read Context. Same model, same tasks, less waste. That’s the whole point of Edgee. And the good news is: Its free!