If you use Claude mostly as a chat box and coding helper, and you click big AI news just to ask whether you are getting a better tool or a tighter set of limits, this is the part worth your time. The easiest mistake is to treat Claude Code and Claude as the same tool and assume the one with the higher score is automatically the better fit. My read is simpler: prompts have become an identity layer, not just a meaning layer.
That matters because the annoying part of launch coverage is not the hype. It is finishing a whole thread and still not knowing whether your daily workflow changed. You came in to see whether the model got stronger, then realized the more important story was the tradeoff nobody put in the headline. The most discussable thing is never that the model got stronger. It is why the strongest version was not put on the table directly.
The leak on March 31, 2026 is why this angle has weight at all. Reports said nearly 2,000 files and about 500,000 lines of Claude Code source were exposed. That does not prove current production behavior, but it does mean people could see implementation choices instead of just marketing claims. One of those choices points straight at request identity.
According to the leaked code summaries, sideQuery places an attributionHeader into a separate system text block, basically a hidden instruction section the user does not see, so the service can parse cc_entrypoint. A related constant, getAttributionHeader, builds a string like x-anthropic-billing-header: cc_version=...; cc_entrypoint=...;. The important detail is that this appears as prompt text, not as a normal HTTP header living outside the prompt.
That is why Claude Code Is Steganographically Marking Requests is more than a dramatic headline. The request is not only carrying what you asked. It is also carrying where the request came from and how the system should treat it. In plain English: prompt space is now doing protocol work. It helps with identity, billing, and routing, not just language understanding. That changes how you should read product differences between tools that look similar from the outside.
So the practical takeaway is not to turn this into a paranoia story. The valuable judgment is smaller and more useful: if you only read launch copy, you can think you bought a stronger assistant when the first thing you may actually hit is a tighter boundary. That is also why my working split is this: Claude Code is better at helping you see the problem first; Claude is better at finishing the downstream work. This reading is based on leaked source code, not a live product test, so it may not match current production. If you know someone still treating Claude Code and Claude as the same product, share this with them before they argue about which one is better.