How much of your AI plan Bravos uses

Threat model and pentest

The annotation campaign wrote 6,564 annotations across 1,618 files, plus test templates for what it found: about 336K tokens and 2.2 model calls per file, 97% of the tokens re-read from the agent's cache. The pentest session tested 10 findings (9 confirmed): about 7.9M tokens per finding. A saved test you replay later uses no AI.

Annotating as you code

In that same session the agent changed code, and Bravos sent it back 3 times to annotate what it had changed. Those replies took 0.7% of the session, because the code was already in the agent's context. One session is a small sample, so read it as an order of magnitude.

From work to a plan's limit

Each bar is the work weighed at the provider's own rates, divided by the plan's allowance. Every bar says where its allowance comes from:

• Measured: our Claude Max weekly meter read 85% after 3.2B tokens that week. Use from other devices was not in our records, so it errs high.
• Published limit: the provider publishes the allowance and how usage counts against it.
• Scaled: the provider publishes a multiple of another plan, such as 5x Pro.
• Estimate: the provider publishes no size, so we assume the same allowance per dollar as our measured Claude Max. We replace these as we measure each plan.