Tokenomics of Agentic Security

"AI writes code now, so why are we still paying for security tools?" If you're a security leader, someone on your board has probably asked you this in the last six months. It's a fair question, and it sounds reasonable right up until you run the numbers. Ask Claude to budget a homegrown AI-powered PR security setup for 150 developers and it comes back with $600K to $1.3M a year in tokens alone. At Claude Fable 5 pricing (double the previous generation), that's roughly $1M to $2.4M. The budget you scoped in January is wrong by June.

So we tested it. We ran 34 realistic AppSec prompts (reachability, exploitability, prioritization, remediation, license compliance) across 13 scopes, through two identical agent setups: same model, same prompts, same step budget. The only difference was context. One agent got precomputed, deterministic evidence from Endor Labs; the other had to work everything out from the raw repo and web searches. The agent working alone burned 79.5M tokens. The one with evidence used 6.6M. That result shouldn't be shocking (give an agent the answer and of course it does better), but the magnitude surprised us: 91.7% fewer tokens, 2.8x faster, 77.6% fewer tool calls. In this webinar, I'll walk through the benchmark, the actual reasoning traces that show where the tokens go, and how to run the same math for your own org.

We'll cover:

1. Why agent workloads cost so much more than chat: An agent re-sends its entire accumulated context at every reasoning step, so agentic workloads burn somewhere between 5 and 30x the tokens of a chat interaction. Design mistakes compound fast; one 50-engineer team got an $8,400 bill in month one from a 50K-token policy injected into every prompt.
2. Where the tokens actually go: The unequipped agent hand-rolled SCA from scratch: 1,985 web searches, 849 greps, 783 manifest reads, about 14 tool calls per question. With evidence, the same questions took about 3 calls each, because a lookup isn't a reasoning problem.
3. Confidently wrong, at 22x the cost: Asked for the highest-severity reachable finding, the agent without call-graph evidence spent 22x the tokens and returned a confident, neatly cited, wrong answer (down to a fabricated version number). Import presence is not reachability.
4. Why the bill is so hard to predict: The unequipped agent's cost tracks your codebase, not your question. The same 34 prompts cost 2.7M tokens on a small project and 10.5M on a large one. With evidence, costs stayed inside a 1.8x band, which is the difference between a budget line and a surprise.
5. Running the math for your org: What an AI reasoning layer covers and what it can't replace (SCA, secrets scanning, and package firewalls aren't reasoning problems), plus how to model token spend against your own developer count and model pricing, so the next time the board asks "can't we build this?", you have a real number.

‍