Cost dashboards tell you what your agents wasted last week. ThumbGate stops the waste before it fires.

Most AI-spend platforms — Finout, Vantage, the new "AI FinOps Assistant" wave — focus on showing you the bill after the agent ran: cost allocation, anomaly detection, unit economics finance can trust. A few (Helicone's rate limits, Revenium's Economic Control) add coarse runtime enforcement keyed off $ thresholds or request counts. None of them stop the wasted tool call by understanding why it would have failed.

That's the layer ThumbGate occupies. Every PreToolUse gate that fires is a Claude / GPT call your agent did not make — input tokens you didn't spend, output tokens you didn't spend, retry loop you didn't trigger. The savings are computable, conservative, and now surfaced as a number on your CLI.

"Measurable" is the operative word. A token-spend dashboard tells finance how much got burned; it doesn't tell the CIO board what was averted. thumbgate cost prints a single conservative dollar figure backed by the gate-block count from your machine — not "what enterprises like you saved." That's the artifact that survives a 2026 budget review.

One command, one number

Once ThumbGate is installed and gates have been firing, this is what an operator sees:

$ thumbgate cost

💰 ThumbGate cost-savings — cumulative
──────────────────────────────────────────────────
  Tool calls blocked : 247
  Tool calls warned  : 12
  Tool calls passed  : 3,401
  Top blocker        : no-mocked-db (138 blocks)

  Tokens you did NOT spend
    Input  : 494K
    Output : 148K
    Total  : 642K

  Estimated $ saved  : $3.95

The methodology is intentionally conservative: 2,000 input + 600 output tokens per blocked call, a Sonnet-heavy model mix (80% Sonnet 4.5, 15% Opus 4.6, 5% Haiku 4.5), Anthropic published prices. The goal is "you almost certainly saved at least this much" — not "let's flatter ourselves." Override the mix with --mix '{"claude-sonnet-4-5":1.0}' if your stack is different.

Prevention vs. reporting

The two layers compose. ThumbGate prevents the wasted spend; a reporting FinOps tool tells finance what the remaining spend was for. Picking ThumbGate doesn't mean you don't also need cost visibility — it means the visibility number gets smaller.

Why the savings are real, not theoretical

1. Every block is one fewer round trip. A blocked tool call doesn't reach the model. There's no "ThumbGate intercepted but the request still cost you" — the agent's tool-call execution is replaced with the gate's verdict, and the agent's next reasoning step takes the verdict as context instead of the failed result.
2. The avoided retry loop is the bulk of the saving. Failed tool calls don't just cost the call — they cost the model's next reasoning turn (which sees the failure and tries again), and often a third turn (which tries a different approach). Conservative 2k input + 600 output assumes one retry; in practice it's often more.
3. The numbers come from your local gate-stats.json. Not from a marketing model, not from "what enterprises like you saved." Your machine, your gates, your blocks.

Get the number on your machine

npx thumbgate init        # wire the PreToolUse hook
# ...let your agent run for a few hours...
npx thumbgate cost        # see what the gates were worth

Or as JSON, if you want to ship it to a dashboard:

npx thumbgate cost --json | jq .savings.dollarsSaved

Related reading

• ThumbGate for Codex in the Enterprise — the same prevention story for the OpenAI×Dell distribution wave.
• ThumbGate for the Agent Manager — the role inside the org that owns "what are our agents costing us."
• Dashboard demo — the $ saved number rendered against demo data, so you can see the shape before installing.