Finic

A decision model fast enough for the real-time path, and agents that do the careful investigation.

[Finic @finicdotai]

$FINIC is live on @ponsdotfamily.

CA : 0x498b5b0d72f6f570e7cdbdc4c05824675eaf9a63

Frontier AI to stop fraud, abuse, and financial crimes. Fast enough for the real-time path, careful enough for the investigation after it.

[Finic @finicdotai]

More to come.

Stay tuned, guys!

[Finic @finicdotai]

Every cohort gets a static metric and a flow metric, split into mitigated vs. unmitigated, side by side.

[Finic @finicdotai]

A 200-user cluster that's already been handled matters less than a 5-user account takeover ring with $500K still live.

Finic's fraud skills rank cohorts by exposure, not headcount.

[Finic @finicdotai]

Every session gets one workspace: each query saved with its result, numbered in run order, and a single investigation.md with the verdict at the top.

Clusters only expand on strong links, or several moderate ones together.

[Finic @finicdotai]

How Finic's fraud investigation skills route your request.

Describe the work in plain language. SKILL.md checks your reference docs, classifies the phase, and sends it to hunting, triage or investigation.

https://t.co/gNdCRORYgv

[Finic @finicdotai]

Password reset two minutes after the new device logged in. Was it initiated by the actual customer? Grayson: No, 89%. Confidence 78%.

[Finic @finicdotai]

Finic Grayson is in early beta: the first universal decision model built for real-time decisions across fraud, risk and financial crimes.

~0.1 s model time for typical requests. $0.035 per million input tokens. Zero data retention available on all plans.

[Finic @finicdotai]

7/7 Finic also open-sourced the fraud investigation skills for AI agents: hunting, triage and investigation, the way real fraud teams work. Works with Claude Code, Claude Cowork and Codex.

https://t.co/gNdCRORYgv

[Finic @finicdotai]

Finic's fraud investigation skills for AI agents are open source.

Drop them into Claude Code, Claude Cowork or Codex and your agent works like a real fraud team: hunting, triage, and investigation down to a written verdict.

https://t.co/gNdCRORYgv

[Finic @finicdotai]

It ships linkage rules, a library of common fraud typologies and the signals each one produces, and a standard workspace layout so every session can be audited later.

Add your own schema and signals under references/.

[Finic @finicdotai]

Out of the box, agents asked to investigate fraud run unstructured queries and flag behavior that's unusual but not actually fraud. These skills add the structure real fraud teams work with.

[Finic @finicdotai]

Three phases: hunting finds new patterns, triage ranks suspicious data into prioritized alerts, and investigation drives a cohort to a written verdict.

Describe the task in plain language and the skill routes it.

[Finic @finicdotai]

Got a use case in mind? Request access to Grayson.

[Finic @finicdotai]

Supply burned.

https://t.co/pRtL0gEXjw

[Finic @finicdotai]

Mean AUC-PR across 6 fraud and risk benchmarks, public and Finic's own. Costs at API list prices. Read the Grayson announcement on https://t.co/wHXa8O5ozo for the details.

[Finic @finicdotai]

Score vs. cost, in one chart 🧵

[Finic @finicdotai]

Claude Sonnet 5.5: 0.540 at $11.75. Jev: 0.470 at $0.16.

[Finic @finicdotai]

Finic Grayson: 0.569 mean AUC-PR at $0.11 per 1,000 requests.

[Finic @finicdotai]

Every team makes two kinds of decisions.

System One: real time. Approve or decline before a payment clears.
System Two: deliberate. Follow the evidence until a case can be closed.

Finic builds AI for both.