Talanton (τάλαντον)

Open-source AI cost tracking, pre-flight token metering & LLM budget guardrails

v0.2.0 on PyPIpythonsub-2ms guardrailslocal-first60+ modelsPyPI (talanton-py) →Live Docs →Repository →

Talanton — Classical Engraving of Icarus: Preventing Runaway AI Overspend

FIG. 01 // WHY TALANTON EXISTS — STOPPING ACCIDENTAL $1,000+ AI BILLSSUB-2MS GUARDRAIL

The Core Problem

When building modern AI-driven applications, prompt costs silently compound. An unmonitored while-loop in an autonomous agent, an unexpected recursion bug, or an employee accidentally submitting a 100-page PDF into an expensive model can burn through thousands of dollars over a single weekend.

Existing monitoring tools (such as Helicone or Langfuse) require proxying all requests through third-party remote clouds. This introduces 40ms to 120ms of artificial network latency to every prompt and forces proprietary user data across the open internet.

The Solution: Local-First Budget Guardrails

Talanton (τάλαντον) is named after the ancient Greek balance scale of honest weight. It runs 100% in-process on the local machine with sub-2ms latency and zero cloud data exfiltration.

Before any request is sent to OpenAI or Anthropic, Talanton meters the exact token count, calculates cost in real time across 60+ frontier models, and compares the projected spend against pre-configured soft warnings and hard circuit-breaker limits.

from talanton import TalantonTracker, BudgetGuardrail
from talanton.guardrails import raise_exception

tracker = TalantonTracker()

# Stop calls if the daily budget hits $10.00
guardrail = BudgetGuardrail(
    tracker=tracker,
    soft_limit=5.00,
    hard_limit=10.00,
    period="day",
    on_hard_limit=raise_exception
)

# Checks spend in 0.08ms before any outbound API call is dispatched
result = guardrail.check(estimated_cost=0.04)
if result.is_blocked:
    print(f"Call blocked! Daily spend reached ${result.current_spend:.2f}")

Drop-in 1-Line SDK Wrappers

Developers can wrap existing standard OpenAI and Anthropic clients with a single line of code. Every chat completion is metered, recorded to an embedded SQLite WAL ledger, and enforced against spending limits with zero boilerplate:

import openai
from talanton import TalantonTracker, BudgetGuardrail
from talanton.integrations import TalantonOpenAI

tracker = TalantonTracker()
guard = BudgetGuardrail(tracker=tracker, hard_limit=25.0, period="day")

# 1-Line Drop-in Client Replacement:
client = TalantonOpenAI(openai.OpenAI(), tracker=tracker, guardrail=guard)

# Standard OpenAI syntax — automatically guarded & tracked:
response = client.chat.completions.create(
    model="gpt-4o",
    messages=[{"role": "user", "content": "Analyze quarterly filings"}]
)

Performance & Scalability Benchmarks

Measured under 25-thread concurrency on Python 3.12 with SQLite WAL persistence:

Pre-Check Latency (P50): 0.08 ms (1,000x faster than cloud proxies).
Pre-Check Latency (P99): 1.94 ms (strict sub-2ms SLA).
Ingestion Throughput: 24,586 calls/sec with zero lock contention.
Data Privacy: 100% local-first — customer prompts never leave your infrastructure.
Model Registry: Exact pricing and tokenization for 61 models (GPT-4o, Claude 3.7 Sonnet, Llama 3.3).

Talanton (τάλαντον)

The Core Problem

The Solution: Local-First Budget Guardrails

Drop-in 1-Line SDK Wrappers

Performance & Scalability Benchmarks

Project Stats

Quick Install

Distribution & Links

Stack & Ecosystem