AI bills are like restaurant bills with no itemization — you only see the total. We break it down per agent, per user, per endpoint. Then tell you where to cut.
Free tier: 100K events/mo. No credit card. Open source SDK (MIT).
A prompt change, a runaway loop, a new feature — and you find out only at month-end. By then the damage is done.
Your provider dashboard shows one big number. Was it the research bot? The classifier loop? The customer support agent? You're guessing.
Should this task run on Haiku or Sonnet? Is prompt caching even working? Without per-call attribution, every change is a coin flip.
Drop-in SDK wraps your Anthropic, OpenAI, or any provider client. Every call gets logged with cost, latency, model, agent name, and user ID. No data leaves your infra unless you opt in.
from acr import track client = track(Anthropic(), agent="research-bot") # Use normally — every call is metered response = client.messages.create(...)
Slice and dice by agent, endpoint, user, customer tier, or any custom tag. Find the 3 calls eating 40% of your budget. See cost-per-customer if you're a B2B SaaS.
We analyze your usage patterns and tell you exactly where Haiku will replace Sonnet without quality loss. Where prompt caching is leaking. Where Batch API gets you 50% off.
| Feature | Cost Radar | Helicone | LangSmith |
|---|---|---|---|
| Starting price | $29/mo | $80/mo | $39/mo (LC only) |
| Free tier | 100K events/mo | 10K events/mo | 5K traces/mo |
| Open source SDK | Yes (MIT) | Partial | No |
| Multi-provider | Yes | Yes | LangChain only |
| Routing recommendations | Yes | No | No |
| Self-host option | Yes (Enterprise) | Yes | No |
| Setup time | 3 lines | ~10 lines + proxy | ~30 min |
Pricing accurate as of 2026-05. Compare directly — links in docs.
Need self-hosted or >5M events/mo? Talk to us about Enterprise ($500+/mo). Annual plans get 20% off.
"Used in production: internal developer tooling — 50 agents, ~1,840 events/day. Caught 3 cost regressions before they hit the bill."
No. By default we only log metadata: model, token counts, latency, cost. Your prompt and response content never leaves your infra. Opt-in content logging is available for debugging if you want it.
Anthropic, OpenAI, Google (Gemini), Mistral, Groq, and any OpenAI-compatible endpoint (local Ollama, vLLM, etc.). Adding a provider takes ~50 lines — PRs welcome.
Yes — Enterprise plan ships a Docker compose stack (Postgres + the dashboard). Air-gapped deployments supported. Talk to us for licensing.
We analyze your call patterns — task type, output structure, latency tolerance, observed quality signals. Then we suggest concrete swaps (Opus → Sonnet, Sonnet → Haiku) with a confidence score. You always approve before any change is enforced; we never change your code.
Account drops back to Free tier — your SDK keeps working, you keep 14 days of history, no data is deleted for 90 days. Full export available anytime as CSV / JSON.
Three lines of code. Five-minute setup. Then never get surprised by an Anthropic bill again.