Feature  ·  Cost Management

Cut your LLM bill before it cuts your runway.

Attribution across 5 dimensions, 30-day forecasting with confidence bands, and AI-generated optimization recommendations with savings estimates.

Works automatically — no extra setup beyond wrapping your LLM clients.

Start free →View docs
Zespan Cost Management
Works withAgent attributionModel attributionUser attributionCost forecastingAI recommendationsQuota enforcement

30 days

forecast horizon

5

attribution dimensions

Top 200

per dimension

1.0  Cost Management

5-Dimension Attribution

What you get

Dimensions: agent, tool, model, user, operationTop 200 entries per dimension for any time rangeTime ranges: 24h, 7d, 30d, 90d

Slice LLM spend by agent, tool, model, user, or operation. For each dimension: total cost, call count, cost per call, average latency, input/output tokens, and error rate. Switch between 24h, 7d, 30d, and 90d windows.

Zespan cost attribution view showing spend breakdown by agent and model

2.0  Cost Management

30-Day Forecasting

What you get

Linear regression over daily cost historyConfidence band: low and high projectionSeasonality detection: per-day-of-week average for scheduling insights

OLS linear regression over daily cost history projects spend 30 days forward with low/high confidence bands. Trend classification (increasing/decreasing/stable) and seasonality detection (day-of-week patterns) surface early warnings.

Zespan 30-day cost forecast with confidence band and trend classification

3.0  Cost Management

AI Cost Recommendations

What you get

Types: model switch, prompt compression, batching, cachingProjected savings: monthly USD estimate per recommendationCode snippet included: copy-paste to implement

Zespan analyzes your trace data and generates ranked, actionable recommendations — each with projected monthly USD savings, confidence level, effort rating, and a copy-paste code snippet to implement it.

4.0  Cost Management

Anomaly Detection & Quota

What you get

Background anomaly detection: flags cost/latency/error spikes automaticallySoft mode: allows overage up to capMultiplier × limit (default 5×)Hard mode: strict cap — free plan always runs hard mode

AI anomaly detection flags unusual cost spikes in real time — catching them the day they happen, not the day the invoice arrives. Per-org monthly quotas enforce spending limits with soft mode (overage allowed up to 5×) or hard mode (strict 429).

Zespan cost explorer showing daily spend trend with anomaly detection

Setup

Under 5 minutes,
two lines of code.

No forking and no architecture changes. Traces appear within seconds of the first agent run, with cost attribution, eval scores, and anomaly alerts on by default.

typescript
// No extra setup — cost tracking is automatic when you wrap your LLM clients.
import { Zespan, wrapOpenAI, wrapAnthropic } from '@zespan/sdk';

const lt = new Zespan({ apiKey: process.env.ZESPAN_API_KEY });

// Every call now attributed by agent, model, user, and operation
const openai = wrapOpenAI(new OpenAI(), lt);
const anthropic = wrapAnthropic(new Anthropic(), lt);

Common questions

How does Zespan know the cost of each LLM call?

The SDK reads token counts from the LLM provider's API response and looks up current pricing for that model. Cost per call is calculated as (input_tokens × input_price) + (output_tokens × output_price). Pricing tables are updated when providers change their rates.

Can I attribute cost to specific product features, not just agents?

Yes. Pass an operation tag in the metadata when making LLM calls — e.g., operation: 'summarize-ticket' or operation: 'generate-reply'. The cost attribution view lets you slice by operation to see which product features drive the most spend.

What's the difference between soft mode and hard mode quota enforcement?

Hard mode strictly caps ingest at the monthly quota limit — any request over quota returns 429. Soft mode allows overage up to a configurable cap multiplier (default 5×) before hard blocking, and overage is billed. Free plans always use hard mode with a 1× multiplier.

How actionable are the AI recommendations?

Each recommendation names the specific agent or operation, explains why the change will help (with supporting evidence from your trace data), gives a projected monthly USD savings figure, and includes a copy-paste code snippet. You can mark them applied, acknowledged, or snooze them.

Explore more features

All features →
TracingSee exactly what your agents are doing.Every LLM call, agent step, and tool invocation captured as a structured trace — with cost, latency, and tokens per span.Agent MonitoringKnow which agents are healthy, and which aren't.Composite health scores, delegation graphs, and per-agent cost attribution — built for systems with many cooperating AI agents.EvaluationsMeasure output quality on every trace, automatically.12 built-in LLM-as-judge templates run on every new trace with no setup. Track quality trends, catch regressions, and run manual eval campaigns.GuardrailsStop bad outputs before they reach users.7 guardrail types run inline on every LLM request — block, warn, redact, or log. PII, toxicity, topic drift, format, cost ceiling, and custom rules.

Your agents are running.
Do you know what they’re doing?

Observe, evaluate, guard, and control every agent, live in minutes. No credit card required.

Start free →Talk to us →
Free tier availableUnder 5 min setupNo credit card