Skip to main content
Token costs have a way of surprising teams that don’t track them closely. A retry loop that fires three extra times per request, an agent that sends unnecessarily large prompts, or a single expensive model standing in for work a smaller model could handle — these patterns are invisible until the bill lands. Veritrix surfaces cost at every level of granularity, from an individual llm.call span all the way up to the total daily spend for an entire agent, so you can spot the runaway loop before the invoice arrives.

​
What Cost Tracking Shows

Veritrix attributes cost at two levels:
  • Per-span cost — every llm.call and llm.retry span carries an inferred dollar cost calculated from its input_tokens, output_tokens, and the current pricing for the model that was called.
  • Per-agent cost — the aggregate spend for all LLM spans that ran under a given agent’s root span, rolled up across any time window you choose.
Together, these two views let you answer both micro questions (“why did this one trace cost $0.40?”) and macro questions (“which agent is driving 80% of our monthly spend?”).

​
How Costs Are Calculated

Veritrix calculates cost from the token counts recorded on each LLM span:
Model pricing is kept up to date by Veritrix automatically — you never need to hard-code token prices or update a configuration file when a provider changes its rates. The model name is captured from the LLM call itself, so multi-model agents are costed correctly even when different agents in the chain use different models.
Cost figures in Veritrix are estimates based on published provider pricing. Actual charges on your provider invoice may differ slightly due to rounding, promotional pricing, or enterprise agreements.

​
Cost Fields on Spans

The following fields on llm.call and llm.retry spans feed directly into cost calculations: You can inspect these fields by clicking any LLM span in the trace waterfall. The inferred_cost is displayed prominently alongside the token counts so you always know what a single call contributed to the total.

​
Aggregate Views

The Cost dashboard aggregates span-level data into time-series charts and per-agent breakdowns. For example:
Research analyst agent — Total spend $3.00 (past 7 days)
  • 1,240 llm.call spans
  • 2.1M input tokens, 890K output tokens
  • Average cost per trace: $0.18
You can slice this view by:
  • Agent — compare spend across all agents in your workspace side by side.
  • Model — see how much each model (e.g. GPT-4o vs. GPT-4o-mini) is contributing to cost.
  • Time window — choose today, the last 7 days, last 30 days, or a custom range.
  • Span type — isolate the cost of retries specifically to understand how much failed calls are costing you.
Sort the per-agent cost table by average cost per trace rather than total spend to identify agents that are expensive per-run, not just agents that run frequently. These are often the best candidates for prompt optimisation or model downgrading.

​
Alerting on Cost Anomalies

Unexpected cost spikes — a retry loop, a prompt that suddenly became 10× larger, a new code path that calls an expensive model — show up in Veritrix before they show up on your provider dashboard. You can configure cost alerts to notify your team when spend crosses a threshold. See the Alerting page for full details on configuring cost alerts, including per-agent thresholds, percentage-change triggers, and notification channels.

​
Fine-Tuning Cheaper Models on Saved Completions

Every completion Veritrix records is available for export. If a particular agent is performing a narrow, well-defined task — classification, extraction, formatting, code review — you can fine-tune a smaller, specialised model on those saved completions and replace the general-purpose model entirely. Teams that do this consistently report cost reductions of up to 25× on the fine-tuned agent compared to running the same task on a frontier model, with equal or better accuracy on the specific task.
1

Identify a high-spend, narrow-task agent

In the Cost dashboard, look for agents with a high volume of similar LLM calls — the same prompt structure, consistent output format, low variance in task type. These are the best candidates for fine-tuning.
2

Export saved completions

Navigate to the agent’s detail page, click Export completions, and choose a date range. Veritrix exports the completions in JSONL format, ready for use with any fine-tuning provider.
3

Fine-tune your model

Use the exported completions to fine-tune a smaller base model with your preferred provider. Veritrix does not lock you into a specific fine-tuning service.
4

Update the agent and compare costs

Deploy the fine-tuned model, point your agent at it, and let Veritrix measure the cost difference. The before/after comparison is visible in the Cost dashboard once the new model has accumulated enough runs.

Traces & Spans

Learn about the token and cost fields recorded on every LLM span.

Alerting

Set up cost anomaly alerts so your team knows the moment spend spikes unexpectedly.