Usage Logs

Usage logs track token consumption and cost per agent run. Use them to understand which agents and workflows drive the most LLM usage.

Viewing usage logs

Open Runs › Usage.

The screens, step by step: Runs and what they cost, in the Guides.

What's tracked

Metric	Description
Total tokens	Input + output + cache read + cache write tokens per run
Input tokens	Cache-miss prompt tokens (billed at the full input rate)
Output tokens	Tokens generated by the model
Cache read tokens	Prompt tokens served from the prompt cache at a discounted rate
Cache write tokens	Prompt tokens written to the cache on first use (Anthropic / Bedrock only; one-time surcharge)
Estimated savings (USD)	What this call would have cost without the cache, minus what it actually cost
Response cache hit	`true` when the LLM call was skipped entirely via the response cache
Cost (USD)	Estimated cost based on the model's pricing, after cache adjustments
Model	The LLM used for the run
Agent / Workflow	What triggered the run
Session ID	Conversation session
Timestamp	When the run occurred

Cache tracking

Every usage record carries four cache fields alongside the usual token counts:

Field	Meaning
`cache_read_tokens`	Tokens served from the prompt cache. Billed at ~10–50% of the input rate, depending on provider
`cache_write_tokens`	Tokens written into the cache on a first hit. Anthropic / Bedrock charge ~1.25× input for this; OpenAI / DeepSeek / Gemini do not
`estimated_savings`	Per-call USD saved vs. paying full input rate for the `cache_read_tokens`
`response_cache_hit`	`true` when the entire LLM call was satisfied by the response cache (no provider call was made)

On a self-hosted install these are summed in the Cache analytics section of Usage:

Saved — sum of estimated_savings across all calls
Cache reads / Cache writes — workspace-wide totals
Response hits — LLM calls served entirely from the response cache, out of all calls, with the rate
A table by Model — cache reads, response hits and savings per model, so you can tell which models actually benefit
Top workflow runs by savings — per-run breakdown, useful for cost-attributing a specific pipeline

Sample log record with cache activity:

{
  "timestamp": "2026-05-23T14:02:11.873Z",
  "model": "claude-opus-5",
  "provider": "anthropic",
  "session_id": "sess-abc123",
  "agent_id": "agent-researcher",
  "source": "chat",
  "input_tokens": 412,
  "output_tokens": 187,
  "cache_read_tokens": 8421,
  "cache_write_tokens": 0,
  "total_tokens": 9020,
  "estimated_cost": 0.00408,
  "estimated_savings": 0.02273,
  "response_cache_hit": false,
  "latency_seconds": 1.84
}

For a response-cache hit, input_tokens and output_tokens reflect the cached completion's original token counts, response_cache_hit is true, and estimated_cost is 0. See Caching for the full design.

Cost tracking in workflows

Workflows track cumulative cost across all steps:

{
  "total_tokens_used": 15234,
  "total_cost_usd": 0.0234
}

max_total_cost_usd can be set on a workflow, but it is not enforced — nothing reads it, so a run is not halted when the figure is passed. Usage is recorded either way. See Cost limits.

Knowledge

A knowledge base's indexing and searches are logged under a session of their own — knowledge: and the base's id, one per base — and a Retrieve step's search under its run, as part of its cost. On SynapseOrch Cloud, Usage names the base instead, with drafted questions and Evaluation run for Evaluate. See Costs and budgets.

Cost estimates

Cost estimates are calculated based on published pricing for each model. They are approximate — actual costs may differ due to caching, batch discounts, or pricing changes. Always verify with your LLM provider's billing dashboard.

A model with no price on the rate card spends $0 as far as these figures know, so it is marked: No price · add one beside it on Usage, and on Home a line saying which models have no price. Adding a price under Usage's Pricing reference clears it. A local Ollama model is free and never marked.

On SynapseOrch Cloud, a platform model's figure is what your model wallet was charged — Charged — never an estimate. A model on your own key is estimated from the rate card; one the rate card has no price for reads No price — its cost is unknown, not zero. Cloud's rate card is ours and read-only, so there is no price to add there.

Log retention

Usage records are rows in the database's usage_logs table. They are kept until you clear them — with Clear logs on the Usage screen, or Usage history in Settings → Memory. Both delete every record; there is no automatic rotation or partial cleanup.