Claude Code (isobel.a@naturebaby.com)
Not available for this agent
- Governance. This agent has not reported a tool call or an approval, so there is nothing for a human to have ruled on.
- Transcripts. Payload capture is off for this organisation. Tokens and costs are still recorded; content is not. Turn it on in Settings if you need transcripts.
Priced by Sentry
$0.6806
List-equivalent, not billed
Floor on this volume
$0.0008
All of it on gpt-5.4-nano
Headroom
$0.6798
The gap routing could close, at most
Cache hit rate
0%
0 tokens read from cache
What this means
This agent moved 112.4k tokens across 1 model, priced at $0.6806. The same token volume on gpt-5.4-nano would price at $0.0008. The difference, $0.6798, is the most routing could ever recover, not a saving on offer. Whether the cheaper model would have done the job is a judgement telemetry cannot make for you, and any vendor quoting you the full number as a saving is selling you arithmetic.
Note: this agent is on a seat, so none of this is money anyone was billed. The figures are list-equivalent. They still matter, because they tell you what the same work would cost the day you move it onto the API.
Where these levers actually come from
- ●Model routing– computed on this page, right now, from this agent's real calls against the price book.
- ○Context re-orchestration, tool-call parallelisation, knowledge-base restructuring, conversation-level token tracking– analysis Sentry delivers as onboarding/consulting work today, not a number this console computes on its own. If a proposal names a saving from one of these, it came from a person reading this agent's transcripts, not from this page.
Where the tokens went
What to do about it
Raise the cache hit rate
Only 0% of input is being served from cache. Cached reads are an order of magnitude cheaper than fresh input, so a stable system prompt and a stable tool set pay for themselves quickly.
One model is doing everything
Every call on this agent went to claude-sonnet-5. That is simple, and it is also the most expensive way to run an agent. Splitting the easy calls onto a smaller model is where the first real reduction comes from.