What Nerve saved us, measured.
We run our own engineering on Nerve: a team of AI agents writing, testing and shipping code every day. These are that instance's real numbers for the last 30 days, every figure labelled measured or estimated.
Last publishedOctober 24, 2026
Saved in 30 days Measured
$20,233
75.9% below what the same usage costs without cached context.
Spend at published API rates
$6,417
Every model call, priced per token at the provider's list rates. Not an invoice.
Input read from cache Measured
99.8%
Up from a 90-day average of 94.5%.
Also saved Estimated
$2,189
Modelled, shown on its own, never added to the measured figure.
Where it came from
Five sources, each with its basis stated. A dash means we do not measure that source yet, so we do not claim it.
- Cached context. Every call records how many input tokens the model provider read from its cache. Cached tokens bill at about a tenth of the normal input price; the saving is the difference, priced per model at published rates.
- Work done without a model call. Many agent steps (reading a file, querying the board, running a check) are handled by deterministic tools instead of a model round. We count those calls and value each at an assumed 1,500 input and 300 output tokens, priced at Claude Sonnet rates. The count is recorded; the per-call value is our assumption.
- Trimmed tool history. Older tool results are summarised instead of re-sent on every round of a long task. We record the characters removed and convert at about 4 characters per token.
- Local models. Tokens served by models you host yourself, valued at cloud prices. Our own instance runs no local models, so this is $0 for us.
- Smaller working memory. Not measured yet. Nerve does page context in and out, but we do not yet record a before-and-after we would stand behind, so we show a dash rather than a number.
Saved in the last 30 days, by source
Last published Oct 24, 2026 UTC
Cached context
$20,233
Work done without a model call
$2,172
Trimmed tool history
$16
Local models
$0
Smaller working memory
not measured
- Measured
- $20,233
- Estimated
- $2,189
Measured figures come from recorded usage. Estimated figures are modelled and shown separately; they are never added to measured ones. A dash means we do not measure that source yet.
Day by day
Each day's spend at API rates, and what the measured saving would have added on top.
The daily chart draws from the live feed. The 30-day totals above are the last published figures.
30 days and 90 days
| Last 30 days | Last 90 days | |
|---|---|---|
| Spend at API rates | $6,417 | $12,371 |
| Saved, measured | $20,233 | $29,135 |
| Below the uncached cost | 75.9% | 70.2% |
| Saved, estimated | $2,189 | $4,734 |
| Input read from cache | 99.8% | 94.5% |
The rise in cache share follows October 5, 2026, when we switched on Nerve's optional conversation caching for our instance.
The same instance, the same 30 days
What that spend bought, and who approved what shipped. These figures come from the live feed only.
- Tasks completed by agents
- —
- Cost per completed task
- —
- Model calls
- —
- Production deploys
- —
- Approved by a person
- —
- Approved by a written policy
- —
- Break-glass
- —
- Shipped with no recorded decision
- —
A dash means the figure has not been published from the live feed yet. Since 8 October 2026, a deploy cannot start without a recorded decision; earlier deploys in the window may show none.
What carries over to your agents
Always on
Each agent's standing instructions and tool definitions go out as an unchanging prefix the model provider can cache, and routine steps run as deterministic tools instead of model calls.
Optional, off by default
Caching the running conversation as well, with a guard that switches it back per agent if it costs more than it saves. We turned it on for our instance on October 5, 2026.
Depends on you
Your model mix, how long your agents work on each task, and whether you already cache. If you do, your saving from this source is smaller, and we will show you by how much.
How we count
Where the numbers come from
Every model call Nerve makes is recorded with its input, output and cache-read tokens and the model used. Spend is those tokens priced at each provider's published per-token rates. Our instance runs on a subscription, so this is what the usage would cost at list price, not what we were invoiced.
What is measured and what is estimated
Cached context is measured: the provider reports cache reads on every call. Work done without a model call and trimmed tool history are estimated: the counts are recorded, but the value of each depends on stated assumptions (1,500 input and 300 output tokens per avoided model round at Claude Sonnet rates; about 4 characters per token). We never add the two kinds together in a headline.
What we leave out
People's time, infrastructure, and Nerve's own fee are not in these figures. Neither is smaller working memory, which Nerve does but we do not yet measure. Leaving a source out can only make the saving look smaller than it is.
How fresh it is
The page refreshes from our instance at most every 10 minutes. If the feed is unavailable it shows the last good figures, marked stale, for up to a week, then falls back to the figures published with the page, marked as such. It never shows zeros in place of a missing number.
See your own numbers.
The two-week, read-only assessment measures this on your agents' real traffic: what you spend, what caching and deterministic tools would save, and what we cannot yet measure.