Concepts

Where a saving comes from

The levers, what each avoids, and why a saving only counts once it can be proved.

Navige exists to make agents cost less. Everything else here — the policies, the audit trail, the approvals queue — is in service of that, because a cost figure nobody can check is a marketing number.

This page is the part people usually ask about last and should read first: where the money actually comes from.

The levers

A lever is something Navige does to a call that makes it cost less than it otherwise would. Six are on unless you switch them off in Cost controls in the console: model routing, prompt caching, context trimming, deduplication, result capping and tool scoping. The output ceiling and the runaway guard are off until you set them up. Levers act on a paid account. Which ones can act depends on which way the call came in: Navige can only change what it carries.

On model calls (the LLM proxy)

  • Model routing. A cheaper model from the same provider answers when the request itself shows a smaller model is enough. The model your agent asked for is the ceiling: a call only ever moves down, never to a model that would refuse the request, and when Navige is unsure it keeps yours. A routed answer is checked before your agent sees it — an empty, cut-off or refused answer, or a malformed tool call, is run again on your model, and that agent stops being routed to a model that keeps failing. The largest lever by a distance.
  • Prompt caching. Marks a repeated prefix so the provider bills it at the cache rate — typically a tenth of the input price for the cached part. The model sees exactly the same request, so the answer is the same.
  • Context trimming. Large tool results from earlier turns are left out; your messages, the system prompt and recent turns never are. The saving is the input tokens never sent, priced at what they would have been billed: an old result usually sits in the cached part of the conversation, so it is priced at the cache-read rate, not the full input rate.
  • Output ceiling. A maximum length for the answer, per workflow, enforced by the provider. It prevents an overrun rather than noticing it afterwards, and is recorded as a fact about the call, never as a saving — nobody can observe what an uncapped answer would have cost.

On tool calls (MCP)

  • Deduplication. A read identical to one made moments ago is answered from the earlier result instead of being run again.
  • Result capping. A tool result larger than the cap is cut down before it reaches the model, so a 400 KB search result does not become 100,000 input tokens on every later turn. Set per tool.
  • Tool scoping. A workflow is shown only the tools it actually uses, learned from its own history, so the model is not paying to read fifty tool descriptions on every call.

Everywhere

  • The runaway guard. An agent stuck in a loop, making no progress, or burning money faster than its own history says is normal is stopped. One warning sign on its own never stops anything; two that agree do. It never changes a request, so once switched on it works for every caller. On Claude Code it is on unless you switch it off.
Claude Code and the desktop apps
Claude Code sends headers and body fields that only work together, so Navige never rewrites its requests: trimming, the output ceiling and prompt caching stay off for it. Model routing reaches it, on by default as Route Claude Code too on the Model routing page, and only on calls that pass the compatibility check. Measurement, budgets, the kill switch and the runaway guard always apply, and with navige connect every tool call it makes is checked too. On the /enforce route Navige sees a decision, not a payload, so no lever can act there.

Why a saving has to be proved

Navige bills a share of measured savings, so the measurement is the product. A number that cannot be defended is worse than no number, because it has to be withdrawn in front of whoever acted on it.

So a saving counts only when both sides were observed: the call that was made, priced against the call that would have been made, from published prices for the models involved. Where a price is unknown the call is recorded unpriced, with its true token counts and no cost — never as zero.

Routing is split in two. The input half of its saving is measured — the same prompt costs what it costs on each model. The output half is modelled, because the cheaper model wrote a different answer and nobody saw the length the original would have written. Only the measured half can reach an invoice.

Note
Three classes of number, and Navige never adds one to another: recorded is what was observed, measured is a saving where both sides were observed, and modelled is an estimate of what did not happen. Only measured savings are billed. See Recorded, modelled, measured.

Telling Navige whether a task worked

Per-call routing needs no outcomes: it judges each request and checks each answer. But only your workflow knows whether a task was actually right — the ticket was categorised correctly, the reconciliation balanced — and that is what experiments, cost-per-successful-task and the quality figures are built on.

There are two ways to tell us, and the first costs no extra request at all — the answer rides along on whatever your agent calls next:

bash
# on any later call, about an earlier task
curl https://api.navige.ai/proxy/openai/v1/chat/completions \
  -H "x-navige-key: $NAVIGE_KEY" \
  -H "x-navige-outcome: $JOB_ID:ok"

# or, if you would rather report it on its own
curl -X POST https://api.navige.ai/v1/transactions/$JOB_ID/outcome \
  -H "x-navige-key: $NAVIGE_KEY" \
  -H "content-type: application/json" \
  -d '{"succeeded": true}'

A failure carries a class rather than a message — $JOB_ID:fail:schema_mismatch — because a message carries your data and is unbounded. You can also judge a task with one click on its page in the console. Correcting a judgement later is a normal event, not an error.

Navige never infers this. It can see that a provider answered; it cannot see whether the answer was right, and a guess here would put evidence nobody produced under a figure that is billed.

Routing never waits for these. It decides each call from the first one; judged tasks make the quality figures and the learning sharper over time.

What Navige will not do to save money

  • Guess a price. An unrecognised model stays unpriced rather than being priced from a near neighbour.
  • Count a modelled saving as a measured one, or bill for it.
  • Move a call to a bigger, different-provider or less capable model than the one your agent chose.
  • Edit a request body where the caller's protocol forbids it, even though that disarms most levers.
  • Ask a model whether to trim your payload — it would have to read the payload first.

Each of those would raise the headline number and make it worth less. The restraint is the reason the figure can be put in front of a finance team.

Something here wrong or missing? Tell us — a person reads it.