Skip to content
IQ

How we measure routing savings.

Routing is one capability inside the IQ dashboard. Here is how we compare its cost: a worked example you can check, and measurements of our own traffic with their baselines named.

Actual savings depend on your workload. The historical 40 to 80 percent band below is not a forecast for your team.

The 60 percent before and after example

One five-step loop, priced two ways. Worked example, not a customer result. October 24, 2026 list prices.

Before: one top-tier model

  1. Planning$0.21
  2. Retrieval$0.19
  3. Tool call$0.17
  4. Synthesis$0.39
  5. Verification$0.35
Loop total$1.31

After: a model for each step

  1. Planning$0.11
  2. Retrieval$0.02
  3. Tool call<$0.01
  4. Synthesis$0.39
  5. Verification$0.01
Loop total$0.53

The same stated token volumes, with step prices rounded to cents. The tool call is billed at less than one cent. This comparison does not measure answer quality. Actual savings depend on your workload.

Check every price and assumption

Worked example using the model lineup and list prices from October 24, 2026. The figures show how the comparison works. Actual savings depend on your workload. The home page shows a short version of this worked example. The full comparison is a single five-step agent loop: planning, retrieval, a tool call, synthesis, and verification. In the Before trace every step runs on the same top-tier reasoning model at high thinking effort, and the five step prices of $0.21, $0.19, $0.17, $0.39, and $0.35 total $1.31 for the loop. This worked example assigns a different model to each step. It compares prices for the stated token volumes and does not measure answer quality. In the After trace, synthesis stays on the same heavy reasoning model at $0.39, planning moves to a cheaper reasoning-heavy model at $0.11, and retrieval, the tool call, and verification use lower-priced models at $0.02, <$0.01, and $0.01. The tool call is under a cent, so it counts as zero cents in the total. Those five total $0.53, so the loop gives back $0.78 of $1.31, which is 60 percent.

This is a worked example rather than a customer result, and it is one representative loop chosen to show the mechanism. The baseline is named in the trace itself: every Before step runs on the same top-tier reasoning model. The per-step prices reflect a specific mix of providers and step types, so a loop with a different shape, a different provider mix, or a higher share of reasoning-heavy steps will land at a different number.

A saving is only a saving against a baseline

The same 1,071 requests of our own organic traffic, compared against three defaults. These re-price the tokens used; they do not re-run the requests.

Routed spend compared with single-model baselines, 1,071 requests
If every request usedRouted spend
A top-tier reasoning model66 percent less
A mid-tier Sonnet-class model43 percent less
gpt-527 percent more

One caveat on both measured samples, stated plainly. The 1,071-request and 6,530-request figures are our own accounts rather than customer traffic; we are signing our first design partners now. The two samples were read at different times and may share some of the same requests, so read the band as two snapshots rather than one continuous curve. Both are real measurements of a real workload rather than a projection, but it is our workload, and we would rather say so than dress it up as somebody else’s.

The measurements behind the 40 to 80 percent band

The number that matters is not how much IQ saves, it is how much it saves against the thing you would otherwise have done. Across 1,071 requests of organic traffic through the production gateway, priced from the gateway’s own billing table with no synthetic load, routed spend came in 66 percent under pinning every request to a top-tier reasoning model, and 43 percent under pinning every request to a mid-tier Sonnet-class model. Against pinning every request to gpt-5, the same traffic came in 27 percent more expensive.

Routing can cost more when your default is already sized close to the work. The Savings page compares your own routed and cached spend against a frontier-pinned baseline, using the tokens your requests actually used.

The 43 percent figure sets the floor of the 40 to 80 percent band, rounded down to the nearest ten. The top comes from a second, larger measurement, on the same gateway billing table but a different, named baseline: $5 per million input tokens and $27.50 per million output tokens, the midpoint of Claude Opus 4.8 ($5/$25) and OpenAI’s gpt-5.6-sol ($5/$30), the savings dashboard’s blend throughout that window, in place since July 10, 2026. Over the 30 days ending August 19, 2026, our account routed 6,530 requests and came in 80.6 percent under that blended baseline. Eighty is 80.6 rounded down to the nearest ten, the same rule the floor uses: neither end of the band claims more than was measured. The savings dashboard has since moved its own baseline to a blend of Claude Opus 4.8 and GPT-6 Astra, which reads higher on the same traffic; this band keeps the older, lower reading.

Both readings share the same limitation, stated plainly. Each one re-prices the tokens a request actually used at its own baseline price; neither re-runs the request against that baseline model, which could tokenize, retry, or answer at a different length than the model that actually served it. It is a real bill compared against published list prices, not a live rehearsal. Which end of the band you land on also depends on how expensive your current default is relative to the work your steps actually do, and the gpt-5 result above is the reminder that a default already sized close to the work does not land inside the band at all. Overall, the band is a rounded envelope of two measured snapshots rather than a target we picked first and then went looking for evidence to support.

Two different things get called a cache hit

Response-cache hits and discounted input tokens use different denominators. They do not add up to a single hit rate.

Response cache

163 of 1,137 requests: 14.3 percent. IQ returned a stored answer without a provider call.

Provider prompt cache

49.1 percent of input tokens were cache reads. The provider still ran the model.

The accounting gap

The June 24 to July 28 sample left out provider cache write charges, so its cost savings read slightly high.

Cache samples, dates and limitations

Conflating them is how a cache rate stops meaning anything, so here is the split. The first is IQ’s own response cache, where the gateway recognises a repeat request and returns the stored answer without calling a provider at all, so the request costs nothing and comes back immediately. Across 1,137 organic requests, a larger sample of the same production traffic behind the routing figures above, 163 were served that way, which is 14.3 percent. The second is provider prompt caching, where the provider still runs the model but bills repeated input tokens at a discount to the normal input rate. Measured as cache-read input tokens over total input tokens presented, that comes to 49.1 percent, and it is the larger of the two effects.

That 49.1 percent carries a date, and the date is part of the number. Provider cache accounting only shipped on July 20, 2026, so the measurement runs from July 21 onward, and ten of the fourteen active days in the sample record zero cache reads because nothing was recording them rather than because nothing was cached. What survives that cut is four days, 200 billed requests, and a single organisation. It is a real measurement rather than a projection, and it is a narrow one, so read it as an early rate rather than a settled one.

Combined against a no-cache counterfactual, the two effects avoided 17.4 percent of what the same traffic would otherwise have cost over the whole period from June 24 to July 28. The sample left out provider cache write charges, so that 17.4 percent reads slightly high. The 49.1 percent may read high too, if cache write tokens were also left out of total input tokens. IQ now includes cache write charges in cost estimates. Since August 12, 2026, request records have kept cache write tokens in a separate field. Overall, both numbers belong on this page, the small one beside the large one, because a cache rate quoted without naming which cache produced it tells a buyer nothing, and we would rather publish the 14.3 percent than let the 49.1 percent do work it has not earned.

Read the result for your workload

Your model mix, cache reuse and the work in each step change the outcome. The dashboard’s Savings page compares your own routed spend against a frontier-pinned baseline.

Why results vary by workload

Three factors drive most of the variance. The first is model mix: the more of your traffic that can run on a cheaper model without losing quality, the larger the cut. The second is cache hit rate: repeated or near-repeated prompts served from the semantic cache cost nothing on the provider side, so workloads with high reuse save more, while first-of-a-kind prompts save less. The third is step composition: agent loops with many light steps (retrieval, tool calls, short formatting) leave more headroom to route down than loops dominated by hard reasoning. Your own savings depend on how these three land for your traffic.