Model use and cost

The dashboard shows what the agents’ model calls used and cost in four places:

A run’s page: at the top, above its tabs, so on every tab, one line and two cards of the run’s model use, after a hill climb’s champion card; its Stats tab has the calls and the table by agent and model.
A campaign’s page, in its sidebar: its model use, summed over all its runs on every rig, with each run.
Metrics, in the main navigation after Rigs (address /ui/metrics): the whole system’s, by model and by campaign, for today, the last 7 days or all.
Settings › Providers, under the providers: Model prices, each model’s price, which the cost uses.

Code counts every call where it is made, never an agent (The model usage records). Control sums the records when a page is drawn. The REST route GET /v1/usage and the MCP tool model_usage give the same numbers. Every label and column with an i says what its number means when you hover it.

The words

Word	Where	What it means
`not reported`	Any figure	The harness that ran the calls does not give that figure, so it is not counted. It is never shown as 0
`on a plan`	Cost	The calls ran on a subscription: a Claude or ChatGPT provider, or Claude Code. Their tokens are counted, not priced
`no price`	Cost	No price is known for the model: its provider’s list gives none and you set none. Never $0
`free`	Cost	The model’s price is 0 in every part: its provider’s list’s, or the one you set
`<$0.01`	Cost	Less than half a cent, but more than nothing

A cost with an i before it leaves something out or adds something. Its i says what: the GPU time on RunPod in it, the tokens on a plan or with no price that are not in it, the free tokens, and the tokens in that were priced as not cached because their harness does not report the cache (Pi, for example, when the provider’s answer does not name its cached tokens).

The six numbers

Metrics starts with six tiles. A run’s page gives four of them, shorter, in two cards: Cost, and Tokens with Output speed. Its calls are on its Stats tab, and its time in calls is in its line.

Tile	What it shows
Cost	The dollars, split by a bar into tokens in (not cached), cached in, out, and GPU time on RunPod. Under it, the tokens with no dollars: `144k tokens on a plan · 350k no price · 3.3k free`
Tokens in	Every token sent to the models, split into `cached` (the provider read them from its cache), `not cached`, and `not reported by` the harnesses that do not say
Tokens out	Every token the models wrote, with the reasoning tokens under it where the provider counts them
Calls	The model calls. Records that do not say how many calls they cover (a Codex turn) count as none, and an i says how many there are. On Metrics, the line under it counts the runs and campaigns, such as `12 runs · 3 campaigns`. The bars under it count the calls in each hour of today, or on each day (each week, past 62 days), once there are two or more to draw: Today has none before 1 a.m.
Output speed	Tokens out a second over each call’s whole time, from the request to the last token: the median, with `p10–p90` under it (8 of 10 calls are in this range)
Time in calls	All calls added together, from the request to the last token, with the median time to the first token under it

A run’s page

At the top of the run’s page, above its tabs, so on every tab, are one line and two cards. A hill climb has a third card first, its champion. A run with no agents and no pod spend on RunPod, such as a try of a kit, has none of them, and no Stats tab.

The line counts the run’s trials and rounds, then gives its time and its time in calls: 48 trials · 16 rounds · 30 h 0 m · 6 h 34 m in calls.

The run’s time is from its start to now, or to its end once it is finished. A paused run’s time goes on.
The time in calls adds all its calls together, from the request to the last token, so calls at the same time each count. Its i says how many records do not report their time, which are not in it. Where no record reports one, the line says time in calls not reported.
A run that has made no model call yet ends the line with no model call yet, and has no cards but a hill climb’s champion.

Under the line:

Champion #N, on a hill climb: the champion’s value of the main metric, with its unit, as the judge measured it when it confirmed the change (the cases’ geometric mean), such as 14.6 tok/s. Under it, its gain over the starting point, measured on this rig, and the starting point’s value: +88.0% over the start, 7.73 tok/s. A champion whose verdict did not record its values (one judged by an older harness) shows an estimate from the starting point’s value and the gain, marked ≈: ≈14.5 tok/s. Before any change is promoted, it shows the starting point’s value and the starting point: no change promoted yet.
Cost: the dollars and their split, as on Metrics, with the pod’s spend while this run was at work on a RunPod rig. The tokens with no dollars are not under it: the Stats tab gives each agent’s and model’s cost, in dollars or in words.
One card with Tokens, in and out together, such as 30.1M in · 995k out, with the split of tokens in under it, as on Metrics; and Output speed. The reasoning tokens are on Metrics.

The Stats tab

Stats, the last of the run’s tabs (?view=stats in the address), has the details:

Calls: the run’s calls, how many calls the workers made in a trial (28 a trial), and the median time to the first token (first token 1.8 s). The bars beside it count the calls in each round, from round 1 to the latest; a run with fewer than two rounds has no bars.
By agent and model: one row for each agent’s calls to each model.

A run that has made no model call yet says No model call recorded yet. Code counts each call of the agents when it ends. A run with no agents and no model use says No model use: this run has no agents.

By agent and model has its rows in groups, as Edit rig names the agents:

The work: Workers, then the Planner.
The advisors: Teachers (with teacher 1 or teacher 2 under it when the run has two teachers and the rig’s settings say which model is which, and merged by for the model that merges their advice), the Auditor, the Review panel and the Synthesizer (with answer cards under it), whose calls write a research run’s answer cards. A research run’s final writer has no row of its own: its calls count under Teachers, whose model it uses.
The rig: on a rig launched on RunPod, the row GPU time on RunPod: the pod’s spend while this run was at work, with its time at the pod’s price an hour, such as 5 h 36 m at $1.65/h.

Column	What it shows
Agent	The agent, and the line under it
Model	The model, with the harness that ran it and the provider that served it under it: `Pi · my-litellm`
Calls	Its calls
In	Its tokens in
Cached	Its tokens in that the provider read from its cache
Not cached	Its tokens in that the provider read again
Out	Its tokens out, with reasoning
tok/s	The median of its calls’ output speed, with p10–p90 under it
Cost	Its tokens at the model’s price, or one of the words
Share	Its tokens in and out, as a part of the run’s

On a narrow screen the table keeps Agent, In, Cached, Out, tok/s and Cost, and shows the model under the agent.

A campaign’s page

The campaign’s sidebar has its own Model use, over all its runs on every rig. Its line on the right counts the runs with records or pod spend.

Cost: the dollars and their split, as on Metrics. The tokens with no dollars are on Metrics.
Tokens in, with the part that was cached under it. Where a harness does not report the cache, the part is of the tokens that were reported, and an i says how many were not.
Tokens out, with the reasoning tokens.
Calls, with the time in calls under it.
Output speed, with the median time to the first token.
A table of its runs: Run (a link to the run’s page, with at work, paused, completed or finished and its rig), In, Out and Cost. A run’s cost includes its pod’s spend on RunPod.
By model, on Metrics › opens Metrics.

A campaign whose runs have made no model call yet says No model call recorded yet.

Metrics

Metrics shows the model use of every campaign and rig. The chips Today, 7 days and All choose the period; 7 days is the default. They are links, such as /ui/metrics?period=today.

Today starts at midnight in your own time zone. The page’s script adds your browser’s zone to the address (tz, such as tz=Europe/Paris); without the script, today starts at midnight UTC. The line under the title says which zone the page uses: every campaign and rig · times in Europe/Paris.
7 days is the 7 days up to now.
All is every record.

With no call in the period, the page says No model call in this period., or for All No model call recorded yet. Code counts each call of the agents when it ends, on every rig.

Under the six numbers:

By model: one row for each provider’s model, with its provider and every harness that ran it under it (my-litellm · Pi, OpenCode v2), its Calls, In, Out, Cached (the part of tokens in), tok/s, Cost and Share. When a RunPod pod spent in the period, the last row is GPU time on RunPod, with the pods’ rigs. On Metrics it includes the time no run was at work, which no campaign counts.
By campaign: one row for each campaign, with its mode, its Runs, then the same columns. A run that control has no campaign for is no campaign.

Model prices

In Settings › Providers, under the providers, Model prices is one table of the models’ prices, in US dollars for a million tokens. It has one row for each model of each provider, except subscriptions: a Claude or ChatGPT provider’s calls are on its plan.

Column	What it shows
Model	The model’s id at its provider
Provider	The provider that serves it
In	Dollars for a million tokens in that the provider did not read from its cache
Cached in	Dollars for a million tokens in that the provider read from its cache. Empty: the In price
Out	Dollars for a million tokens out, with reasoning
Price from	Where the price comes from: `provider's list`, `yours`, `free` or `no price`

On a narrow screen the provider shows under the model.

Price from	Meaning
`provider's list`	The provider’s model list gives the price. Control keeps it up to date with the list. Its prices show as text, not as fields, and can’t be changed
`free`	The list gives a price of 0
`yours`	You set it here, or with `POST /v1/model-prices` or `set_model_price`
`no price`	Neither: the model’s calls show `no price`

A model with no listed price has three fields. Save prices, under the table, saves every row at once, and waits until a price changes. To set a price, fill in In and Out; to remove yours, empty all three. The line beside Save prices then says Saved. Every cost follows.: a price you set or change costs every call of the model, the earlier ones too. With no change it says Nothing changed. Each change is in the audit log as model_price_set or model_price_cleared, by owner:ui.

The line beside Save prices says why a row can’t be saved, with the model and its provider: as you type it starts with Not yet., and after you press Save prices with Not saved., such as Not saved. small-model of lab: set In and Out, or leave all three empty. Rows without a problem are saved even then, and the line ends with how many, such as The other 2 changes are saved.

Message	Why
`a price is a number, such as 0.60.`	A field holds something other than a number of US dollars
`set In and Out, or leave all three empty.`	Only one of In and Out is set, or Cached in without In