What Is Metered

Insulin measures six resources. Each is recorded as it is consumed, priced against the rate in effect for that month, and taken off your credit balance.


The six metered resources

These are the rows you see under Resource in the usage breakdown, with the unit each quantity is counted in.

ResourceUnitMeasured
AI input tokenstokensEverything sent to the model on each request
AI output tokenstokensEverything the model produces
Sandbox processorprocessor-secondsWhile a sandbox is running
Sandbox memoryGB-secondsWhile a sandbox is running
Sandbox storageGB-secondsFor as long as the sandbox exists, running or stopped
Object storageMiB-monthsFiles held in your organization’s storage

Tokens

Input and output are counted separately because every provider charges differently for each direction, so each direction needs its own rate.

Cached context is counted, at rates of its own. A long conversation resends earlier turns on every request, and providers charge for those tokens even when they serve them from cache. They are recorded as input tokens, but priced by what the cache did: reading from it costs 10% of the model’s input price, and writing to it 125% — unless the supplier publishes its own cache price for that model, in which case that price is used.

Reasoning tokens count as output. Models that think before answering bill that thinking at the output rate, and Fours records it the same way.

Failed and cancelled requests are still billed. A generation that fails partway has already been run, and the provider charges Fours for it whether or not the answer reached you. Whatever token count the provider reports for the failed call is what is metered.

AI that an app runs is billed like any other AI. When an app you build in Insulin calls a model at runtime, those tokens are metered and priced the same way a chat turn’s are. This used to run only on your own keys and record $0; it can now run on a Fours-hosted model, which Fours pays the vendor for — so it is charged rather than recorded for free. If your organization has no model such a call can run on, the call is refused with “No usable model provider configured” rather than silently returning nothing.

What the Fours-hosted models cost Fours

The Fours-hosted pool runs on one supplier. These are that supplier’s published prices per million tokens, read from Fours’ own rate table — the rate you are charged is these plus the margin described under How prices are set.

Fours-hosted modelInputOutput
Gemini 3.7 Flash$0.75$3.75
Gemma 4 31B$0.20$0.40
DeepSeek V4.1 Flash$0.20$0.60
GLM 5.3 Flash$0.15$0.50
DeepSeek V4 Flash$0.09$0.18

Gemini 3.7 Flash is the pool’s quality tier; Gemma 4 31B, DeepSeek V4.1 Flash and GLM 5.3 Flash are its light tier. DeepSeek V4 Flash is no longer offered: it serves only an agent or app that was already set to it. Which of them a turn may use is not only a question of price — see the tier step-down.

Sandboxes

A sandbox bills three things independently, because the sandbox provider bills Fours for three things independently.

  • Processor and memory accrue while the sandbox is running. Stopping it ends both.
  • Storage accrues for as long as the sandbox exists. Stopping does not release the volume — it is still reserved for you, and Fours is still charged for it. Deleting the sandbox is what stops the charge.

Accrual is measured in windows rather than in one lump at the end, so a sandbox left running for weeks shows up in the usage breakdown as it goes rather than arriving as a surprise.

Object storage

Object storage is not consumed, it is held — so it is measured as a gauge rather than counted. Once a day, Fours records how much your organization is holding, across shared organization storage and each member’s own storage together.

The month is billed on its highest daily reading. Not the average, and not the figure at month end. If you hold 500 GB for a single day and delete it, that month bills 500 GB: the capacity was reserved on the day you used it, and the supplier charged for it then.

From consumption to a number on your bill

The two arrows out of the balance are why the page can look inconsistent for a short while: the Credit balance moves within about a minute of a request, while the usage breakdown is re-aggregated every couple of hours. A balance that has dropped against a table that has not caught up is the normal, expected state — not a missing charge. See Reading your usage.

How prices are set

Fours prices everything; the payment provider only collects. Stripe holds no rate card and performs no pricing — it receives an already-priced invoice. That is what makes the invoiced total equal the metered total.

  • Rates are effective-dated and immutable. A price change applies to the month it takes effect in and to later months. It never reaches backwards: re-pricing a closed period reproduces the original amount.
  • Model prices come from one reviewed price list. Fours keeps each supplier’s published price for every model it prices — with an introductory price and the date it ends, the higher rate some suppliers charge for a very long request, and the supplier’s own cache prices where it lists them — and publishes from that list a rate for each model, for Fours-hosted use and for your own key, and for each kind of token, ahead of each month. A request long enough to cross a model’s long-context threshold is priced at that model’s long-context rate.
  • Other rates are refreshed monthly, ahead of the month they apply to, from the supplier’s published prices.
  • An automatic price rise is capped. A refreshed rate more than 25% above the previous one is rejected and the previous price is kept, so a misread supplier page cannot reach an invoice. Price cuts are accepted at any size — providers cut sharply and often, and refusing a real cut would keep overcharging you. Model prices from the reviewed list are exempt: they are reviewed before they are published, and a known change — an introductory price ending — can double a price overnight.
  • A model with no published rate is still billed, at a documented fallback. The fallback supplier cost is $5.00 per million input tokens and $10.00 per million output tokens; with the margin below applied, the charged rate is $7.50 per million input tokens and $15.00 per million output tokens. Cached input takes the cache share of the input fallback — $0.75 per million to read from the cache, $9.38 to write to it. On your own key the fallback is charged as the platform fee instead: $0.50 per million input tokens and $1.00 per million output tokens, and $0.05 and $0.63 for cache reads and writes. Newly released models are the usual case, and consumption is never silently dropped for want of a price.

Fours’ price is the supplier’s cost plus a margin: 50% on model calls Fours supplies the key for, and 30% on sandbox and storage. Prices are in US dollars and exclude tax.

Using your own provider keys (BYOK)

If you connect your own model provider key, that provider bills you directly for the tokens, under your own contract. Fours’ charge is then a platform fee of 10% of the supplier’s price for those tokens, with no margin on top — the platform still routes and secures every request and meters the token usage the provider reports. Each kind of token takes 10% of its own price, so a cache read costs 10% of the cache-read price. The fee is rounded to the nearest micro-cent and is never less than $0.01 per million tokens, so no token is free.

Two consequences worth knowing:

  • BYOK usage still appears in your usage breakdown, attributed to the provider whose key was used, whenever the provider reports how many tokens the request used. For streamed replies, Fours does not currently ask Cloudflare Workers AI, DigitalOcean GradientAI, or Together AI for that count, so a reply streamed from one of them may be missing from the breakdown.
  • A zero balance still pauses BYOK requests, because the platform fee is still owed. See When AI pauses.

What is never billed

Billed AI is not limited to your chats and agents. These also run for your organization, and each is billed under its own name on the Agent tab of the usage breakdown:

  • Helpers in Insulin — Knowledge base (embedding documents as they are indexed, and each search), Thread titles, AI text generation, Skill drafting, Inbox, PRM and Deal registration.
  • AI features elsewhere in the Fours console — Co-sell quality score, Co-sell referral AI fix, Co-sell business problem, Co-sell engagement score, Co-sell contact score, Marketplace insights, AI field mapping, SCA extraction, AWS invoice extraction, and the workflow AI steps: Workflow AI email, Workflow AI Slack message, Workflow AI Agent node and Workflow AI test event.

What is not billed:

  • AI Fours runs on its own behalf. A model call Fours makes for itself rather than for you — such as rewriting a marketplace’s error message on an offer or entitlement page you merely opened — is Fours’ cost. It is recorded for Fours’ own monitoring, and never billed or refused.
  • Reading what you already have. Opening, listing or searching your stored files, and starting, stopping or connecting to an existing sandbox, are never refused for billing reasons and are not separately charged.
  • Database storage. Not metered.

How long the detail is kept

RecordKept for
Individual consumption records90 days
The attributed usage breakdown you see in the console13 months
Period totalsIndefinitely

After 13 months the breakdown is collapsed rather than deleted: the attribution — which user, which app, which model — is dropped, and a daily total per resource remains. A history that ended abruptly would read as data loss.

The period selector offers the last 12 months, which is inside the window where the full attributed breakdown still exists.

Spotted something wrong or out of date on this page? Tell us and we'll correct it.