Skip to content

Plans, billing and Compute Units

Runink TIDE is sold on its own, separately from FACE and PULSE. FORGE is part of TIDE: a TIDE plan includes it. Every product has the same three plan categories, and the category says where the work runs, which decides how compute is charged.

The price list is served by the console at GET /api/plans?product=core, which needs no sign-in.

The three categories

TIDE LiteTIDE DedicatedTIDE Enterprise
Where the work runsRunink’s shared hardware, spot-stylerunners in your own cloud account: GCP, Databricks or Snowflakeyour premises (on-prem runners) or your own cloud account
Priced perseat (a person), 1–9 seatsnode (a machine), from 1 nodecustom quote
Price$85 a seat a month billed yearly, or $99 a seat month to month$279 a month for the first node, then $169 a month for each extra node; yearly contractcustom
Compute18,000 CU a seat a month, meteredunlimited CU, on your own cloud billunlimited CU
Other chargesCU over the allowance (below)1.5% of the runners’ cloud costnone on on-prem runners

There is no node fee and no flat platform fee on any plan: the plan’s platformMonthlyCents is 0. On-prem runners are Enterprise only. A yearly price is charged as twelve times the monthly rate.

The seat or node count must fit the plan’s band, or the quote is refused with the band’s minSeats and maxSeats. A Dedicated plan offered month to month is refused with “this plan is annual-only; retry with annual=true”. Enterprise returns no quote: /api/checkout answers custom: true and directs you to sales.

Compute Units

1 CU is one second of a model slot, so 3,600 CU is one slot-hour. A slot is one request’s hold on the shared model plane. CU is charged only on Lite, where the work runs on Runink’s hardware. Dedicated and Enterprise are unmetered: their compute runs on your own infrastructure.

On LiteRule
Allowance18,000 CU a seat a month (5 slot-hours), pooled across the subscription’s seats
Over the allowancepay as you go at $0.12 per 100 CU, billed per started block of 100 CU
Prepaid CUbought in advance for a month at 5% off, or for a year at 10% off. Unused prepaid CU lapses at the end of its period. Prepaid CU extends a Lite allowance only
Fair useat most 900 CU a seat in any rolling hour (15 slot-minutes). Work past that queues: it is never refused and never billed extra

How CU is measured

The model router (see Models and inference) is the one place that sees both when a request was given a slot on the model plane and when its response finished, so it is the one meter. It measures how long each request held its slot; the time a request spent waiting for a slot is never billed. Billing sums the period’s slot time and converts it to CU once, rounding up, so a period with any usage at all bills at least 1 CU.

Controls on that path:

  • The router never charges anyone. It measures; it does not know who pays.
  • Metering never blocks a request. If a measurement cannot reach the caller (for example, the client disconnected mid-stream), the request still completes and the gap is counted. A missing measurement records nothing: never a guessed figure.
  • One rule for every product. FACE and TIDE record and read compute units with the same shared recorder and meter, so a slot-second costs the same CU in both.
  • Who pays for the console’s own model calls (Ask TIDE, the grounded Ask, the console chat, and voice transcription). By default, the signed-in person’s purchaser: their own subscription, or the subscription their seat is on. This is the same account the Billing page shows, so every person’s usage lands on the allowance they see. A single-purchaser installation can instead bill every call to one account, whoever asked.
  • The house account. A signed-in person with no subscription and no seat is billed to the installation’s house billing account when it names one, as FACE bills an unplaced member to its purchaser file’s default.
  • Unattributed usage. A call with no signed-in person (or one the console cannot resolve: an identity that is not an e-mail address, an unreadable subscription table) is counted as unattributed, against the platform account. It stays visible in the console’s metering counters, and it is never invoiced to anyone, the house account included.
  • Unmeasured sites are listed below. An unmeasured site records nothing; it is never estimated.
  • Each product is metered separately. A TIDE subscription’s Billing page counts only TIDE’s usage (the console and FORGE). A purchaser’s FACE or PULSE usage counts against that product’s own allowance and is never pooled into TIDE’s.
  • The Billing page shows given against used for the current anniversary period: the plan’s CU allowance, the CU measured so far, and the period’s start and end. With no current period on record, the page says compute units are not measured rather than showing a lifetime total. An app instance TIDE runs for a subscriber (FACE, PULSE) shows the same period on its own Billing page; an instance with no period on record shows usage as unmeasured.
  • Platform work is never billed. TIDE’s own agents on their own triggers (the reviewer, fixer, curator and the rest, on a schedule, a pull request or a mention) are never metered.
  • Agents a person dispatches or schedules are billed to that person’s purchaser. A DataEx › Agents › Trust dispatch, or an Atlas playbook run a person started, carries a short-lived billing grant; the purchaser’s identity stays in the console and never reaches the Git host. A retried report is never counted twice, and a grant cannot record more time than the run could have used. A playbook run a schedule or an event started is billed to the admin who activated that playbook, as FACE bills a scheduled brief to its owner. A run nobody attributable stands behind (a built-in playbook, a person with no TIDE subscription or seat) is unmeasured: never billed to anyone.

Where compute units are measured

SiteStateNotes
Ask TIDE, the grounded Ask, console chatMeasuredthrough the model router, against the signed-in person’s purchaser
Voice transcriptionMeasuredthrough the model router, the same rule as chat
Agents a person dispatches (Harness, playbooks)Measuredthrough the model router, against the purchaser of the person who dispatched the run or activated its playbook
Text-to-speechNot measuredspeech synthesis does not run through the model router, so there is no slot time to bill
FORGE verify runsNot measuredbuild-runner time, not model time
Agents on their own triggers (cron, PR events, mentions)Not measuredthe platform’s own work: never billed

CU is not spec.computeUnits

ClientInstance.spec.computeUnits is a sizing budget for an instance’s pod, not the billed unit: the operator maps it to bounded CPU and memory limits. The FinOps view on the Intelligence overview’s CFO Cockpit multiplies that allocation by the $0.12 per 100 CU rate and labels the result cu-estimate with measure: allocation: it is showback of what was allocated, an estimate, never a charge, and never the metered usage the Billing page shows.

Currency and billing periods

  • One list, in US dollars. /api/plans and the /api/checkout quote are in US cents.
  • Your regional currency. Billing converts the list into the currency of the purchaser’s country with the European Central Bank’s euro reference rates. The rate is the one in force when the month opens: the last ECB publication before the 1st, fixed for the whole month. Seat prices round to a whole unit of the local currency; usage lines and fees round to its minor unit (cents). A country without a mapped currency is billed in US dollars.
  • A missing or stale rate stops the bill. It is never billed as 0 and never at another rate.
  • Anniversary periods. A period runs from the day of the month your subscription started, 00:00 UTC, to the same day the next month. A start on the 31st bills to the last day of a shorter month (February 28, or 29 in a leap year), then returns to the 31st.

Seats, subscriptions and the Billing page

  • Seats are managed on the instance’s Admin page, through the billing seat routes (/api/billing/seats). Nobody changes their own access, the last admin stays, and a full seat pool refuses an invite. See Agents › Access.
  • One subscription per email. Every route that starts a checkout (/api/signup, /api/checkout and the Billing page) refuses a second subscription for the same purchaser. The public signup is gated by reCAPTCHA when it is configured.
  • Checkout goes through Stripe, and only when the installation configures it. Without it the quote still works and the page says billing is not configured.
  • A Dedicated plan can name its cloud account at checkout (cloudTenant: a GCP project, a Databricks workspace or a Snowflake account). A Lite plan naming one is refused, because Lite runs on shared hardware and has no runners of its own.
  • Entitlement fails closed. A subscription is entitled only while it is active, trialing or past due; anything else, an unknown plan, or an unreadable subscription record grants no plan, never Lite and never a free tier.
  • The Billing page (the Billing row of the rail footer) is the shared billing component. It shows a usage figure only where a meter feeds it: an unmeasured value is shown as unmeasured, never as zero, and an estimate is never shown as a charge.

For where TIDE can deploy a runner, and what a connected runner does, see Runners.