What do answers, workflows and agent tasks cost together?
- Historical prices with sources
- Less model work through reuse
- Team costs alongside company use
Illustrative calculation · no measured DataInbox runs
Your usage mix
Start with the work, not tokens.
Enter the monthly number of questions, workflows and agent tasks you expect. Use one category or combine them. The starting volumes are examples you can change.
Questions to answer
A colleague or customer asks about an order, policy or product.
DataInbox selects permitted, current sources and keeps the answer linked to its evidence.
Workflow requests
A request needs checks, routing and perhaps an approval.
DataInbox holds the state, rules and approval boundary; AI helps where interpretation is needed.
Agent tasks
An agent gathers context, proposes a next step and uses an approved action.
DataInbox limits the agent's context and authority, and records the action and result.
Estimated model cost
380 expected questions or tasks per month, using the reference assumptions below.
€52.94/ month
About US$59.22 in Standard text tokens.
This estimates provider usage if every entered use follows the reference model-call pattern, not savings. Some workflows need no model. Subscription, implementation and other costs are separate.
Provider prices checked 9 Oct 2026 · ECB rate dated 8 Oct 2026.
How is this estimate calculated?
The three kinds of work use fixed reference profiles. Model calls, tokens, cache reads and extra attempts per use are assumptions, not measurements of DataInbox or your organisation. A pilot replaces these profiles with observed data.
| Work type | Calls | Input / call | Output / call | Cache reads | Extra attempts |
|---|---|---|---|---|---|
| Questions to answer | 3 | 18,000 | 700 | 30% | 6% |
| Workflow requests | 8 | 14,000 | 350 | 15% | 12% |
| Agent tasks | 10 | 20,000 | 900 | 10% | 15% |
For GPT-5.6 Terra, this estimate uses $2 input, $0.2 successful cache reads and $12 output per million tokens. Check the linked provider source for context limits, introductory rates and cache conditions.
For each work type: uses × calls × tokens per call × (1 + extra attempts). Uncached input, successful cache reads and output each use the published Standard rate per million tokens. USD is divided by the ECB reference rate (€1 = $1.1186). The three amounts are added.
Excluded: cache writes and storage time, tools, grounding, storage, regional or priority processing, tax, DataInbox subscription, implementation, human review and failure costs. More tokens or a lower token price do not prove a better outcome.
Sources: OpenAI GPT-5.6 Terra, OpenAI, Anthropic, Google, ECB.
The AI Companion governance layer is included in the DataInbox subscription; model consumption follows the provider contract. See DataInbox pricing.
The same questions, changing model prices
What have 1,000 questions cost over time?
Compare the published cost of the same workload: 10,000 input tokens and 500 output tokens per question. See how model releases and price changes affect model costs.
Cost per 1,000 questions. The price axis is logarithmic: equal steps show equal ratios, keeping lower costs readable. Points are actual release or price events; lines connect different models and are not daily measurements.
Claude Sonnet 5.5 · per 1,000 questions
€22.37
Standard price per million tokens: $2 input · $0.1 cache read · $10 output.
Only cache reads became cheaper. This historical workload uses no cache, so its USD cost is unchanged. The EUR view uses this event's exchange rate.
€1 = $1.1177. USD cost divided by this rate = EUR cost.
OpenAI
First recorded
€307.35
GPT-4
14 Mar 2023
Latest recorded
€22.02
GPT-6.1 Sol
29 Sept 2026
Anthropic
First recorded
€83.72
Claude 2
11 Jul 2023
Latest recorded
€22.37
Claude Sonnet 5.5
7 Oct 2026
First recorded
€37.08
Gemini 1.5 Pro
23 May 2024
Latest recorded
€8.10
Gemini 3.8 Flash
2 Sept 2026
Locally stored price history from 14 Mar 2023 to 7 Oct 2026, checked on 9 Oct 2026. EUR amounts use the ECB rate for each event, not today's rate. The calculation uses uncached input, standard processing and no batch discount. Equal token counts do not prove equal model quality. Tools, grounding, storage, regional processing and taxes are excluded.
Where less model work can come from
Resolve the business facts once. Give AI the work that needs it.
An order question can lead to an answer, a workflow or an agent task. DataInbox can keep current state, permissions and rules available across those interactions, so each model request does not need to reconstruct them.
“Can this order still be changed?”
Architecture example · savings must be measured in your workflow
Repeated model orchestration
When every interaction starts from scratch
Collect context again
Send order data, policies and conversation history into the model.
Reconstruct the decision
Ask the model to work out status, applicable rules and the next step.
Prepare the answer or action
Draft a response, then check what may actually be done.
Repeat on the next request
Another person or agent can repeat the same retrieval and model work.
A comparison pattern, not a claim that every agent currently works this way.
With a governed DataInbox flow
Current business context stays with the Inbox
Resolve facts and rules
Look up current order state; check identity, permissions and the change window.
Use AI where needed
Give the model relevant, permitted facts to interpret the request or write the answer.
Control the next action
Use an authorised action or route the proposal for human approval, as configured.
Keep the result as evidence
Record the answer, action and resulting state for the next interaction.
Requires connected sources, configured rules, action permissions and a clear owner.
Less context per request
Possible benefit: Select the facts needed for this question instead of repeatedly sending a broad history or knowledge base.
Measure in a pilot
Input tokens per correct answer, source coverage and answer quality.
Fewer unnecessary model steps
Possible benefit: Known checks and routing can run through rules. This can reduce repeated model calls and manual handoffs.
Measure in a pilot
Model calls per completed case, retries, handoffs and total processing cost.
Reuse an answer when it still fits
Possible benefit: An approved answer may serve a repeated question when audience, permissions, source state and validity still match.
Measure in a pilot
Eligible reuse, stale or rejected answers, rework and time to a usable response.
Cache discounts and avoided work are different mechanisms
Provider prompt cache
A provider may charge less for repeated input that meets its cache conditions. The request can still contain the same number of input tokens and still invoke the model. Cache writes, storage and expiry can affect the price.
Less work sent to the model
Selecting less context or handling a known check with rules can reduce input tokens or model calls. Reusing a validated answer requires an explicit policy; a stored answer alone does not make reuse safe or automatic.
What this page establishes
The diagram explains a design supported by the documented DataInbox capabilities. The benchmark profiles are illustrative assumptions, with no measured runs. They do not prove a saving for your implementation.
Team accounts and company use
What do team accounts mean for your total budget?
Personal AI workspaces support individual colleagues. DataInbox organises shared answers, workflows and agents for the company. Compare the costs and choose which personal accounts remain useful.
$20 / person per month · billed annually · minimum 2 accounts.
The starting calculation retains every account. A retained subscription bills at least 2 accounts; zero assumes the subscription has been fully cancelled.
DataInbox · Starter
€295.00 per month for the company, with unlimited users. Model use is charged separately.
The usage mix above represents an estimated 1,300 business messages per month; workload assumptions are not a measurement.
Both amounts use the same monthly scale. They pay for different functions: personal workspaces and shared business operations.
Current personal workspaces
€893.97
- Personal AI accounts · 50
- €893.97
The selected team subscription only. Existing external API costs are not included here.
DataInbox + model use + retained accounts
€1,241.92
- Personal AI accounts · 50
- €893.97
- DataInbox company platform
- €295.00
- Estimated model API use
- €52.94
One company subscription plus estimated API consumption from your usage mix and the personal accounts you retain.
Difference from the current account bill
+€347.94/ month
This is a budget scenario. It does not establish savings, equivalent capabilities or replacement of personal workspaces.
Annual prices are monthly equivalents under an annual commitment; Claude bills this upfront. Fewer accounts affect your bill only when the contract and billing terms allow it; removing accounts does not automatically produce a refund. Confirm this with your provider.
Provider amounts are USD list prices converted at the ECB rate dated 8 Oct 2026: €1 = $1.1186. This is an indicative EUR conversion; the local invoice can differ. Prices exclude tax. Extra credits, tools, grounding, storage, implementation and other provider charges are not included.
Team subscriptions checked: 9 Oct 2026. Official sources: ChatGPT Business Standard · DataInbox pricing and capacity · ECB.
Business evidence
Measure the result, not just model usage.
A useful comparison runs the same questions and tasks before and after setup. Only when quality and actual usage are known can you judge total cost and value per completed task.
Questions to answer
Measure: Correct answers with usable sources.
Workflow requests
Measure: Completed cases, handoffs and rework.
Agent tasks
Measure: Tasks completed within the permitted scope.
Include human time, waiting, rework, subscriptions and implementation. Fewer model calls create business value only if the task is completed at least as well.
Test your real workload
Turn this estimate into a testable pilot.
Send your usage mix. Together we can define the source, outcome, approval and cost measures for a like-for-like trial.