AI cost · Model choice
MixedFinance & ITKeeping AI model costs low without making every answer worse
AI bills grow quietly: a premium model on every step, a few video generations nobody checked. This guide compares the levers that keep spend down, and which to pull first.
Valnora & Elric team6 min read
Compared in this guide
- AOne premium model for everything
- BBrain Auto routing in Chat
- CCheaper, faster or Free models per workflow block
- DPinned and default media models
- ECost confirmation threshold
- FDenied models and allowed providers
- GBYOK versus platform key
Most AI bills do not grow because of one large decision. They grow because a premium model is the default on every workflow step, because Chat sends a one-line rewrite to the same model as a long analysis, and because a few people try every video model in the catalogue. None of these is wrong on its own. Together they add up.
This guide compares the levers Agent Serve gives you to keep spend down, what each one costs you in quality or effort, and which to pull first. No prices are quoted: they change too often, and any figures you see here are illustrative.
The options side by side
Ways to keep AI spend down, side by side
| Approach | What it saves | Trade-off | Best for |
|---|---|---|---|
| One premium model for everything | Nothing. It is the baseline most teams start from. | Simple, but every small task pays the top rate. | Short pilots only |
| Brain Auto routing in Chat | Simple requests go to a lighter model without anyone choosing. | Less control over which model answers a given request. | Everyday Chat use across a team |
| Cheaper, faster or Free models per workflow block | Routine steps such as tagging and extraction cost less. | Needs a test set to confirm quality holds for each step. | High-volume workflows |
| Pinned and default media models | Stops people reaching for the priciest video model by habit. | Some jobs need a stronger model; members can still ask. | Marketing and creative teams |
| Cost confirmation threshold | Expensive generations need a deliberate yes. | An extra click on large jobs. | Any workspace with media generation on |
| Denied models and allowed providers | Removes models you have decided are not worth the cost. | Admins must review the list as the catalogue changes. | Organisations with a model standard |
| BYOK versus platform key | BYOK uses rates you negotiated; the platform key needs no setup. | BYOK means managing keys and provider bills yourself. | Teams with existing provider agreements |
These levers stack. Most teams end up using four or five. The useful question is the order.
Start with routing in Chat
In Chat, the model picker offers Brain Auto. It routes each request to a suitable model, so a short rewrite does not go to the same model as a multi-step analysis. For a team that uses Chat all day, this is the change with the least effort and the widest effect.
People who need a particular model can still choose it under More models. That list shows many reasoning models, tagged Free or ZDR where that applies.
Then fix the high-volume workflow steps
Workflows are where volume hides. A workflow that tags every support email or extracts fields from every invoice may run thousands of times a month. In the workflow builder, each AI block has its own model, so the step that classifies text can use a cheaper, faster or Free model while the step that writes the final reply keeps a stronger one.
Do not guess. Run 20 to 50 real examples through the cheaper model and compare the results with what you get today. If a person cannot tell the difference for that step, switch it. Logs show each block’s inputs and outputs for every run, which makes the comparison quick.
Put limits on media generation
Images, video, speech and music cost more per request than text, and video costs the most. Three settings keep that in check:
- Default media models. In Access policies, admins set default image, video, speech and music models per permission group. Pick a Fast or lighter variant as the default.
- Pinned media models. In Chat, + → Media models pins a video, image or audio model for that chat, separate from the reasoning model. A creative team can pin the model they agreed on rather than switching each time.
- Cost confirmation. Any generation estimated above the threshold asks before it runs. The default is US$1, set by admins under Access policies → "Cost confirmation (USD)".
Every generated file lands in the Library under Generated, with its prompt, model and cost, so you can see what was spent on what.
What the solution produces
Starter cost policy for a marketing workspace (illustrative)
Checklist- Chat default: Brain Auto. Members may pick other reasoning models from More models.
- Workflows: a fast model for tagging and extraction; a stronger model only for final copy.
- Default image model set; default video model set to a Fast variant.
- Cost confirmation (USD): 1, the default.
- Denied models: two premium video models, reviewed every quarter.
- Keys: platform key while testing; company provider keys once volume is steady.
Remove what you do not want paid for
Access policies also hold allowed model providers and denied models. If your team agreed that two premium video models are not worth it for social content, deny them for that permission group. If a team should not generate media at all, turn off the Media Generation switch. Policies apply per permission group, so finance and marketing can have different rules.
Choose between your keys and the platform key
Under Settings → Model API keys, you can bring your own keys (BYOK) for model providers and some tools. Calls then bill to your provider account, at whatever rates you have agreed. Without your own key, the platform key is used and the usage is metered to the workspace.
The platform key is the easier start: nothing to set up, and each workspace’s usage is counted separately. BYOK makes sense once volume is steady and you already have provider agreements, or when security wants calls to run under your own contracts. For that side of the decision, see model data security and zero data retention.
Controls and governance
Roles and permission groups
Admin, write and read roles per workspace, plus permission groups that limit which models, blocks and integrations a team can use.
Bring your own keys
Use your organisation’s own model provider keys at organisation or workspace level.
Approvals before action
Relays never send email or book meetings without an approval, under any autonomy setting. An approval covers the exact payload the reviewer saw and expires if nobody decides.
Run logs
Every workflow run records each block’s inputs, outputs and errors, so a reviewer can trace how a result was reached.
Audit trail
Proposals, approvals, declines and executed actions are written to the workspace audit log, which admins can export.
Zero data retention
Zero data retention with supported model providers, backed by your own provider keys and agreements or self-hosted models. Some provider features keep data by design, such as Gemini Deep Research, so they are outside zero data retention.
More on how Agent Serve handles access, data and audit on the security page.
Which to do first
- Set Chat to Brain Auto as the default.
- Find your three highest-volume workflow steps and test a cheaper model on each.
- Set default media models and keep the US$1 cost confirmation, or lower it.
- Review usage per workspace after a month, and add denied models where the spend is not earning its keep.
- Move to BYOK once volume and provider agreements make it worthwhile.
Questions teams ask
Does Brain Auto always pick the cheapest model?
No. It routes each request to a model suited to it, which for simple requests is usually a lighter one. If you need a specific model for a task, choose it from More models.
What happens when a generation would cost more than the threshold?
It asks before running, with the estimate shown. Admins set the threshold in Access policies under Cost confirmation (USD). A running generation also shows a progress card with a cancel button.
Are prices listed in this guide?
No. Model prices change often and depend on the provider and on your agreements. Check the model catalogue and your provider contracts for current rates.
Where cost is decided in a workspace
People · Admin policy
- 01Access policiesAllowed providers, denied models, default media models and the Media Generation switch per permission group.
- 02Cost confirmationGenerations estimated above the threshold, US$1 by default, ask before running.
- 03Model API keysYour own provider keys, or the platform key metered to the workspace.
Workspace · Everyday use
- 04Brain Auto in ChatRoutes each request to a suitable model instead of the most expensive one.
- 05Model per blockRoutine steps such as classifying or extracting use a cheaper, faster model.
- 06Library provenanceEvery generated file records its prompt, model and cost.
What you need
- A list of your heaviest AI uses: which workflows, which teams in Chat, which media jobs
- A small test set per high-volume workflow step, to check a cheaper model gives good enough results
- Permission groups per team, so policies can differ between marketing, finance and support
- A decision on the cost confirmation threshold if US$1 does not suit you
- Provider agreements and keys, if you plan to bring your own
- A monthly owner for reviewing denied models and defaults as the catalogue changes