Models · Right model per job
MixedAll teamsWhich model for which job, from tagging emails to rendering video
Tagging an email, reading an invoice, writing a board memo and making a product video are different jobs. This guide matches each to a kind of model, and shows when to let Brain Auto route and when to pin a model yourself.
Valnora & Elric team6 min read
Compared in this guide
- AClassification: tag a ticket, route an email, label intent
- BExtraction: fields from invoices, forms, contracts
- CWriting: customer replies, follow-ups, summaries
- DLong reasoning: analysis, memos, multi-step plans
- EImages: product shots, ad variants, edits
- FVideo: short ads, product clips
Most teams pick one model they trust and use it for everything. It works, but it means tagging an email costs the same as writing a board memo, and a slow model sits in the middle of a workflow that runs five thousand times a month. The opposite mistake is a cheap model on a job that needs care, such as reading totals off invoices.
This guide matches common jobs to the kind of model that fits, and shows where in Agent Serve each choice is made. It does not quote prices, which change too often. Model names are examples from a catalogue that changes; any figures are illustrative. For the wider list of cost levers, see keeping AI model costs low.
The jobs side by side
Matching the job to a model
| Job | What matters | Start with | Route or pin |
|---|---|---|---|
| Classification: tag a ticket, route an email, label intent | Speed, cost per call and a fixed list of labels | A small, fast model with the label list in the prompt | Pin it on the workflow block. Volume is high and the job never changes. |
| Extraction: fields from invoices, forms, contracts | Exact values, layout handling, structured output | A parsing block (Textract, Reducto or Mistral Parse) for the document, then a mid-size model | Pin both. Check totals and dates against a test set. |
| Writing: customer replies, follow-ups, summaries | Tone, accuracy against your sources, short output | A mid-size model with the Knowledge Base for facts | Pin in workflows; Brain Auto is fine in Chat. |
| Long reasoning: analysis, memos, multi-step plans | Quality over speed, long context, citing sources | A strong reasoning model | Pin in Chat for that piece of work, or on the one workflow block that needs it. |
| Images: product shots, ad variants, edits | Following the reference image, text in image, cost | An image model such as Nano Banana 2.1, Seedream 5.0, FLUX.2 or GPT Image 2 | Pin via + → Media models per chat; admins set the team default. |
| Video: short ads, product clips | Motion quality, length, cost per second | A fast variant such as Veo 3.1 Fast for drafts; Veo 3.1, Kling v3.0 or Seedance 2.5 for finals | Pin per chat. Keep the cost confirmation on. |
The pattern is simple: the more often a job runs and the narrower it is, the smaller the model can be. The rarer and more open-ended the job, the more a strong model earns its cost.
Route or pin: the real decision
Brain Auto in the Chat model picker routes each request to a suitable model. It is the right default for everyday use, where one person might ask for a rewrite, then a spreadsheet formula, then a strategy outline in ten minutes. Nobody should have to choose a model for each.
Pinning makes sense when the job is known in advance:
- In workflows, pin every AI block. A workflow does the same job each run, so choose the model once, test it and leave it. In the workflow builder, each block has its own model, so a triage workflow can classify with a small model and write the escalation summary with a strong one.
- In Chat, pin for a piece of work. Choose a model under More models for a long analysis or a document you will iterate on for an hour. Reasoning models there are tagged ZDR and Free where that applies.
- For media, pin per chat. + → Media models pins an image, video or audio model for that chat, separate from the reasoning model. A creative team pins the model they agreed on, rather than switching each time.
Test before you switch
Do not choose by reputation. For each high-volume job, keep 20 to 50 real examples with the answer a person expects, in a table or a file. Run them through the current model and the candidate, and compare. Logs record each block’s inputs and outputs, so the side-by-side takes minutes.
What the solution produces
A model choice recorded for one workflow (illustrative)
Checklist- Support triage workflow · 3 AI blocks
- Classify intent (9 labels): small fast model · 47 of 50 test emails matched the agent’s label
- Extract order number and product: same small model · 50 of 50 matched
- Draft reply from the Knowledge Base: mid-size model · 2 of 20 drafts needed edits, same as the larger model
- Escalation summary for a lead: strong reasoning model · kept, low volume
- Review again when the catalogue changes or the label list grows
For extraction, test the hard cases on purpose: scanned pages, two-column layouts, totals that span pages. A parsing block such as Textract, Reducto or Mistral Parse often matters more than the model that reads its output.
Media: draft cheap, finish carefully
Images and video cost more per request than text, and video the most. A practical routine is to draft with a fast variant, such as Veo 3.1 Fast, and only render the chosen version with a heavier model. Product photos and reference images can be passed in, as image-to-image or as the first frame for image-to-video, which cuts the number of attempts more than any model change.
Any generation estimated above the cost confirmation threshold asks before it runs. The default is US$1, set by admins under Access policies → "Cost confirmation (USD)". A running generation shows a progress card with cancel. Every file lands in the Library under Generated, with its prompt, model and cost, so a team can see which model earned its keep.
ZDR narrows the menu
Zero data retention (ZDR) means the provider does not keep the prompt or answer after the request. If a team handles customer records, an admin can turn off Models Without Zero Data Retention for its permission group, and only ZDR models will run.
That changes the choice above. Pick classification, extraction and writing models from the ZDR-tagged ones, and expect media generation to be unavailable for that group, since media models have no ZDR endpoints. Many organisations run ZDR-only for finance or support and leave marketing on the wider catalogue. For the full security picture, see model data security and ZDR.
Your keys or the platform key
Under Settings → Model API keys, you can bring your own keys (BYOK) for model providers and some tools. Calls then run under your provider account and agreements. Without your own key, the platform key is used and usage is metered to the workspace. BYOK does not change which model is best for a job; it changes whose contract the call runs under, which matters most for the high-volume, pinned workflow steps.
Controls and governance
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.
Bring your own keys
Use your organisation’s own model provider keys at organisation or workspace level.
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.
Run logs
Every workflow run records each block’s inputs, outputs and errors, so a reviewer can trace how a result was reached.
Deployed versions
Workflows run from a deployed version, so a change is reviewed before it reaches production runs.
PII detection
A Guardrails block detects personal data and blocks or masks it before text reaches the next step.
More on how Agent Serve handles access, data and audit on the security page.
A sensible order
- Leave Brain Auto on in Chat.
- List your workflow AI blocks by monthly volume and test a smaller model on the top three.
- Set default media models per team and keep the cost confirmation on.
- Decide which teams are ZDR-only, then pick their pinned models from the ZDR list.
- Record each choice and its test result in the workflow description, and revisit quarterly.
Questions teams ask
Should we turn Brain Auto off?
Usually not. In Chat, most requests are short and mixed, which is what routing is for. Pin a model when you have a reason: a long piece of work, a model your team tested, or a ZDR requirement.
Does pinning a media model change the reasoning model?
No. + → Media models pins an image, video or audio model for that chat, separately from the reasoning model that answers.
How often should we revisit our choices?
When the catalogue adds a model for a job you run at volume, and at least quarterly. Rerun the same test set, so the comparison is fair.
Where each model choice is made
Chat · Everyday requests
- 01Brain AutoRoutes each request to a suitable model, so short asks do not use the heaviest model.
- 02Pinned modelsMore models for a specific reasoning model; + → Media models to pin image, video or audio.
Workflows · Repeated jobs
- 03Model per blockThe tagging step, the extraction step and the writing step each get their own model.
- 04Test setTwenty to fifty real examples with the answer a person expects.
- 05Run logsInputs and outputs for each block, to compare one model against another.
Admins · Access policies
- 06Defaults and limitsDefault media models, denied models and the cost confirmation threshold.
- 07ZDR and keysThe ZDR-only switch per team, and your own model keys under Settings.
What you need
- A list of your AI jobs by volume: what runs thousands of times a month, what runs a few times
- Twenty to fifty real examples per high-volume job, with the answer a person expects
- Agreement on which teams handle data that must stay on ZDR models
- An admin to set default media models and the cost confirmation threshold per permission group
- Provider agreements and keys, if you plan to bring your own