Choosing an AI model for a worker

What this is

A worker is an identity: its name, role, manager, memory and history. The model is what it thinks with, and you can change it at any time without changing who the worker is. This page is about choosing that model for a kind of work. It is general guidance; the provider pages linked at the bottom are the authority on names, capabilities and prices, and they change often.

Who it helps

An owner deciding which provider account to connect first, and a team lead choosing a model per worker rather than one for everyone.

The shape of every provider's line-up

Each major provider sells a small ladder. The names differ, the shape does not:

• A frontier tier for demanding reasoning, long tasks and careful judgement. Slowest and most expensive per word.
• A balanced tier with most of the capability at a fraction of the cost and latency. The right default for most business work.
• A fast, economical tier for high-volume, simple tasks: classification, extraction, monitoring, first-pass drafts.

On the reviewed date the provider pages named, for example, Claude Fable 5.1, Claude Opus 5, Claude Sonnet 5 and Claude Haiku 4.5 at Anthropic; the GPT-6 and GPT-5.6 families at OpenAI; Gemini 3.1 Pro and the Gemini 3 Flash and Flash-Lite models at Google; and the Grok 4 series at xAI. Names and tiers move; check the links rather than this paragraph.

Matching the tier to the work

Use a different model per worker. A monitor and a manager have different jobs; giving them the same expensive model wastes money, and giving them the same cheap one wastes your time.

What drives the cost

Providers charge per token, roughly per word, separately for what the model reads (input) and what it writes (output). Three things move the bill:

1. How much the worker reads. Long documents, long histories and long web pages all count. Routines that re-read the same pages every day add up.
2. How much it writes. Verbose output costs more than a five-bullet brief. Say what "done" looks like.
3. Which tier. The frontier tier can cost several times the balanced tier per token. Reserve it for work that needs it.

Every provider publishes its per-token prices; the current figures are on the pages linked below and are not repeated here because they change. In Korps you connect your own provider account and pay the provider directly; Korps shows usage in dollars against a ceiling you set. A Korps-managed AI option is planned and its providers are not decided.

Which providers Korps connects to

Direct connections on your own account: Anthropic (Claude), OpenAI, xAI (Grok), Google (Gemini), OpenRouter, and a local Ollama host. Families such as Llama, Mistral, DeepSeek and Qwen are reached through your OpenRouter key or your own Ollama server rather than as separate Korps integrations. A Cursor plan powers Cursor workers. Korps stores no provider credential: the API refuses a request that carries one, and the key stays on the machine your workers run on.

Data and residency

Where your data goes is a model decision as much as a tool decision. A provider's own terms say whether prompts are retained or used for training, and where they are processed; read them for the account type you have. For regulated data, the model, the provider and the route all need to be approved before a worker reads it. A local Ollama server keeps inference on your own machine at the cost of capability.

Requirements and availability

• General guidance. Model tiers, costs and fit are educational; the provider pages are the authority.
• Available in Korps: direct connections to the providers above, per-worker model choice, changing a worker's model without losing its identity, and usage shown in dollars against a ceiling.
• Planned for Korps: a Korps-managed AI usage option.
• A provider account with billing enabled is yours to set up; Korps cannot create one for you.

Limitations and approval points

• No model choice removes the need for a link on every claim and a person on every send. Stronger models are wrong less often, not never.
• Benchmarks published by providers measure their tests, not your work. Try two tiers on a real assignment and compare.
• Prices, names and capabilities on this page were checked on the reviewed date and will drift; treat the linked sources as current and this page as orientation.