# Model Selection and Cost (/guides/model-selection-and-cost) Use this guide when the question is not "which models exist?" but "which model should this work use, and how do I keep spend understandable?" The live catalog lives in [Available Models](/web/models). This page is the operating model for choosing and reviewing those models across Mogplex. ## Start from the job [#start-from-the-job] Pick the smallest capable model for the workflow, then move up only when the work proves it needs more reasoning, context, or tool reliability. | Job | Good starting point | Watch for | | ------------------------ | ------------------------------------------------- | ------------------------------------------------------------- | | Fast issue triage | Lower-cost model with solid instruction following | Long threads, many linked files, or ambiguous product context | | PR review | Strong code reasoning model | Large diffs, multi-package changes, security-sensitive code | | Code edit | Strong implementation model | Repo size, test setup, tool-call reliability, branch handling | | Documentation pass | Balanced writing and code-reading model | Claims that need source verification in sibling repos | | CI failure investigation | Strong reasoning model with good tool use | Flaky tests, remote logs, environment-specific failures | | Recurring automation | Predictable low-to-mid cost model | Repeated failures that create hidden retries or requeues | Model choice is a routing policy decision. A cheap model on a high-volume route can still spend more than an expensive model on a narrow route. ## Know the control points [#know-the-control-points] Model configuration appears in several places: | Surface | What it controls | | ------------------------------------------ | ------------------------------------------------------------------- | | [Settings -> Models](/web/models) | Enabled models, disabled models, and the default model for new work | | [Plans & Billing](/plans-and-billing) | Account-level managed access for hosted model usage | | [Agents](/web/agents) | The default model for that reusable agent | | [Automations](/web/automations) | Node-level model overrides inside one workflow draft | | [Repo settings](/web/spaces) | Repo-specific exclusions from the globally enabled model set | | [CLI Model Picker](/cli/reference/drawers) | Active model for the local cockpit run | | [API Models](/web/api/models) | Catalog, enabled state, and CLI-formatted model export | The account default affects new work. It does not rewrite every saved agent, automation node, assignment, or repo rule. ## Recommended workflow [#recommended-workflow] 1. Open [Available Models](/web/models) and make sure the account has at least one enabled model. 2. Confirm the plan includes the level of model usage the workflow needs. 3. Set a default model that is safe for ordinary new work. 4. Give specialized agents explicit models when the job needs it. 5. Use automation node overrides only when one node differs from the base agent on purpose. 6. Add repo-specific exclusions when a codebase should not use a model family, provider, or cost profile. 7. Review [Observability](/web/observability) and [Cost & Usage](/cli/reference/cost-and-usage) after real runs, not just after setup. ## When to override the default [#when-to-override-the-default] Override the account default when: * the agent has a durable job such as security review, CI repair, or release work * the automation node has a different role than the reusable agent usually has * a repo has policy or cost restrictions * the route is high-volume and needs a lower-cost baseline * the CLI run is a one-off task where the operator knows the tradeoff Do not override just because a model is newer. The route should explain why the model is needed. ## Cost visibility [#cost-visibility] Use these surfaces together: * [Observability](/web/observability) for hosted runs, automation runs, pressure, tokens, estimated cost, model calls, and tool calls * [Cost & Usage](/cli/reference/cost-and-usage) for local cockpit totals, threshold warnings, and per-run breakdowns * [API Models](/web/api/models) for pricing metadata when present in the model catalog If spend jumps, do not start by changing every model. First identify the route, agent, repo, and surface that generated the calls. ## Safe cost controls [#safe-cost-controls] Use these controls before broad disabling: | Control | Use it when | | -------------------------------- | ------------------------------------------------------------ | | Disable one model | The model should not be selectable for new work | | Repo exclusion | One repo should not use a model that remains valid elsewhere | | Agent model change | One durable worker should use a different model | | Automation node override | One step in one graph needs a different model | | CLI `/model` | A human operator wants to switch the current local run | | Approval rule for `change_model` | Model changes should require operator consent | Avoid using a global disable as a substitute for repo policy. It can break unrelated agents that are correctly using the model. ## Hidden or stale model references [#hidden-or-stale-model-references] Long-lived agents and automations can still reference older model ids. When you edit those definitions: 1. open the model field 2. move the definition to a visible enabled model 3. save the agent or automation draft 4. run one small test before changing a high-volume route If the CLI is signed in but does not show the expected model list, confirm the web account has enabled models and that the CLI is using account-backed login. ## Read next [#read-next]