Guides · 4 min read ·

Custom Models in Cursor: Setup, Split Billing and Routing

Custom Models in Cursor: Setup, Split Billing and Routing

Cursor's best-kept secret is that its agent side accepts custom models: your key, your billing, your model choice, inside the editor you already live in. Most users never open that settings page and keep paying bundled rates for work a metered key would serve cheaper. This guide shows how to wire custom models into Cursor cleanly, split completion from agents across billing, compare keys fairly, and keep every configuration reversible in minutes.

Wire a custom model in minutes

Open Cursor's model provider settings and add a custom endpoint: base URL from the provider docs, your API key, and the exact model string from the catalog. Assign the custom model to agent sessions while leaving inline completion on defaults — this split preserves Cursor's best feature while moving the expensive work to billing you control. Run one familiar ticket to confirm diffs, approvals and behavior feel identical.

Verify methodically: wrong model strings fail confusingly, malformed base URLs time out mysteriously, keys without quota die silently. Check each against provider documentation character by character before blaming the editor. The setup takes minutes when done in order and consumes afternoons when done hopefully — invest the order.

Split completion from agents

Keep inline completion on Cursor's own models — that integration is genuinely hard to reproduce and worth its price. Move agent sessions to your key, where per-token metering beats bundled rates for bursty work. The two bills stay separate, so each side earns its keep transparently instead of hiding inside one subscription.

This split mirrors how the product is actually built: completion and agents are different systems with different economics stapled together. Unstapling the billing lets each compete honestly. Most developers find completion worth paying for and agents worth metering — a conclusion reachable only by separating them first. Revisit the split quarterly; both sides evolve.

Compare keys without fooling yourself

Run identical agent tickets on the default billing and your custom key for a week. Score task success first — a cheap key that stalls tickets is the most expensive option available. Then compare tokens burned per ticket, because providers differ in context efficiency even serving the same nominal model. Finally multiply by rates for the honest ranking.

Require two consecutive winning weeks before changing primaries; single weeks can be luck. Keep both routes configured permanently — providers have outage days and deprecations, and mutual fallback costs nothing idle. Document which key serves by default and when to switch, so crunch weeks run on rails instead of adrenaline.

Multi-provider Cursor without chaos

Configure several providers simultaneously and switch per session or ticket type: flagship for hard problems, cheap fast models for bulk, specialty models for niche work. Cursor handles multiple custom endpoints gracefully; the discipline is yours to maintain. Without written routing, traffic drifts to whatever was configured last — convenience routing that quietly overpays.

Write the routing down: default key, escalation triggers, overflow order for outages. Review quarterly as prices and generations move. The overhead is one document; the payoff is every ticket landing on its best-value route indefinitely. Multi-provider setups look complex and operate simply — the complexity lives in the doc, the simplicity in daily use.

Operate custom models permanently

Review keys and model assignments quarterly against current prices; the value ranking rotates fast enough to obsolete careful choices silently. Set per-key budgets with alerts before team rollout — metered agent work without guardrails funds expensive lessons overnight. Test fallback routes routinely under calm conditions.

Maintain the routing document as living infrastructure with reasons attached. New teammates inherit economics on day one instead of improvising for months. The maintenance burden is an hour per quarter protecting savings that compound every ticket. Custom models in Cursor reward the organized permanently and tax the improvising continuously — choose deliberately.

Frequently asked questions

How do I add a custom model to Cursor?

Open Cursor settings, find the model provider section, add a custom endpoint with base URL and key, then select it for agent sessions. Completion can stay on the default while agents use your key — split billing without split workflows.

Does custom model work for Tab completion too?

Agent sessions accept custom providers; inline completion typically stays on Cursor's own models. Plan the split accordingly: custom keys for agents where the money is, defaults for completion where the magic is.

How do I compare two keys in Cursor?

Compare identical agent tickets on each key for a week: success rate first, then tokens burned, then dollars. Watch model behavior differences — same nominal model can vary by provider.

Can I keep multiple providers configured?

Keep both configured and switch per session or ticket. Provider outages, deprecations and price moves all favor optionality; single-key setups convert each into an incident.

Custom model fails — what first?

Verify the exact model string against the provider catalog, check the base URL character by character, and confirm the key has quota. Nine of ten custom-model failures are naming, URL or billing — in that order.

Related guides

Keep your coding workflow moving

Kintio provides an Anthropic-compatible API for Claude Code and other developer tools. Review the fit for your workload before switching.

Explore Kintio →

Last updated: