---
title: "Switch Models Fast"
description: "Run one prompt across several candidate models, compare their responses and costs, then update the production model string after the evaluation."
canonical_url: "https://vercel.com/academy/ai-gateway/switch-models"
md_url: "https://vercel.com/academy/ai-gateway/switch-models.md"
docset_id: "vercel-academy"
doc_version: "1.0"
last_updated: "2026-08-08T23:27:13.238Z"
content_type: "lesson"
course: "ai-gateway"
course_title: "Using AI Gateway in Production"
prerequisites:  []
---

<agent-instructions>
Vercel Academy — structured learning, not reference docs.
Lessons are sequenced.
Adapt commands to the human's actual environment (OS, package manager, shell, editor) — detect from project context or ask, don't assume.
The lesson shows one path; if the human's project diverges, adapt concepts to their setup.
Preserve the learning goal over literal steps.
Quizzes are pedagogical — engage, don't spoil.
Quiz answers are included for your reference.
</agent-instructions>

# Switch Models Fast

# How Do I Switch Models Without Rewriting My App?

When a new model becomes available, your team may need to compare its quality, latency, and cost with the current model before deciding whether to adopt it.

With direct integrations, that comparison can require another SDK or changed response handling. Through the Gateway, the application change is the model string.

\*\*Note: Quick Answer\*\*

AI Gateway models use the `creator/model-name` format. The AI SDK request and response shape stays consistent when that string changes. Evaluate a candidate with representative prompts or shadow traffic before changing production; use the `models` array for outage fallback, not model evaluation.

## Outcome

Run the order-confirmation prompt across three candidate models, compare the responses and costs, then make the production model change in one place.

## Fast Track

1. Run `pnpm switch-models`
2. Read the three answers and three costs
3. Change the order chat's model string in `app/api/order/route.ts` if a candidate wins

## Hands-on exercise

Requirements:

- A `scripts/switch-models.ts` with a `candidates` array of three model strings: the incumbent `openai/gpt-5.4-mini`, plus `anthropic/claude-sonnet-4.6` and one more from the model list (browse and pick; that's part of the exercise)
- Same order-confirmation prompt for all three, with the menu attached
- Print each answer and its actual cost from the metadata

The shared request shape keeps this comparison in one loop, so the evaluation can focus on output quality, latency, and cost.

\*\*Note: Test before the switch\*\*

A one-line change is still a change to production. Run representative prompts or shadow traffic against the candidate first, then change the primary model when it wins. You can keep the proven model in the `models` fallback array for outages, but fallback traffic does not evaluate a healthy new primary.

## Try It

```bash
pnpm switch-models
```

```
=== openai/gpt-5.4-mini ===
"Two Birria Eclipses and an elote — $13.50. Consommé's on the side, napkins are on you."
Cost: $0.0000702

=== anthropic/claude-sonnet-4.6 ===
"Confirmed: two Birria Eclipse tacos with consommé and one Elote Clásico, $13.50 all in."
Cost: $0.0021675

=== google/gemini-3.1-pro-preview ===
"Order confirmed — two Birria Eclipse ($5.00 each), one Elote Clásico ($3.50): $13.50."
Cost: $0.0009480
```

All three candidates used the same prompt and request code. Choose the model that meets the feature's requirements, then update the production string.

Two issues you may encounter:

**A candidate model errors with "not found."** Model catalogs move. Check the exact string on the model list; and if you're on the free tier, remember it covers a subset, so a brand-new model may need purchased credits to try.

**The new model answers differently than your prompts assume.** Handle that difference through prompt and behavior testing. The integration itself remains unchanged, so your evaluation effort can focus on output quality.

## Commit

```bash
git commit -m "feat(models): add candidate comparison for the order confirmation prompt"
```

## Done-When

- [ ] Three models from at least two companies answered via one loop
- [ ] You know each answer's actual cost
- [ ] You can explain why evaluation traffic tests a candidate while fallback traffic only handles failures

## Solution

```ts filename="scripts/switch-models.ts"
import { generateText } from "ai";
import { MENU } from "../lib/menu";

const candidates = [
  "openai/gpt-5.4-mini", // the incumbent
  "anthropic/claude-sonnet-4.6",
  "google/gemini-3.1-pro-preview",
];

const prompt =
  "Confirm this order back to the customer with a total: two Birria Eclipse " +
  `and one Elote Clásico.\n\n${MENU}`;

for (const model of candidates) {
  const result = await generateText({ model, prompt });
  const cost = result.finalStep.providerMetadata?.gateway?.cost;

  console.log(`=== ${model} ===`);
  console.log(`"${result.text.trim()}"`);
  console.log(`Cost: $${cost}`);
  console.log();
}
```

After the evaluation, update the production model in `app/api/order/route.ts`:

```diff
-    model: "openai/gpt-5.4-mini",
+    model: "anthropic/claude-sonnet-4.6",
```

The code change is small because the evaluation happened first.

## Related Questions

- [How do I survive a provider outage?](/ai-gateway/stay-reliable/provider-outage)
- [How do I route to the cheapest provider automatically?](/ai-gateway/save-money/cheapest-provider-routing)
- [Will AI Gateway work with my existing AI stack?](/ai-gateway/stay-reliable/existing-ai-stack)


---

[Full course index](/academy/llms.txt) · [Sitemap](/academy/sitemap.md)