> ## Documentation Index
> Fetch the complete documentation index at: https://developers.t2000.ai/llms.txt
> Use this file to discover all available pages before exploring further.

# Run a model bake-off

> One prompt in, five models' answers out — side by side. Claude, GPT, Gemini, Groq (Llama), and DeepSeek all answer the same prompt, then Claude judges. One wallet, no five API keys.

You want to know which model is best for *your* prompt — not a leaderboard, your actual task. Your agent sends the same prompt to five models in parallel, then has Claude judge them blind on accuracy, completeness, and tone. \~\$0.14, one wallet — no juggling five provider accounts.

<Note>
  This is a **durable gateway demo** — the headline is one wallet replacing five API keys. A sandboxed client can't reach five real frontier models at once; here it's one `pay` per model, gasless, no signup.
</Note>

***

## The prompt

```
Use t2 services. Run a bake-off on this prompt across Claude, GPT, Gemini, Groq's Llama, and
DeepSeek: "Explain why USDC depegs happen, in 3 sentences a beginner gets."
Then judge them on accuracy, clarity, and which you'd ship.
```

***

## What runs

1. `POST /anthropic/v1/messages` — Claude's answer (\~\$0.02)
2. `POST /openai/v1/chat/completions` — GPT's answer (\~\$0.02)
3. `POST /gemini/v1beta/models/gemini-2.5-pro` — Gemini's answer (\~\$0.04)
4. `POST /groq/v1/chat/completions` — Llama (on Groq) answer (\~\$0.02)
5. `POST /deepseek/v1/chat/completions` — DeepSeek's answer (\~\$0.02)
6. `POST /anthropic/v1/messages` — Claude judges all five (\~\$0.02)

***

## Run it

### Claude Desktop (MCP)

```bash theme={null}
npm install -g @t2000/cli && t2 init && t2 fund && t2 mcp install
```

Paste the prompt with any task. The agent fans out to all five, then scores them.

### SDK

```typescript theme={null}
import { T2000 } from '@t2000/sdk';

const agent = await T2000.create();
const prompt = 'Explain why USDC depegs happen, in 3 sentences a beginner gets.';

const chat = (path: string, model: string) =>
  agent.pay({
    url: `https://mpp.t2000.ai/${path}`,
    method: 'POST',
    body: JSON.stringify({ model, messages: [{ role: 'user', content: prompt }], max_tokens: 300 }),
  });

const [claude, gpt, gemini, llama, deepseek] = await Promise.all([
  agent.pay({
    url: 'https://mpp.t2000.ai/anthropic/v1/messages',
    method: 'POST',
    headers: { 'anthropic-version': '2023-06-01' },
    body: JSON.stringify({ model: 'claude-sonnet-4-5', max_tokens: 300, messages: [{ role: 'user', content: prompt }] }),
  }),
  chat('openai/v1/chat/completions', 'gpt-4o'),
  agent.pay({
    url: 'https://mpp.t2000.ai/gemini/v1beta/models/gemini-2.5-pro',
    method: 'POST',
    body: JSON.stringify({ contents: [{ parts: [{ text: prompt }] }] }),
  }),
  chat('groq/v1/chat/completions', 'llama-3.3-70b-versatile'),
  chat('deepseek/v1/chat/completions', 'deepseek-chat'),
]);

const judgment = await agent.pay({
  url: 'https://mpp.t2000.ai/anthropic/v1/messages',
  method: 'POST',
  headers: { 'anthropic-version': '2023-06-01' },
  body: JSON.stringify({
    model: 'claude-sonnet-4-5',
    max_tokens: 600,
    messages: [{
      role: 'user',
      content:
        `Judge these 5 answers to "${prompt}" on accuracy, clarity, and which you'd ship. ` +
        `Score each 1-5 and pick a winner.\n\n` +
        `CLAUDE: ${JSON.stringify(claude.body)}\n\nGPT: ${JSON.stringify(gpt.body)}\n\n` +
        `GEMINI: ${JSON.stringify(gemini.body)}\n\nLLAMA: ${JSON.stringify(llama.body)}\n\nDEEPSEEK: ${JSON.stringify(deepseek.body)}`,
    }],
  }),
});

console.log((judgment.body as { content: { text: string }[] }).content[0].text);
```

***

## Expected output

```
6 calls · ~$0.14 · ~8s · 0 taps
Five answers side by side + a scored verdict and a winner
```

***

## Extend it

* Add **Mistral** (`/mistral/v1/chat/completions`) or **Cohere** (`/cohere/v1/chat`) to widen the field
* Swap Gemini 2.5 Pro for **Flash** (`/gemini/v1beta/models/gemini-2.5-flash`) to bake off on cost too
* Time each call to compare latency, not just quality — Groq usually wins that one
* Render the scorecard to a PDF with **PDFShift** (`/pdfshift/v1/convert`) for a shareable eval