mistral-cost-tuning

# Mistral AI Cost Tuning

## Overview
Optimize Mistral AI costs through model selection, token management, caching, batch inference, and budget monitoring. Mistral offers the best price-performance in the market with models from $0.1/M tokens (Ministral/Small) to $0.5/M tokens (Large).

## Prerequisites
- Access to [Mistral AI console](https://console.mistral.ai/) for usage data
- Understanding of current usage patterns
- Database for usage tracking (optional)

## Pricing Reference (as of 2025)

| Model | Input $/M tokens | Output $/M tokens | Best For |
|-------|------------------|--------------------|----------|
| `ministral-latest` (3B) | $0.10 | $0.10 | Simple tasks, edge |
| `mistral-small-latest` | $0.10 | $0.30 | General purpose, fast |
| `codestral-latest` | $0.30 | $0.90 | Code generation |
| `mistral-large-latest` | $0.50 | $1.50 | Complex reasoning |
| `pixtral-large-latest` | $2.00 | $6.00 | Vision + text |
| `mistral-embed` | $0.10 | — | Embeddings |
| Batch API (any model) | **50% off** | **50% off** | Non-realtime bulk |

**Always check [docs.mistral.ai/deployment/laplateforme/pricing](https://docs.mistral.ai/deployment/laplateforme/pricing/) for current rates.**

## Instructions

### Step 1: Cost Calculator

```typescript
const PRICING: Record<string, { input: number; output: number }> = {
  'ministral-latest':      { input: 0.10, output: 0.10 },
  'mistral-small-latest':  { input: 0.10, output: 0.30 },
  'codestral-latest':      { input: 0.30, output: 0.90 },
  'mistral-large-latest':  { input: 0.50, output: 1.50 },
  'pixtral-large-latest':  { input: 2.00, output: 6.00 },
  'mistral-embed':         { input: 0.10, output: 0 },
};

function calculateCost(
  model: string,
  inputTokens: number,
  outputTokens: number,
  isBatch = false,
): number {
  const p = PRICING[model] ?? PRICING['mistral-small-latest'];
  const multiplier = isBatch ? 0.5 : 1.0;
  return ((inputTokens / 1e6) * p.input + (outputTokens / 1e6) * p.output) * multiplier;
}

// Example: 100K requests/month, avg 500 in + 200 out tokens
const monthlySmall = calculateCost('mistral-small-latest', 50_000_000, 20_000_000);
const monthlyLarge = calculateCost('mistral-large-latest', 50_000_000, 20_000_000);
console.log(`Small: $${monthlySmall.toFixed(2)}/month`);  // $11.00
console.log(`Large: $${monthlyLarge.toFixed(2)}/month`);  // $55.00
```

### Step 2: Smart Model Router

```typescript
type TaskComplexity = 'trivial' | 'simple' | 'moderate' | 'complex';

function selectModel(complexity: TaskComplexity): string {
  switch (complexity) {
    case 'trivial':  return 'ministral-latest';       // $0.10/M — yes/no, extract, format
    case 'simple':   return 'mistral-small-latest';   // $0.10/M — classify, summarize, Q&A
    case 'moderate': return 'codestral-latest';       // $0.30/M — code gen, moderate reasoning
    case 'complex':  return 'mistral-large-latest';   // $0.50/M — multi-step reasoning, analysis
  }
}

// Auto-detect complexity by prompt characteristics
function estimateComplexity(prompt: string): TaskComplexity {
  const tokens = Math.ceil(prompt.length / 4);
  if (tokens < 50) return 'trivial';
  if (tokens < 200) return 'simple';
  if (prompt.includes('code') || prompt.includes('analyze')) return 'moderate';
  return 'complex';
}
```

### Step 3: Token Budget Manager

```typescript
class BudgetManager {
  private dailyBudgetUsd: number;
  private monthlyBudgetUsd: number;
  private dailySpend = 0;
  private monthlySpend = 0;
  private lastResetDay = new Date().getDate();

  constructor(dailyBudget: number, monthlyBudget: number) {
    this.dailyBudgetUsd = dailyBudget;
    this.monthlyBudgetUsd = monthlyBudget;
  }

  canAfford(model: string, estimatedInputTokens: number, estimatedOutputTokens: number): boolean {
    const cost = calculateCost(model, estimatedInputTokens, estimatedOutputTokens);
    this.maybeResetDaily();
    return this.dailySpend + cost <= this.dailyBudgetUsd
        && this.monthlySpend + cost <= this.monthlyBudgetUsd;
  }

  recordSpend(model: string, usage: { promptTokens: number; completionTokens: number }): void {
    const cost = calculateCost(model, usage.promptTokens, usage.completionTokens);
    this.dailySpend += cost;
    this.monthlySpend += cost;
    this.checkAlerts();
  }

  private checkAlerts(): void {
    const monthPct = (this.monthlySpend / this.monthlyBudgetUsd) * 100;
    if (monthPct > 90) console.error(`BUDGET CRITICAL: ${monthPct.toFixed(1)}% of monthly budget`);
    else if (monthPct > 80) console.warn(`Budget warning: ${monthPct.toFixed(1)}% of monthly budget`);
  }

  private maybeResetDaily(): void {
    const today = new Date().getDate();
    if (today !== this.lastResetDay) {
      this.dailySpend = 0;
      this.lastResetDay = today;
    }
  }

  report() {
    return {
      daily: { spent: this.dailySpend, budget: this.dailyBudgetUsd },
      monthly: { spent: this.monthlySpend, budget: this.monthlyBudgetUsd },
    };
  }
}
```

### Step 4: Prompt Optimization

```typescript
// Reduce tokens = reduce cost directly
function optimizeForCost(systemPrompt: string): string {
  // Remove filler words
  return systemPrompt
    .replace(/please\s+/gi, '')
    .replace(/I would like you to\s+/gi, '')
    .replace(/\s+/g, ' ')
    .trim();
}

// Before: "I would like you to please provide a comprehensive and detailed explanation of how REST APIs work." (~25 tokens)
// After: "Explain REST APIs concisely." (~6 tokens, 76% reduction)

// Set maxTokens to prevent runaway output
const response = await client.chat.complete({
  model: 'mistral-small-latest',
  messages,
  maxTokens: 200, // Cap output — prevents 4000-token essays
});
```

### Step 5: Batch API for Bulk Workloads

```typescript
// Batch API = 50% cost reduction for non-realtime processing
// Instead of 100K individual API calls at $11/month (small)
// Use batch: $5.50/month for the same work

// Supported endpoints:
// /v1/chat/completions, /v1/embeddings, /v1/fim/completions,
// /v1/moderations, /v1/ocr, /v1/classifications

// See mistral-webhooks-events for implementation details
```

### Step 6: Usage Tracking SQL

```sql
CREATE TABLE mistral_usage (
  id SERIAL PRIMARY KEY,
  model VARCHAR(50) NOT NULL,
  input_tokens INTEGER NOT NULL,
  output_tokens INTEGER NOT NULL,
  cost_usd DECIMAL(10, 6) NOT NULL,
  is_batch BOOLEAN DEFAULT FALSE,
  endpoint VARCHAR(50),
  user_id VARCHAR(50),
  created_at TIMESTAMP DEFAULT NOW()
);

-- Daily cost report
SELECT
  DATE(created_at) AS day,
  model,
  SUM(input_tokens) AS total_input,
  SUM(output_tokens) AS total_output,
  SUM(cost_usd) AS total_cost
FROM mistral_usage
WHERE created_at >= NOW() - INTERVAL '30 days'
GROUP BY 1, 2
ORDER BY 1 DESC, 5 DESC;

-- Highest-cost users
SELECT user_id, SUM(cost_usd) AS cost, COUNT(*) AS requests
FROM mistral_usage
WHERE created_at >= DATE_TRUNC('month', NOW())
GROUP BY 1 ORDER BY 2 DESC LIMIT 10;
```

## Cost Reduction Strategies

| Strategy | Savings | Effort |
|----------|---------|--------|
| Use mistral-small instead of large | 80% | Low |
| Batch API for bulk | 50% | Medium |
| Response caching (temp=0) | 30-80% | Medium |
| Prompt optimization | 20-50% | Low |
| Set maxTokens | 10-40% | Low |
| Use ministral for simple tasks | 90% vs large | Low |

## Error Handling
| Issue | Cause | Solution |
|-------|-------|----------|
| Unexpected costs | Untracked usage | Implement BudgetManager |
| Budget exceeded | No alerts | Set alerts at 80% and 90% |
| Wrong model | No routing logic | Use complexity-based model selection |
| Long responses | No maxTokens | Always set maxTokens |

## Resources
- [Mistral Pricing](https://docs.mistral.ai/deployment/laplateforme/pricing/)
- [Mistral Console](https://console.mistral.ai/)
- [Batch Inference](https://docs.mistral.ai/capabilities/batch/)

## Output
- Cost calculator with current pricing
- Smart model router by task complexity
- Token budget manager with alerts
- Prompt optimization patterns
- Usage tracking SQL schema