| Usage Scenario |
Requests/Month |
Input Tokens |
Output Tokens |
Est. Cost/Month |
| Light (1-2 staff, occasional use) |
500 |
50K |
25K |
$0.05 |
| Moderate (3-5 staff, daily use) |
2,000 |
200K |
100K |
$0.20 |
| Heavy (10+ staff, frequent use) |
10,000 |
1M |
500K |
$1.00 |
| Enterprise (50+ staff, all-day use) |
50,000 |
5M |
2.5M |
$5.00 |
Pricing (us-east-1):
- Input: $0.25 per million tokens
- Output: $1.25 per million tokens
- Average request: ~100 input tokens + ~50 output tokens
| Component |
Light |
Moderate |
Heavy |
| ECS Fargate (0.25 vCPU / 512 MB) |
$4.38/mo |
$4.38/mo |
$4.38/mo |
| ALB (always-on) |
$16.43/mo |
$16.43/mo |
$16.43/mo |
| CloudWatch Logs |
~$0.50/mo |
~$1.00/mo |
~$3.00/mo |
| Total Infrastructure |
~$21/mo |
~$22/mo |
~$24/mo |
| Scenario |
Bedrock API |
Infrastructure |
Total |
| Light |
$0.05 |
$21 |
~$21 |
| Moderate |
$0.20 |
$22 |
~$22 |
| Heavy |
$1.00 |
$24 |
~$25 |
Note: The ALB is the dominant cost. For very light usage, consider running the bot on a single EC2 t4g.nano instance (~$4/mo) without a load balancer.
| Model |
Input $/1M tokens |
Output $/1M tokens |
Cost vs Haiku |
| Claude 3.5 Haiku |
$0.25 |
$1.25 |
— (baseline) |
| Claude 3.5 Sonnet |
$3.00 |
$15.00 |
12x more |
| Llama 3.1 70B |
$0.59 |
$0.79 |
~2x more |
| Llama 3.1 8B |
$0.10 |
$0.10 |
~0.5x less |
| GPT-4o-mini |
$0.15 |
$0.60 |
~0.5x less |
- Use Haiku for most requests — It’s fast enough for content management
- Keep responses concise — The system prompt limits output to 1024 tokens
- Monitor with CloudWatch — Set a budget alarm at $50/mo
- Consider Llama 3.1 8B — If you need even lower cost, it’s ~50% cheaper
- Cache common responses — For FAQ-type queries, cache the response
- Use AWS Compute Savings Plans — For 24/7 Fargate usage, save ~20%