Llama 3.3 70B
Amazon Bedrock token pricing and cost optimization reference for Llama 3.3 70B.Rates normalized to USD per 1 Million Tokens ($/1M) with prompt caching and batch API discounts across 3 AWS regions.
meta.llama-3-3-70b-v1:0On Google Cloud Vertex AI
The same model is sold through Vertex AI at one global rate per price type. These are Google's published rates, not the AWS rates above; both hosts are shown because the model is sold on both.
| Rate | Vertex AI, USD per 1M tokens |
|---|---|
| Input | $0.72 |
| Output | $0.72 |
| Cache read | — |
| Batch input | — |
| Batch output | — |
| Input, long context | — |
| Output, long context | — |
| Output, thinking | — |
Vertex AI rates are global: Google publishes one rate covering every region where the model runs, against the per-region AWS prices above. A dash means Google publishes no rate for that price type.
Alternatives to Llama 3.3 70B
| Option | Price vs Llama 3.3 70B | Why look at it |
|---|---|---|
| Llama 4 Maverick 17BCompare → | input-67%-$0.48/1M output+35%+$0.25/1M Standard tier, per 1M tokens | Next step down in price from Meta (lightweight tier) |
| Llama 3.1 70B Latency OptimizedCompare → | input+25%+$0.18/1M output+25%+$0.18/1M Standard tier, per 1M tokens | Next step up in price from Meta |
| Llama 3.2 90BCompare → | inputsame price outputsame price Standard tier, per 1M tokens | Same Llama line, different version (general tier) |
| Llama 3 8BCompare → | input-58%-$0.42/1M output-17%-$0.12/1M Standard tier, per 1M tokens | Also from Meta (lightweight tier) |
Llama 3.3 70B Spend Estimator
Model inference spend by prompt volume, cache hit rate, and batch processing.
Key Cost Levers & Architecture
Practical mechanisms to reduce inference costs with Llama 3.3 70B on AWSCaches repetitive system instructions, schemas, and document prefixes in memory with 5-minute and extended 1-hour TTL tiers.
- Read Rate: $0.18 / 1M tokens
- Write Rate (5m TTL): $0.90 / 1M tokens
- Ideal for: Multi-turn conversations, RAG document queries, structured schemas
Asynchronous bulk processing submitted via Amazon S3 with a standard 24-hour turnaround SLA.
- Batch Input: $0.36 / 1M tokens
- Batch Output: $0.36 / 1M tokens
- Ideal for: Offline classification, synthetic data generation, benchmark evals
AWS Bedrock supports standard on-demand inference, priority reserved throughput, and cross-region routing.
- Standard: Lowest latency, pay-per-token with zero commitments
- Cross-Region Routing: Automatically burst traffic to available regional capacity
- GovCloud: Dedicated isolated compliance partitions with +20% uplift
Rates by region
Live rates across all 3 available AWS cloud regionsDeveloper Quickstart
Invoke Llama 3.3 70B via Amazon Bedrock Converse APIimport boto3
# Amazon Bedrock Converse API invocation
client = boto3.client("bedrock-runtime", region_name="us-east-1")
response = client.converse(
modelId="meta.llama-3-3-70b-v1:0",
messages=[
{
"role": "user",
"content": [{"text": "Summarize key features and cost levers of Llama 3.3 70B."}]
}
],
inferenceConfig={
"maxTokens": 1024,
"temperature": 0.7
}
)
output_text = response["output"]["message"]["content"][0]["text"]
usage = response["usage"]
print(f"Response: {output_text}")
print(f"Usage: {usage['inputTokens']} in, {usage['outputTokens']} out")