DataLLM LabAWS Bedrock Pricing in 2026: Modes, Real Costs & Break-Even Math
AWS Bedrock pricing has three modes — on-demand, provisioned throughput, and batch — and picking the wrong one is the most common way teams overpay. Bedrock passes through each model's per-token price, so the model you pick and the mode you run it in are what drive cost. This guide breaks down the three modes, models what popular models actually cost per month, works through the provisioned-throughput break-even, quantifies the batch discount, lists the traps, and shows when a gateway is the cheaper route.
The three pricing modes
| Mode | How you pay | Best for |
|---|---|---|
| On-demand | Per token, no commitment | Most workloads, bursty traffic |
| Provisioned throughput | Reserved capacity, hourly (1/6-mo discounts) | Steady high-volume production |
| Batch | ~50% off on-demand, async | Offline/bulk, latency-tolerant jobs |
What models cost
Bedrock passes through each model's per-token price, so cost is driven by which model you choose. Claude on Bedrock, by tier:
Modeled monthly cost
Per-token rates are abstract; here's what Claude on Bedrock costs per month across five workloads (on-demand):
| Monthly workload | Claude Opus 4.7 | Claude Sonnet 4.6 | Claude Haiku 4.5 |
|---|---|---|---|
| Support chatbot | $500 | $300 | $100 |
| RAG / knowledge base | $1,500 | $900 | $300 |
| Coding agent | $1,025 | $615 | $205 |
| Batch extraction | $950 | $570 | $190 |
| Content generation | $1,100 | $660 | $220 |
The biggest lever within Claude on Bedrock is tier choice — Haiku runs a RAG workload for $300 versus Opus at $1,500. Mode choice (below) is the second lever.
Provisioned throughput break-even
Provisioned throughput reserves dedicated capacity billed by the hour (cheaper with a 1- or 6-month commitment). The trap: you pay for the reservation whether you use it or not. So the break-even is about utilization, not raw volume:
- Rule of thumb: provisioned only beats on-demand above roughly 50-70% sustained utilization of the reserved capacity.
- Worked logic: if your traffic is steady 24/7 near the reserved throughput, the committed hourly rate (with discount) undercuts paying per token. If it's bursty — busy 8 hours, idle 16 — you're paying for ~16 idle hours, and on-demand (which charges nothing while idle) wins.
- Decision: measure your actual hourly utilization for a week on on-demand first. Commit to provisioned only once you can see sustained, predictable load near capacity.
In short: start on-demand, graduate to provisioned when the utilization data justifies it — never the reverse.
The batch discount, quantified
Bedrock's batch inference runs asynchronously at roughly 50% of on-demand. Concretely, the modeled batch-extraction workload that costs $950/month on-demand on Claude Opus drops to about $475 on batch. Any latency-tolerant job — overnight enrichment, evaluation runs, bulk generation — left on on-demand is paying roughly double for no benefit.
Which mode to choose
Default On-demand
- Bursty or unproven traffic. Pay only for what you use; no commitment.
Steady scale Provisioned
- Predictable 24/7 load near capacity (>50-70% utilization). Commit for the discount.
Offline Batch
- Latency-tolerant bulk jobs → ~50% off automatically.
Lowest price Cheaper model
- Route easy work to Haiku/Nova/Llama instead of Opus — bigger than any mode discount.
Cost traps to avoid
- Provisioned for bursty traffic — idle reserved capacity is the #1 overspend.
- Ignoring batch — any latency-tolerant job left on on-demand pays roughly double.
- Flagship for everything — route easy work to cheaper Bedrock models (Nova, Llama, Haiku).
- Forgetting surrounding AWS costs — data transfer, storage, and orchestration add up beyond token price.
Bedrock vs direct vs gateway
| AWS Bedrock | Direct API | Gateway | |
|---|---|---|---|
| Token price | Provider rate | Provider rate | Route to cheapest |
| AWS-native (IAM, VPC) | Yes | No | No |
| Cross-provider routing | Within Bedrock | No | Yes, 300+ models |
| Best for | AWS compliance | One provider | Lowest price + flexibility |
Bedrock wins on AWS-native compliance; a gateway wins when you want the cheapest equivalent model across providers with one key.
Want the cheapest equivalent, not just AWS-native?
DataLLM Lab routes across 300+ models with live price comparison — pick the lowest-cost model that clears your quality bar, one OpenAI-compatible key.
FAQ
How does AWS Bedrock pricing work?
Per token at each model's rate, in three modes: on-demand (no commitment), provisioned throughput (reserved capacity hourly), and batch (~50% off, async). Plus normal AWS costs for surrounding services.
Is Bedrock cheaper than the provider directly?
Usually the per-token price matches the provider's direct rate — no token markup for models like Claude. Choose Bedrock for AWS-native integration; a gateway for the lowest cross-provider price.
When does provisioned throughput pay off?
Above ~50-70% sustained utilization — you pay for the reservation whether used or not. For bursty/low traffic, on-demand is far cheaper. Measure utilization on on-demand first, then commit.
Does Bedrock have a batch discount?
Yes — async batch is ~50% of on-demand. A $950/mo batch-extraction workload on Opus drops to ~$475 on batch.
Which models are on Bedrock?
Anthropic (Claude), Meta (Llama), Amazon (Nova/Titan), Mistral, Cohere, and more. Note: Anthropic server-side tools and Managed Agents aren't available via Bedrock.
How much does Claude cost on Bedrock?
Passthrough Anthropic rates — Opus $5/$25, Sonnet $3/$15, Haiku $1/$5. On RAG, ~$1,500/$900/$300 per month respectively. Tier choice is the biggest lever.
When is a gateway cheaper than Bedrock?
When you want the cheapest equivalent across providers — a gateway like DataLLM Lab routes to the lowest-cost model with one key, no provisioned commitment.
Are there hidden costs in Bedrock pricing?
Yes — data transfer, knowledge-base storage, orchestration, and especially idle provisioned capacity. Model the full picture, not just the token rate.