The on demand token pricing curve, the provisioned throughput unit commitment, the model selection and switching cost, the customization and fine tuning fee structure, the data retention and privacy clauses, and the AWS EDP commit interaction for the AWS Bedrock enterprise commitment.
A working playbook for CIOs, CFOs, and software asset managers committing to AWS Bedrock in 2026, with the six dimension framework, the model version lock mechanics, and the multi cloud AI moves that recover twenty two to thirty eight percent against the AWS account team's opening commitment proposal.
AWS Bedrock is the load bearing AI inference layer for enterprises that have standardized on AWS. The licensing surface is unfamiliar to most CIOs because Bedrock combines four commercial mechanics that look nothing like the traditional software licensing motion: token metered inference pricing, provisioned throughput commitments, customization and fine tuning fees, and the AWS EDP commit overlay. Each of these mechanics carries discount levers, but the levers are visible only when the buyer separates the four mechanics at the contract negotiation and prices each one against documented engagement benchmarks rather than the AWS account team's preferred bundled commitment.
This paper sets out the Redress Compliance buyer side framework across the six dimensions of the Bedrock commercial surface: the on demand token pricing curve, the provisioned throughput unit commitment, the model selection and switching cost, the customization and fine tuning fee structure, the data retention and privacy clause, and the AWS EDP commit interaction. Each dimension is run against documented engagement evidence from more than five hundred enterprise software negotiations across the practice. Read the related AWS services practice, the AWS EDP negotiation guide, the GenAI vendors services practice, and the AI platform contract negotiation download.
Used across the practice, the framework typically delivers twenty two to thirty eight percent savings against the AWS account team's opening Bedrock commitment proposal, with the upper end of the range available when the buyer credibly opens the multi cloud AI conversation against Azure OpenAI, Google Vertex AI, or Anthropic direct at the same EDP renewal cycle.
The enterprise generative AI procurement conversation reached an inflection point in 2026. The early adoption phase, where lines of business signed individual Anthropic, OpenAI, or Google Vertex AI contracts outside the central IT procurement function, has given way to a consolidation phase where the CIO and the central procurement team are pulling AI spend into the established hyperscaler agreements. AWS Bedrock sits at the center of this consolidation because Bedrock surfaces multiple model families through a single API and runs inside the existing AWS account structure, which means Bedrock spend can be folded into the existing AWS EDP commitment rather than negotiated as a separate vendor contract.
The Bedrock model portfolio in 2026 includes the Anthropic Claude family, the Amazon Nova family, the Meta Llama family, the Mistral family, the Cohere family, the AI21 Jurassic family, and the Stability AI family. AWS charges for inference at a per token rate that varies by model and by region. The published per token rates for the Claude Sonnet class models sit at three dollars per million input tokens and fifteen dollars per million output tokens at the small footprint, with negotiated rates moving below the published rates at the higher commitment scale. The Amazon Nova family carries lower per token rates, with Amazon Nova Pro at three dollars per million input tokens and twelve dollars per million output tokens at the published rate.
The financial stakes are substantial and growing. A mid market enterprise running a single high traffic GenAI application can clear two to four million dollars per year in Bedrock inference spend at the published rates. A large enterprise running multiple production GenAI applications across customer service, code generation, document processing, and internal knowledge retrieval can clear ten to thirty million dollars per year in Bedrock inference spend. The AWS account team consistently pushes the enterprise toward the provisioned throughput commitment model at the higher spend tiers, with the framing that provisioned throughput delivers predictable cost and lower per token rates. The framing is partially true and substantially incomplete.
The provisioned throughput model carries documented trap mechanics the buyer side framework needs to anchor against. Provisioned throughput model units are purchased in one month, six month, or twelve month commitment terms. The shorter terms carry a substantial premium against the longer terms. The longer terms lock the customer into a specific model version, and AWS does not allow free switching between model versions inside the commitment term. A customer that commits to twelve months of Claude Sonnet 3.5 provisioned throughput and decides six months later to move to Claude Sonnet 4 or Claude Opus 4 typically forfeits the remaining commitment or pays a stacked commitment until the original term ends. The AWS account team frames this constraint as a technical limitation rather than a commercial one, but it is unambiguously a commercial commitment trap.
The competitive pressure on AWS Bedrock is also real. Azure OpenAI runs a similar provisioned throughput unit model under the name Provisioned Throughput Units. Google Vertex AI runs a similar commitment model under the name Provisioned Throughput. Anthropic direct, OpenAI direct, and Cohere direct all offer enterprise contracts that compete with the hyperscaler resold versions of their own models. The CIO who runs the Bedrock commitment alongside the Azure OpenAI commitment, the Vertex AI commitment, and the direct vendor relationships has a real multi cloud AI conversation to anchor against, and the practice has documented engagements where credible multi cloud AI pressure recovered twenty three to thirty one percent at the Bedrock commitment without any actual workload migration. Read the related AWS Azure GCP competitive framework, the AWS EDP negotiation guide, and the GenAI knowledge hub.
The buyer side framework therefore runs against three structural realities. First, the Bedrock commitment is not a software license. It is a metered consumption contract with provisioned throughput overlays, and the commercial levers are different from the traditional enterprise software motion. Second, the provisioned throughput model carries commitment trap mechanics that the buyer needs to neutralize at the contract negotiation. Third, the multi cloud AI alternatives are real and the AWS account team will discount aggressively when the buyer credibly opens the multi cloud conversation.
The first commercial dimension at AWS Bedrock is the pricing model selection. AWS offers two principal pricing models on Bedrock: on demand token metered pricing and provisioned throughput unit commitment pricing. The two models are commercially asymmetric and the AWS account team will frame the choice in the publisher's favor unless the buyer separates the two models at the contract negotiation.
On demand token pricing meters the inference call at the input token and output token level, with the per token rate set by the model family and the region. The published per token rates for the principal model families in early 2026 sit in the following ranges. The Claude Sonnet class models run at three dollars per million input tokens and fifteen dollars per million output tokens. The Claude Opus class models run at fifteen dollars per million input tokens and seventy five dollars per million output tokens. The Claude Haiku class models run at twenty five cents per million input tokens and one dollar twenty five cents per million output tokens. The Amazon Nova Pro model runs at three dollars per million input tokens and twelve dollars per million output tokens. The Llama 3 70B model runs at two dollars sixty five cents per million input tokens and three dollars fifty cents per million output tokens.
The on demand model has two structural advantages for the buyer. First, the customer pays only for actual token consumption with no commitment overhead. Second, the customer can switch between model families and model versions inside the same API without commitment penalties. The on demand model also has two structural disadvantages. First, the per token rate is higher than the equivalent provisioned throughput rate at the higher commitment scale. Second, the on demand model is subject to AWS service quotas and rate limits that constrain high throughput applications during peak load.
Provisioned throughput unit pricing commits the customer to a specified number of model units at a fixed hourly rate for a one month, six month, or twelve month term. The model unit corresponds to a specific throughput capacity expressed in tokens per minute or in inference requests per minute, depending on the model family. The published per model unit rates vary substantially across the model families. The Claude Sonnet 3.5 provisioned throughput rates run at thirty nine dollars per model unit per hour at the six month commitment and at twenty two dollars per model unit per hour at the twelve month commitment. Higher commitment terms deliver per unit rates that sit approximately forty to fifty five percent below the on demand rate at the same throughput level.
The buyer side response anchors the pricing model selection against the customer's actual inference workload pattern. A bursty workload pattern with substantial peak to trough variance favors the on demand model because the provisioned throughput model bills for the committed capacity regardless of actual utilization. A steady workload pattern with a predictable token rate favors the provisioned throughput model at the higher commitment scale. The practice has documented engagements where the AWS account team pushed the customer toward the twelve month provisioned throughput commitment when the customer's actual workload pattern was bursty and the on demand model would have delivered lower total spend even at the higher per token rate.
The second commercial dimension is the model selection, model version lock, and switching cost. This is the dimension where the AWS account team holds the most leverage because the provisioned throughput commitment locks the customer to a specific model version for the contracted term, and the model version landscape in 2026 changes substantially across a twelve month horizon.
The model families surfaced through Bedrock evolve at different cadences. The Anthropic Claude family released Claude Sonnet 3.5 in mid 2024, Claude Sonnet 4 in early 2025, Claude Sonnet 4.5 in mid 2025, Claude Opus 4 in early 2026, and Claude Opus 4.6 in mid 2026. The Amazon Nova family released Nova Pro in late 2024 and Nova Premier in early 2026. The Meta Llama family released Llama 3.1 in late 2024, Llama 3.3 in early 2025, and Llama 4 in early 2026. The pace of model release means a twelve month provisioned throughput commitment locked at the start of 2026 to Claude Sonnet 3.5 will be running on a two year old model by the end of the commitment term, with the customer paying provisioned throughput rates while newer model versions are available at lower on demand rates.
The provisioned throughput commitment specifies the model version, not the model family. A commitment to Claude Sonnet 3.5 provisioned throughput does not entitle the customer to Claude Sonnet 4 or Claude Opus 4 inference inside the commitment. A customer that decides to upgrade the production application to the newer model typically faces three options. The first option is to forfeit the remaining provisioned throughput commitment and absorb the loss. The second option is to stack a new provisioned throughput commitment on top of the existing commitment, doubling the throughput spend for the overlapping period. The third option is to negotiate a commitment conversion clause with the AWS account team, which is not part of the standard contract and which the AWS account team typically declines unless the buyer raises the conversion clause at the original contract negotiation.
The buyer side response neutralizes the version lock mechanic at the original contract negotiation with three specific contract clauses. First, the commitment conversion clause that allows the customer to convert a portion of the committed throughput to a newer model version once per quarter at no additional cost. Second, the commitment downgrade clause that allows the customer to scale down the committed throughput at the contracted anniversary up to a defined percentage of the original commitment. Third, the rate match clause that obligates AWS to match the published on demand rate for the newer model version when the published on demand rate falls below the contracted provisioned throughput rate. The practice has documented engagements where all three clauses were inserted at the original contract negotiation, with the AWS account team agreeing to the clauses when the buyer credibly opened the Azure OpenAI multi cloud conversation. Read the related AWS Azure GCP competitive framework.
The third commercial dimension is the model customization, fine tuning, and data retention surface. This dimension carries hidden cost mechanics that show up only when the contract is read in detail, and the AWS account team typically does not lead with these mechanics in the commercial pitch.
AWS Bedrock supports custom model import for the Llama, Mistral, and Mixtral model families, and supports fine tuning for the Claude Haiku, Amazon Titan, and Cohere Command families. The custom model import service carries an hourly fee per imported model in the storage tier and a per token inference rate when the imported model is invoked. The fine tuning service carries a per token training fee, an hourly storage fee per fine tuned model, and a separate per token inference rate for the fine tuned model. The fine tuning per token training fee sits at twelve cents per million tokens for the Claude Haiku family at the published rate, and the fine tuned model inference rate sits approximately fifteen to twenty percent above the base on demand rate for the same model family.
The buyer side response anchors the fine tuning and custom model spend at the original contract negotiation. The practice has documented engagements where the AWS account team did not surface the fine tuned model inference premium in the original commitment conversation, and the customer discovered the premium only after the fine tuned model went into production. The corrective move is to price the fine tuned model inference rate as a separate line item at the original contract negotiation and to require AWS to commit to a defined premium ceiling above the base on demand rate for the same model family.
AWS Bedrock layers three managed services above the inference layer: Knowledge Bases for managed retrieval augmented generation, Agents for managed tool use, and Guardrails for managed safety filtering. Each of these services carries a separate metered consumption fee on top of the inference fee. The Knowledge Bases service charges for vector storage and for retrieval requests. The Agents service charges for agent invocation and for the underlying inference. The Guardrails service charges for content evaluation at a per token rate. The combined fee for a production GenAI application using all three managed services typically sits at thirty to fifty percent above the base inference fee at the same token volume.
The buyer side response prices the managed services separately at the original contract negotiation and inserts a managed service commitment cap that limits the combined managed service fee to a defined percentage of the underlying inference spend. The cap recovers measurable spend across the contracted term.
The data retention and privacy clauses are the most easily overlooked commercial dimension. The standard Bedrock terms allow AWS to retain inference logs for service operation purposes and to use anonymized inference data for service improvement at AWS's discretion. The standard terms also allow Anthropic, Meta, Mistral, and other model providers to retain a portion of the inference data for service improvement under the model provider agreement. Customers in regulated financial services, healthcare, life sciences, and government industries typically need stricter data retention and privacy terms than the standard Bedrock contract provides.
The buyer side response inserts three specific data clauses at the original contract negotiation. First, a zero retention clause that requires AWS to delete inference logs within a defined retention window. Second, a no training clause that prohibits AWS and the model provider from using the customer's inference data for service improvement or model training. Third, a data residency clause that requires AWS to process inference requests in a defined set of regions. The practice has documented engagements where all three clauses were inserted at the original contract negotiation. Read the related AI platform contract negotiation download.
The fourth commercial dimension is the interaction between Bedrock spend and the AWS Enterprise Discount Program commit. This is the dimension where the AWS account team holds the most structural leverage and where the buyer can recover the most measurable spend with disciplined commitment scoping.
Bedrock spend rolls up into the AWS EDP commitment by default, which means a customer with a fifty million dollar three year EDP commitment counts Bedrock inference fees, provisioned throughput fees, customization fees, and managed service fees against the EDP commit. The rollup is commercially attractive because it converts variable Bedrock spend into commitment burn down, but the rollup also masks the actual Bedrock cost profile inside the broader EDP report and makes it harder to negotiate the Bedrock specific discount layer at the next EDP renewal.
The buyer side response separates the Bedrock spend from the EDP commit at the reporting layer even when the spend rolls up at the billing layer. The practice has documented engagements where the buyer required the AWS account team to provide a separate Bedrock spend report on a quarterly cadence and to negotiate the Bedrock specific discount layer as a distinct line item at the EDP renewal.
AWS offers a Bedrock specific discount layer above the standard EDP discount band when the Bedrock spend exceeds defined thresholds inside the EDP commitment. The published discount layer carries a five to twelve percent additional discount above the EDP rate at the higher Bedrock spend tier. The discount layer is not advertised in the standard EDP commercial pitch, and the AWS account team typically does not surface the discount layer unless the buyer raises Bedrock spend as a separate commitment conversation.
The buyer side response raises the Bedrock specific discount layer at the EDP renewal and requires the AWS account team to apply the discount layer retroactively for the prior twelve months of Bedrock spend. The practice has documented engagements where the retroactive discount recovered six hundred thousand to two million dollars per year on Bedrock spend across the prior commitment cycle. Read the related AWS EDP negotiation download, the AWS EDP flexibility provisions download, and the AWS EDP commitment calculator.
The EDP shortfall risk is the final structural mechanic. A customer with a fifty million dollar three year EDP commitment that falls below the contracted commitment burn rate faces a stacked commitment risk at the EDP renewal, where AWS uses the prior commitment shortfall as the anchor for the next commitment proposal. Bedrock spend interacts with the shortfall risk because Bedrock spend is one of the few AWS spend categories where the customer can credibly accelerate spend in the final year of the EDP term to clear the commitment without ordering unused capacity. The buyer side response uses the Bedrock spend acceleration as a leverage point at the EDP renewal rather than as an emergency tactic in the final commitment year.
The standard provisioned throughput commitment delivers forty to fifty five percent below the on demand rate at the same throughput level. The Bedrock specific discount layer adds a further five to twelve percent on the EDP rolled up spend at the higher Bedrock spend tier. The buyer side framework anchors both layers at the contract negotiation rather than at the EDP renewal alone.
Not without a commitment conversion clause. The standard provisioned throughput commitment specifies the model version, not the model family. The buyer side response inserts a commitment conversion clause that allows the customer to convert a portion of the committed throughput to a newer model version once per quarter at no additional cost.
Bedrock spend rolls up into the AWS EDP commit by default, which means inference fees, provisioned throughput fees, and customization fees count against the EDP commit. The buyer side response separates Bedrock spend at the reporting layer and negotiates the Bedrock specific discount layer as a distinct line item at the EDP renewal.
The standard Bedrock terms allow AWS to retain inference logs for service operation and to use anonymized inference data for service improvement. The model providers also retain a portion of inference data under the model provider agreement. The buyer side response inserts zero retention, no training, and data residency clauses at the original contract negotiation.
The Bedrock commitment negotiation should start one hundred twenty days before the EDP renewal or the standalone Bedrock commitment renewal. The multi cloud AI benchmark needs at least sixty days to complete, and the contract redlines typically require four to six weeks to clear the AWS commercial approval cycle.
The fine tuned model inference rate typically sits fifteen to twenty percent above the base on demand rate for the same model family, plus the customer pays the per token training fee and the hourly fine tuned model storage fee. The buyer side response prices the fine tuned model inference rate as a separate line item at the original contract negotiation.
Yes. Azure OpenAI, Google Vertex AI, and Anthropic direct all offer enterprise contracts that compete with the Bedrock resold versions of the same models. The practice has documented engagements where credible multi cloud AI pressure recovered twenty three to thirty one percent at the Bedrock commitment without any actual workload migration.
The three managed services layer on top of the base inference fee at a combined rate of thirty to fifty percent of the underlying inference spend. The buyer side response prices each managed service as a separate line item and inserts a combined managed service commitment cap at a defined percentage of the inference spend.
The practice runs four engagement models against the AWS Bedrock commitment. The Vendor Shield always on advisory subscription covers the Bedrock commitment alongside the broader AWS estate. The Renewal Program runs a structured twelve month managed sequence around the EDP renewal and the Bedrock commitment cycle. The Benchmark Program sizes the commitment against more than five hundred documented engagements. The software spend assessment sizes the Bedrock commitment alongside the broader AWS EDP footprint. Read the related AWS services practice, the AWS EDP negotiation guide, the AWS Azure GCP competitive framework, and the GenAI knowledge hub.
The AWS EDP commit framework with the Bedrock interaction, the multi cloud AI competitive conversation, the shortfall risk framework, and the buyer side moves at the renewal cycle.
Used across more than five hundred enterprise software engagements. Independent. Buyer side. Built for CIOs running the next AWS EDP renewal alongside the Bedrock and managed AI commitment.
The AWS team pushed a twelve month Claude Sonnet provisioned throughput commitment as the only path. Redress inserted the commitment conversion clause, the managed services cap, and the multi cloud benchmark in parallel. Twenty nine percent saving against the opening Bedrock proposal at the EDP renewal.
We work for the buyer. Always. There is no other side of our table.
AWS EDP framework signals, Bedrock commitment signals, multi cloud AI signals, and the broader enterprise AI procurement signals from the Redress Compliance practice.
Once a month. Audit patterns, renewal benchmarks, vendor commercial signals across Oracle, Microsoft, SAP, Salesforce, IBM, Broadcom, AWS, Google Cloud, ServiceNow, Workday, Cisco, and the GenAI vendors. No follow up sales pressure.
Free providers (Gmail, Yahoo, Outlook) cannot subscribe. Work email only. Unsubscribe in one click.
