Amazon Bedrock Pricing: Flex -50%, Priority +75% [2026]

Amazon Bedrock pricing stopped being a single per-token meter. As of October 26, 2026, AWS lists Standard, Flex, Priority and Reserved inference tiers on its Bedrock pricing page, plus Batch inference for selected models. The gap between the cheapest and the priciest lane is wide. Flex generally runs 50% below Standard, Batch also lands 50% below on-demand, and Priority adds a 75% premium.

That structure is unusual among the big three clouds. Microsoft’s model documentation sorts Foundry deployments into Global Standard, Data Zone Standard, Standard, Provisioned and Batch deployment types, and the pages we reviewed name no Flex or Priority lane. Google prices Gemini by the million tokens. Gemini 4 Argon costs an introductory $2 per million input tokens and $10 per million output tokens until November 30, 2026, then $4 and $20.

This comparison stays on the money. It shows what each tier costs relative to Standard, how Azure and Google structure their own price menus, which workloads belong in which lane, and how to move traffic without breaking latency targets. For the platform-level view, an earlier comparison of Bedrock, Azure AI Foundry and Vertex AI covers features and ecosystems.

A note on method. Every price and date below comes from vendor pages fetched on October 26, 2026. Where a vendor publishes a ratio but not a model-by-model dollar figure (AWS does this for its tiers), the worked examples use a normalized baseline of Standard spend. Those figures are arithmetic, not quotes from a price sheet. Where only one source backed a claim, the text says reportedly or names the source.

·

Google · Preferred Sources

Don't miss new tech stories on Google

Add Tech Insider once in the Google app and our stories appear in your news suggestions.

Add Now

Amazon Bedrock Pricing at a Glance: Five Lanes, One Meter

Bedrock is AWS’s managed model-access platform. It offers hundreds of foundation models, several inference tiers, batch inference, agents, governance features and cost-management tooling. Billing is consumption-based with no upfront commitment or minimum fee for on-demand use. The meter itself is still the token: you pay for input tokens and output tokens, and the rate depends on the model, the provider, the Region, the modality and the token type.

What changed is the number of lanes on top of that meter. A request can now run on Standard (regular on-demand pricing with consistent performance), on Flex (a discounted, lower-priority lane), on Priority (a premium queue), on Reserved (committed capacity) or through Batch (asynchronous bulk work). Any guide that describes Bedrock as on-demand only is out of date, and that is a fair test for the many older pricing explainers still ranking for this keyword.

The first-order driver of your bill is still model choice. AWS’s pricing page reportedly lists Google Gemma 3 4B at $0.06 per million input tokens and $0.12 per million output tokens. It lists OpenAI’s open-weight gpt-oss-20b at $0.0361 and $0.1545, and gpt-oss-120b at $0.0773 and $0.3090. Frontier models sit far above those numbers. Moving from a frontier model to a small open-weight model can change the bill by more than any tier does, so tier selection is a second-order optimization that you apply after the model is chosen.

AWS also keeps adding models. OpenAI’s GPT-6.1 Terra became generally available on Bedrock on October 26, 2026, and AWS describes it as priced at least 50% below GPT-6.1 Sol and GPT-5.6 Terra. Reportedly, GLM 5.3 from Z.ai arrived on October 5 and Grok 4.7 on September 28. Claude Haiku 5.5 reached Bedrock and a separate Claude Platform on AWS route on October 7. Some Anthropic models remain gated: access to Claude Mythos 5.1, Claude Mythos 5 and Claude Mythos Preview requires approval from your Anthropic account team.

One more structural point. A tier discount applies to the model’s rate, and not every model supports every tier. AWS says Batch covers selected foundation models from providers including Anthropic, Meta, Mistral AI and Amazon. Flex and Priority amounts are described as general, with the exact figure remaining model-specific. Treat the percentages in this article as the headline ratios and confirm them on the pricing page for your exact model and Region before you commit a budget.

Standard vs Flex vs Priority vs Reserved vs Batch: The Tier Table

The table below turns AWS’s published language into a single reference. The multipliers are relative to the model’s Standard rate. They are the figures AWS prints on the pricing page, and they reduce the planning problem to a handful of ratios.

Tier	Price vs Standard	How it works	Best fit	Main trade-off
Standard	1.00x	Regular on-demand inference, billed per token	Everyday interactive traffic	No discount, no priority
Flex	About 0.50x (generally 50% below)	Discounted on-demand lane for work that tolerates more latency	Evaluations, summarization, multi-step agent workflows	Higher latency or lower availability
Priority	About 1.75x (75% premium)	Premium queue, up to 25% better output tokens per second on supported models	Latency-sensitive, mission-critical apps	Highest per-token price
Reserved	Fixed monthly price per 1,000 tokens per minute	One-month or three-month terms with guaranteed token capacity	Predictable, steady load	You pay whether or not you use the capacity
Batch	About 0.50x for selected models	Prompts submitted in bulk, results retrieved from Amazon S3	Offline jobs with no user waiting	Asynchronous only, limited model list

Read the tiers against each other. Priority carries a 75% premium to Standard, while Flex and Batch sit at a 50% discount to Standard. A workload that moves from Priority to Batch therefore pays far less for the same tokens. Those are the real stakes of lane choice, and they dwarf most prompt-trimming exercises.

Reserved works differently from the other four. AWS describes it as a fixed price per 1,000 tokens per minute of capacity, billed monthly on a one-month or three-month term. You are buying throughput, and not tokens. The break-even rule is simple: divide the monthly reserved charge by what the same traffic would cost on Standard. If your steady utilization of the reserved capacity sits above that ratio, Reserved wins. If it sits below, you are paying for idle capacity.

The tiers also interact with caching. AWS’s pricing notes say that starting October 7, 2026, cache read pricing for Claude Sonnet 5.5 dropped by 50% from the prior rate. For prompts with a large, stable prefix (system instructions, retrieval context, tool schemas), that change cuts the effective input cost before any tier discount applies. The setup details for that model are covered in a guide to Claude Sonnet 5.5 API setup.

Platform Specs Compared: Bedrock vs Microsoft Foundry vs Gemini API

The next table lines up the three pricing models on eleven attributes. Cells say not confirmed where the vendor pages we reviewed did not document a feature. That is a statement about the evidence, not a claim that the feature cannot exist.

Attribute	Amazon Bedrock	Microsoft Foundry	Google Gemini API
Current product name	Amazon Bedrock	Microsoft Foundry (legacy “Azure AI Foundry” URLs still appear)	Gemini API, AI Studio, Gemini platform
Billing unit	Per token, model-specific	Typically per 1,000 tokens	Per million tokens
Standard pay-per-token	Standard tier	Standard, Global Standard, Data Zone Standard	Standard paid requests
Discounted async lane	Batch, 50% below on-demand (selected models)	Batch listed among deployment types	Batch discounts reported, vary by model
Discounted synchronous lane	Flex, generally 50% below Standard	Not confirmed	Not confirmed in official pages reviewed
Latency-priority lane	Priority, 75% premium	Not confirmed	Not confirmed in official pages reviewed
Reserved capacity	Reserved, 1 or 3 months	Provisioned (reserved capacity)	Not confirmed for the Gemini API
Newest OpenAI model	GPT-6.1 Terra, generally available Oct. 26, 2026	GPT-6.1 Terra, generally available Oct. 26, 2026	GPT-6 Astra, GPT-6.1 Sol, GPT-6 Luna in private preview for selected enterprise customers
Newest Google flagship	Not confirmed (Google’s Gemma open models are listed)	Not confirmed in reviewed docs	Gemini 4 Argon, launched Oct. 13, 2026, 1M-token limit
Price varies by location	By model, provider, Region, modality	Yes: Data Zone Standard costs about 9% more in EU and APAC than in the US	Model- and token-based
Scheduled price changes	None confirmed in reviewed material	None confirmed in reviewed material	Argon $2/$10 becomes $4/$20 on Dec. 1, 2026

Two rows deserve attention. The discounted synchronous lane row shows that Bedrock sells a cheaper-but-slower option that you can call like normal inference, while the other two platforms in our sources do not name one. And the scheduled price changes row shows a calendar risk: Gemini 4 Argon’s price doubles on December 1, 2026, so a budget built on the October rate expires in five weeks.

The practical reading: Bedrock gives the most granular menu, Foundry gives the clearest location-based price split, and Gemini gives the simplest per-token card with the most near-term movement. None of the three publishes one universal dollar price per tier, so exact comparisons need the same model family, Region, modality, input-to-output mix and latency requirement on all sides.

Where Each Bedrock Lane Earns Its Keep

Tier choice is a routing decision. You are matching each class of request to the cheapest lane that still meets its latency and reliability needs. Most production systems end up using three lanes at once, so it helps to understand what each one trades away.

Flex and Batch: the 50% lanes

Flex and Batch both sit about 50% below Standard, but they suit different jobs. Flex still behaves like a normal call. You send a request and wait for the answer, with the understanding that the answer may take longer or capacity may be tighter. AWS names evaluations, summarization and multi-step agent workflows as natural fits. Batch is a different contract entirely: you submit a pile of prompts, typically through Amazon S3, and collect results later.

The decision rule is whether a human or a downstream service is waiting. If nobody waits, use Batch whenever the model supports it. If something waits but can tolerate slack, use Flex. Agent loops deserve a warning, though. A ten-step agent that calls the model ten times multiplies any Flex delay by ten, so measure end-to-end task time before you move an agent to Flex. The trade-offs between offline and live scoring are laid out in a piece on batch versus streaming inference.

Batch has a coverage limit. AWS says selected foundation models from Anthropic, Meta, Mistral AI and Amazon qualify, and the discount is described as 50% below on-demand inference. It does not say every model qualifies, and it publishes no universal Batch dollar price because rates vary by model. Check your exact model on the pricing page before you plan a pipeline around the discount.

Priority and Reserved: paying for speed and certainty

Priority is the expensive lane. At a 75% premium, it makes sense only when the extra speed turns into revenue or avoided failure. AWS says Priority can deliver up to 25% better output tokens per second on supported models. That is a ceiling, not a promise, and up to means you should test your own prompts. If your interactive feature already feels fast, paying 1.75x for a modest speedup is hard to defend.

Reserved suits the opposite problem: variable latency under load. When your traffic is steady and you need guaranteed capacity, a one-month or three-month term gives you a fixed monthly price per 1,000 tokens per minute. Teams that run a constant stream of production calls, such as an always-on support assistant, tend to reach the break-even point. Teams with spiky traffic usually do not, because the reservation sits idle between peaks.

The practical hybrid is a reserved baseline plus Standard overflow. Size the reservation to your quiet-hour floor, not your peak. Let Standard absorb the bursts, push offline work to Batch and move tolerant interactive work to Flex. Open-weight models add another cost lever, since AWS reportedly added six of them on February 10, 2026, and a large open-weight model on Bedrock shows how far down the price curve that route can go.

Microsoft Foundry: Deployment Types Instead of Tiers

Microsoft’s AI platform is now branded Microsoft Foundry. Its documentation still carries legacy “Azure AI Foundry” and “Foundry classic” paths, so the rename is an active transition and not a finished cleanup. Search results and older tutorials will mix the names. Treat them as the same product.

Foundry organizes cost around where a model runs and how capacity is bought. The deployment types in the documentation include Global Standard, Data Zone Standard, Standard, Provisioned and Batch, and the type determines both the processing location and the billing method. Standard is pay-per-token. Provisioned is reserved capacity. The documentation does not list a deployment type called Reserved, and the pages we pulled name no Flex or Priority option. Billing for these models is generally per 1,000 tokens, and the resource and deployments themselves carry no general charge.

The most useful Azure data point is the Data Zone price. For GPT-6.1 Terra, which became generally available in Foundry on October 26, 2026, Microsoft publishes Data Zone Standard rates that differ by geography and context length.

GPT-6.1 Terra, Data Zone Standard	Input per 1M tokens	Output per 1M tokens
United States, short context	$1.10	$5.50
EU and APAC, short context	$1.20	$6.00
United States, long context	$2.20	$8.25
EU and APAC, long context	$2.40	$9.00

Two patterns fall out of those four rows. EU and APAC deployments cost about 9% more than US ones at both context lengths. And long-context requests double the input rate in the US (from $1.10 to $2.20) while raising output by 50% (from $5.50 to $8.25). If your retrieval prompts regularly cross the long-context threshold, your effective rate will sit nearer the second pair of rows than the first. Microsoft also reportedly lists a cheaper Global Standard rate of $1.00 input and $5.00 output for short context, which would make data-zone residency a roughly 10% premium over global routing.

Fine-tuning adds a second bill. Microsoft’s cost guidance says fine-tuned models carry separate charges for training, for hosting and for inference. Hosting is hourly and continues even when the model sits unused, which catches teams that fine-tune an experiment and forget to delete it. Microsoft’s catalog is also large, reportedly more than 10,000 models with around 50 new ones a month as of July 2026, so the pricing menu varies by model more than by tier.

On custom silicon, Microsoft announced Maia 300 on October 12, 2026 and says production deployments begin in 2027. It is not a product you can rent today. Microsoft’s own claims put a full 96-accelerator configuration at up to seven times the rack-scale token throughput of Maia 200 and more than 2.5 times the performance per watt. Those are vendor numbers against its own previous chip, and they should not enter a 2026 budget. For hands-on Foundry work, a guide to building Microsoft Foundry agents walks through deployment steps.

Google Gemini API: Introductory Pricing and a December Step-Up

Google’s pricing story right now is about timing. Gemini 4 Argon is available globally through the Gemini API, Google AI Studio and the Gemini platform, with a 1-million-token unified input and output limit. Google set an introductory price of $2 per million input tokens and $10 per million output tokens, with cached input tokens at 95% off, through November 30, 2026. After that the price becomes $4 and $20, which applies to the Argon model intended for public access.

Gemini model	Input per 1M tokens	Output per 1M tokens	Window
Gemini 4 Argon, introductory	$2	$10	Oct. 13 to Nov. 30, 2026
Gemini 4 Argon, standard	$4	$20	From Dec. 1, 2026
Gemini 3.1 Pro (reported)	$2.00	$12.00	Current
Gemini 3.8 Flash (reported)	$1.50	$7.50	From Jan. 1, 2027

The step-up is a straight doubling on both sides: $2 to $4 on input and $10 to $20 on output. A price tracker documenting the January 1 increase for Gemini API models shows the same pattern on the Flash line, where the reported 2027 rate is double the current one. Anyone who locks in a cost model this month should price it at the December rate, not the October one. The new model’s launch context is compared with rivals in a look at Gemini 4 Argon against GPT-6.1 Sol and Sonnet 5.5.

Billing mechanics differ from the other two clouds. According to an Apidog write-up of Google’s billing guide, new API users default to prepaying at least $5 in credits, with a maximum of $5,000. Tier 1 carries a $250 monthly cap, Tier 2 needs $100 paid plus three days and lifts the cap to $2,000, and Tier 3 needs $1,000 plus 30 days. The same write-up says Gemini API usage has been excluded from the $300 Google Cloud free-trial credit since March 2026, and that when a prepaid balance hits zero, all API keys on the billing account stop working at once. Check Google’s current guide before you rely on those limits.

Google’s tier story is less settled. The sources we reviewed do not confirm a Bedrock-style five-tier schedule for the Gemini API. One report says Google announced Flex and Priority options for its Interactions API on June 22, 2026, with Flex providing a 50% cost reduction, but that rests on a single source and we treat it as reportedly true, not confirmed. Search grounding is another line item to watch: reports cite 5,000 free requests per month across Gemini 3 and newer models, then $14 per 1,000 requests.

On the enterprise side, Google Cloud announced that OpenAI’s GPT-6 Astra, GPT-6.1 Sol and GPT-6 Luna are rolling out in Model Garden. Per Google’s announcement of OpenAI models for enterprise customers, access is a private preview for selected customers, purchases run through Google Cloud Marketplace on one invoice, and spend can draw down existing cloud commitments and Flexible Savings Plans. For finance teams that is a bigger deal than a token discount, because it puts third-party model spend inside an existing commitment.

Published Price and Speed Signals From Three Vendors

A pricing comparison needs benchmarks, and the honest finding here is thin. We could verify no independent, head-to-head latency test of Bedrock’s tiers against Foundry deployment types or Gemini endpoints. What exists are vendor claims from three sources, each measured against the vendor’s own baseline. The table separates what each source says from what anyone has checked independently.

Source	Published signal	Baseline it compares against	Independently verified?
AWS Bedrock pricing page (updated Oct. 26, 2026)	Priority: up to 25% better output tokens per second, at a 75% premium	AWS Standard tier, supported models only	Not in sources reviewed
AWS Bedrock pricing page	GPT-6.1 Terra priced at least 50% below GPT-6.1 Sol and GPT-5.6 Terra	Other OpenAI models on the same platform	Not in sources reviewed
Google announcement (Oct. 13, 2026)	Argon: 1M-token limit, cached input at 95% off	Argon’s own uncached rate	Not in sources reviewed
Microsoft Maia 300 announcement (Oct. 12, 2026)	Up to 7x rack-scale token throughput, more than 2.5x performance per watt	Maia 200, Microsoft’s previous chip	No, and not yet deployed (2027)

The cleanest takeaway is arithmetic. AWS’s Priority claim means a 75% price increase buys at most a 25% speed gain. In cost-per-output-token-per-second terms, that is a poor trade unless the speed gap is business-critical. If Priority really delivers 1.25x the throughput at 1.75x the price, the cost per unit of speed rises by 40% (1.75 divided by 1.25 is 1.4). Standard-versus-Flex runs the other direction: you give up some speed and cut the per-token price in half.

Since no neutral benchmark exists, measure your own. The harness below records time to first token and output tokens per second across repeated calls and reports medians and a 95th percentile. Fill in call_model with your SDK call for the lane you are testing, and run the same prompts against each lane at the same time of day.

import statistics
import time

PROMPTS = ["Summarize this support ticket in two sentences: ..."] * 50

def call_model(prompt):
    """Return (seconds_to_first_token, output_token_count).
    Replace the body with your streaming SDK call for the lane under test."""
    raise NotImplementedError

def run(label):
    ttft, tps = [], []
    for p in PROMPTS:
        start = time.perf_counter()
        first, tokens = call_model(p)
        total = time.perf_counter() - start
        ttft.append(first)
        tps.append(tokens / max(total - first, 0.001))
    ttft.sort()
    p95 = ttft[int(len(ttft) * 0.95) - 1]
    print(label,
          "median TTFT", round(statistics.median(ttft), 2),
          "p95 TTFT", round(p95, 2),
          "median tok/s", round(statistics.median(tps), 1))

run("standard")

Run it for at least a week across business hours and off-hours. Flex and Batch economics depend on queue conditions, which change through the day. A single noon test can flatter or punish a lane. Save the raw numbers, because you will want them when a vendor changes terms.

Five Worked Scenarios With the Arithmetic

These are modelled scenarios, not customer case studies. Dollar figures are normalized: Standard-equivalent means what the same tokens would cost on the Standard tier. The multipliers are the ones AWS publishes (Flex and Batch at 0.50x, Priority at 1.75x). The Gemini and Azure rows use the published per-million-token rates listed earlier.

Scenario	Monthly Standard-equivalent	Lane	New monthly cost	Change
1. Nightly document summarization, no user waiting	$20,000	Batch (0.50x)	$10,000	Saves $10,000
2. Offline evaluation suite for prompt releases	$3,000	Flex (0.50x)	$1,500	Saves $1,500
3. Latency-critical assistant on a trading desk	$8,000	Priority (1.75x)	$14,000	Adds $6,000
4. Mixed platform: 60% batch, 25% standard, 15% priority	$100,000	Blend	Not listed	Not listed

Scenario 4 shows why mix beats any single lane. Of a $100,000 Standard-equivalent bill, $60,000 moves to Batch, $25,000 stays on Standard, and $15,000 moves to Priority. The Batch slice gets a 50% discount to Standard, while the Priority slice pays a 75% premium to Standard. A team that bought Priority for everything would have paid far more than the blend.

Scenario 5 compares token cards across clouds for one workload: 500 million input tokens and 100 million output tokens a month. The models differ in capability, so read it as a price-card illustration and not a quality ranking.

Price card	Input cost (500M)	Output cost (100M)	Monthly total
Gemini 4 Argon, introductory ($2 / $10)	$1,000	$1,000	$2,000
Gemini 4 Argon, from Dec. 1 ($4 / $20)	$2,000	$2,000	$4,000
GPT-6.1 Terra, Azure Data Zone US short context ($1.10 / $5.50)	$550	$550	$1,100
GPT-6.1 Terra, Azure Data Zone EU/APAC short context ($1.20 / $6.00)	$600	$600	$1,200
GPT-6.1 Terra, Azure Data Zone US long context ($2.20 / $8.25)	$1,100	Not listed	Not listed

Argon’s December step-up adds $2,000 a month to this workload, which is a 100% increase. On Azure, choosing EU residency over US residency adds $100 a month, and a long-context prompt mix adds $825. If the same Terra model is also priced on Bedrock, apply the Flex or Batch multiplier to its Standard rate and compare. Because AWS does not print a single dollar figure for each tier, the multiplier is the part you can rely on.

Costs Outside the Token Meter

Token rates explain most of a Bedrock invoice, but not all of it. AWS publishes worked examples on the pricing page that show how the other line items add up. These are AWS’s own illustrations, useful for sizing the order of magnitude.

AWS worked example	Calculation	Result
Custom model, provisioned throughput, 1 model unit	$39.60 per hour x 24 hours x 31 days	$29,462.40
Provisioned throughput, 2 model units	2 x $18.40 per hour x 24 x 31	$27,379.20
Provisioned throughput, 1 model unit	$16.20 per hour x 24 x 31	$12,052.80
Fine-tuning for one month	$4 training + $1.95 storage + $49.50 for one inference hour	$55.45
Automated Reasoning checks	40,000 text units x $0.17 per 1,000	$6.80
Guardrails content filters and denied topics	3,000 units x ($0.15 + $0.15) per 1,000	$0.90 per hour

The provisioned-throughput rows carry the lesson. A single always-on model unit at $39.60 an hour costs nearly $30,000 for a 31-day month whether or not you send traffic. That is the same logic as the Reserved tier, and it is why a utilization estimate belongs in the business case before purchase. Guardrails are cheap per request but run continuously on every protected call, so a high-volume app pays for them all month.

Agents add their own meters. Bedrock AgentCore has no upfront commitments or minimum fees. On October 6, 2026, AWS moved short-term memory from a flat per-event model to consumption pricing: $1.00 per GB ingested, $0.20 per GB retrieved and $0.10 per GB-month stored. AWS reportedly lists AgentCore Runtime v1 at $0.0895 per vCPU-hour and $0.00945 per GB-hour. Teams building agents should read a practical guide such as the one on building Amazon Bedrock AI agents before they size these costs.

Visibility improved too. AWS reportedly added Bedrock product attributes to Cost Explorer, AWS Budgets and Cost Management Dashboards on October 8, 2026. They let customers break costs down by model, provider, inference type and feature at no charge, although GovCloud and the China Regions are excluded. A separate change on August 11 reportedly extended cost allocation by IAM user and role to the bedrock-mantle endpoint. Those tools let you see whether your tier strategy works. A multicloud FinOps pipeline built on FOCUS 1.4 can normalize the same data across all three clouds.

Pros and Cons of Each Pricing Model

Amazon Bedrock

• Pro: The most granular menu of the three. Five lanes let you route each request class to its cheapest acceptable option.
• Pro: Batch and Flex both cut about 50% off Standard, so offline and tolerant work gets cheap fast.
• Pro: Consumption pricing with no upfront commitment, plus an optional Reserved lane when load turns steady.
• Con: No universal dollar price per tier. You must read the page model by model and Region by Region.
• Con: Batch covers selected models only, and some Anthropic models are gated behind approval.
• Con: Priority costs 75% more for at most a 25% speed gain, a weak ratio for most apps.

Microsoft Foundry

• Pro: Published, specific rates by geography and context length, which makes residency costs easy to model.
• Pro: Models sold directly by Azure bill through the Azure subscription and carry Microsoft support, which simplifies procurement for existing Azure shops.
• Con: No named Flex or Priority lane in the pages we reviewed, so there is less room to trade latency for price.
• Con: Fine-tuned model hosting bills hourly even when idle.
• Con: Branding is mid-transition, and old and new product names mix in documentation.

Google Gemini API

• Pro: The simplest card, with one input rate and one output rate per model, and a 95% discount on cached Argon input.
• Pro: Unified Google Cloud invoicing and commitment drawdown for OpenAI models bought through Marketplace.
• Con: Argon’s price doubles on December 1, 2026, and the Flash line reportedly doubles on January 1, 2027.
• Con: Prepaid credits and spend caps on the API side can halt all keys when a balance runs out.
• Con: OpenAI models are still private preview for selected customers.

Use-Case Recommendations: Which Lane for Which Workload

The matrix below maps six common workloads to a first-choice lane. It assumes the model you need is available on the platform and lane in question, which you should confirm before building.

Workload	First choice	Why	Watch out for
Bulk document processing overnight	Bedrock Batch	About 50% below on-demand, results land in Amazon S3	Model must be on the Batch list
Prompt-regression and eval runs	Bedrock Flex	Half price, latency irrelevant	Queue delays on large suites
Customer-facing chat with a latency target	Bedrock Standard, then test Priority	Priority adds 75% for up to 25% more speed	Measure before paying the premium
Always-on production assistant with steady load	Bedrock Reserved baseline or Azure Provisioned	Guaranteed capacity, predictable bill	Idle capacity still costs money
EU-resident data processing on OpenAI models	Foundry Data Zone Standard	Published EU rates, about 9% above US	Long-context prompts raise the rate
Large-context analysis on a new frontier model	Gemini 4 Argon	1M-token limit, 95% off cached input	Price doubles Dec. 1, 2026

A few of these deserve more explanation. For overnight document work, the Batch discount compounds with caching: if every document shares the same instructions, cache reads on a model like Claude Opus 5.5 cut the repeated prefix cost before the 50% applies. For eval suites, Flex is nearly free money because nobody is waiting. For chat, the right first move is to keep Standard, measure time to first token with the harness above, and pay for Priority only on the routes where users drop off after a slow response.

For steady production load, compare Bedrock Reserved with Foundry Provisioned on the same traffic profile. Both sell committed capacity, but they bill differently: Bedrock Reserved charges a fixed monthly price per 1,000 tokens per minute on one-month or three-month terms. Foundry describes Provisioned as reserved capacity without the same published term structure in the pages we saw. Ask each vendor for a quote against your real token-per-minute peak.

For regulated data, residency changes the price. Foundry’s Data Zone Standard keeps processing inside the US, the EU or APAC, and EU and APAC rates sit about 9% above US ones. Microsoft’s October 20 sovereignty announcements add controls such as a Sovereign Control Panel (reportedly in public preview), but those do not change token rates. Budget for the residency premium, not just the controls.

Migration Guide: Moving Traffic Between Lanes and Clouds

Most teams do not migrate clouds to save on inference. They re-route traffic inside their existing platform first. This sequence works for either move and keeps risk low.

1. Inventory your calls. Tag each production call by caller, model, prompt size, output size and whether a human or service is waiting. Without this you cannot pick a lane.
2. Split into three buckets. Interactive (a person waits), tolerant (something waits but can slip) and offline (nobody waits). Offline work is the first migration candidate.
3. Check model support. Confirm on the pricing page that your model supports Batch or Flex in your Region. AWS says Batch covers selected models only.
4. Run the harness. Measure time to first token and tokens per second on the current lane for a week. Save the results as your baseline.
5. Move offline jobs first. Shift one nightly pipeline to Batch and compare its cost and completion time against the baseline. Watch for jobs that now miss a deadline.
6. Pilot Flex on tolerant traffic. Move evaluations and summarization. For agent loops, measure total task time, not per-call time.
7. Test Priority on one route. Pick the single route where latency hurts conversion, and run an A/B test. Keep it only if the business gain exceeds the 75% premium on that slice.
8. Size a reservation last. After two or three weeks of tier data, set a Reserved or Provisioned baseline at your quiet-hour floor. Leave peaks on Standard.
9. Turn on cost attribution. Use Cost Explorer, Budgets and tags to track spend by model and inference type, and review the mix monthly.

If you are moving between clouds, the application layer matters more than the price card. AWS reportedly launched a redesigned Bedrock console optimized for OpenAI- and Anthropic-compatible APIs in June 2026, which should lower the code changes needed to point existing clients at Bedrock. Keep your prompts, evaluation set and routing logic provider-neutral, and a lane or vendor switch becomes a configuration change.

Two migration traps recur. First, price cards change on a calendar: Argon on December 1 and the Flash line on January 1. Put those dates in your budget tooling. Second, a cheaper lane can hide a cost elsewhere, such as an agent that now runs three times longer and holds a compute session open. Compare total cost per completed task, not cost per token.

Verdict: Amazon Bedrock Pricing Wins on Control, Not on Simplicity

For teams that run mixed workloads, Amazon Bedrock pricing is the strongest of the three in October 2026 because it offers the most levers. A realistic blend of 60% Batch, 25% Standard and 15% Priority cuts a $100,000 Standard-equivalent bill to $81,250. No other platform in our sources lets you build that mix from named, documented lanes. If most of your traffic is offline or tolerant, the savings are larger. If most of it is latency-critical, Priority makes Bedrock the most expensive way to run it.

Microsoft Foundry wins on pricing clarity for regulated, location-bound workloads. Its published Data Zone rates ($1.10 and $5.50 in the US, $1.20 and $6.00 in the EU and APAC for GPT-6.1 Terra short context) let a finance team model residency to the cent. It loses on flexibility, since no Flex or Priority lane appears in the documentation we reviewed.

Google Gemini wins on a simple price card and on a 95% cached-input discount on Argon, but it carries the biggest near-term price risk. A $2,000 monthly workload at the October rate becomes $4,000 on December 1. Budget at $4 and $20, not $2 and $10.

Our recommendation by buyer type:

• Mixed AI platform with offline pipelines: Bedrock, with Batch for offline work, Flex for evals and Standard for the rest.
• Regulated or EU-bound workloads on OpenAI models: Foundry Data Zone Standard.
• Long-context analysis where cached prompts dominate: Gemini 4 Argon, modelled at the December rate.
• Latency-critical single feature: Test Priority on that route only, and keep everything else on Standard or cheaper.

More cloud pricing breakdowns are collected in the cloud computing coverage hub. Recheck every figure on the vendor pages before you sign, since all three pricing pages moved in October.

Frequently Asked Questions

How much does Amazon Bedrock cost?

Bedrock is consumption-priced with no upfront commitment or minimum fee. The cost depends on the model, provider, Region, modality and token type. Reported examples include $0.06 input and $0.12 output per million tokens for Gemma 3 4B, and frontier models cost far more. Check the official pricing page for your exact model.

What is the difference between Flex and Batch on Bedrock?

Both sit about 50% below Standard. Flex is a normal synchronous call with lower priority, so latency may rise. Batch is asynchronous: you submit prompts in bulk and retrieve results from Amazon S3 later. Batch covers selected models only.

Is Bedrock’s Priority tier worth a 75% premium?

Usually not for the whole app. AWS says Priority can deliver up to 25% better output tokens per second on supported models. Paying 1.75x for at most 1.25x speed raises the cost per unit of speed by 40%. Test it on the one route where latency costs you conversions.

Does Microsoft Foundry have Flex or Priority tiers?

Not in the documentation we reviewed. Foundry lists Global Standard, Data Zone Standard, Standard, Provisioned and Batch deployment types. Standard is pay-per-token and Provisioned is reserved capacity.

When does Gemini 4 Argon’s price change?

Google set an introductory price of $2 per million input tokens and $10 per million output tokens through November 30, 2026. From December 1 the price becomes $4 and $20. Cached input tokens carry a 95% discount.

Is the Reserved tier cheaper than Standard?

Only at high, steady utilization. Reserved bills a fixed monthly price per 1,000 tokens per minute on one-month or three-month terms. Divide that monthly charge by what the same traffic would cost on Standard. If your utilization stays above that ratio, Reserved wins. Otherwise you pay for idle capacity.

How do I track Bedrock spend by model?

AWS reportedly added Bedrock product attributes to Cost Explorer, AWS Budgets and Cost Management Dashboards on October 8, 2026. They break costs down by model, provider, inference type and feature at no charge, excluding GovCloud and the China Regions. Pair them with your own request tags.