AI Cloud Pricing Comparison: AWS, Azure, Google and Alibaba

AI cloud pricing is comparable only when the exact model, service, region, deployment mode, context tier and workload are stated. This guide uses current first-party list prices and a transparent calculation method so that unlike prices are not presented as equivalent.

AI cloud pricing: the quick answer

There is no universally cheapest AI cloud. A low input-token rate can be offset by a higher output rate, long-context tier, reasoning-token volume, tool charges, retries or idle provisioned capacity. Models also differ in quality, latency, features and tokenization, so a price table is not a performance ranking.

For a defensible comparison, first choose models that pass the same task-level evaluation. Then price the measured input, cached input and output tokens for the exact deployment route. Compare cost per successful task, not a vendor’s lowest advertised number.

Contents

• Methodology and scope
• Current cloud price examples
• Token-cost formula and worked example
• AWS, Azure, Google and Alibaba pricing details
• Direct API reference prices
• Costs that basic token tables miss
• Hosted API vs self-hosting
• How to choose a provider
• Frequently asked questions
• Official sources

Methodology: what this AI cloud pricing comparison measures

The main comparison uses publicly listed, pay-as-you-go prices for managed model inference. Each numeric row identifies an exact model, deployment scope or region, pricing mode and context band. Cached input is shown separately because it is not interchangeable with uncached input.

Included	Excluded or separated
Standard or on-demand text-model inference	Consumer chat subscriptions
Uncached input, cached input and text output	Image, video, audio, embedding and reranking charges
Exact model and documented scope	Temporary promotions and introductory credits
USD list prices from first-party sources	Taxes, exchange rates and negotiated contracts
A declared token-mix example	Universal “cheapest” or ROI claims
Managed API charges	GPU VM estimates and self-hosting TCO

A context window is a capacity limit, not a usage charge. A model that supports a large context does not automatically bill the maximum context on every request. Conversely, some providers move an entire request into a higher price band after its input crosses a threshold.

Current AI cloud price examples

The following rows are reproducible examples from the providers’ current official tables. They are not claims that the listed models are equivalent. Prices are USD per 1 million tokens.

Cloud service	Exact model	Scope and pricing mode	Uncached input	Cached input	Output	Official source
Amazon Bedrock	openai.gpt-5.6-terra	In-region, Standard On-Demand; US East (N. Virginia and Ohio)	$2.75	$0.28 cache read; $3.44 30-minute cache write	$16.50	AWS
Amazon Bedrock	qwen.qwen3-next-80b-a3b	On-Demand; N. Virginia, Ohio and Oregon	$0.15	Not included in the cited row	$1.20	AWS
Microsoft Foundry	gpt-5.6-terra	Standard Global	$2.50	Not listed in the cited table	$15.00	Microsoft
Google Vertex AI	gemini-3.6-flash	Standard pay-as-you-go	$1.50	$0.15	$7.50	Google Cloud
Alibaba Cloud Model Studio	qwen3.7-plus	Singapore, International deployment; no more than 256K input tokens per request	$0.40	Model- and cache-mode-specific	$1.60	Alibaba Cloud
Alibaba Cloud Model Studio	qwen3.7-max	Singapore, International deployment; documented tier up to 1M input tokens	$2.50	Model- and cache-mode-specific	$7.50	Alibaba Cloud

How to calculate LLM API cost correctly

Calculate input and output independently. Add cached-input, cache-write, tool, storage or other charges only when they apply.

Total cost =
(uncached input tokens ÷ 1,000,000 × uncached input rate)
+ (cached input tokens ÷ 1,000,000 × cached input rate)
+ (output tokens ÷ 1,000,000 × output rate)
+ tool, storage, fine-tuning or provisioned-capacity charges

Worked workload: 1M input tokens and 200K output tokens

This example aggregates many standard, uncached text requests. Every individual qwen3.7-plus request remains at or below 256K input tokens. It excludes cache, tools, promotions, taxes and retries. The workload contains 1.2 million total tokens; it is not mislabeled as “1 million total tokens.”

Cloud row	Calculation	Estimated charge
AWS Bedrock — Qwen3 Next 80B A3B	(1 × $0.15) + (0.2 × $1.20)	$0.39
AWS Bedrock — GPT-5.6 Terra	(1 × $2.75) + (0.2 × $16.50)	$6.05
Microsoft Foundry — GPT-5.6 Terra	(1 × $2.50) + (0.2 × $15.00)	$5.50
Google Vertex AI — Gemini 3.6 Flash	(1 × $1.50) + (0.2 × $7.50)	$3.00
Alibaba Model Studio — Qwen3.7 Plus	(1 × $0.40) + (0.2 × $1.60)	$0.72
Alibaba Model Studio — Qwen3.7 Max	(1 × $2.50) + (0.2 × $7.50)	$4.00

Real workloads rarely have identical output ratios. Measure token usage from provider responses, separate requests by model and context band, and calculate each group before summing the monthly total.

AWS vs Azure vs Google vs Alibaba: pricing details

Amazon Bedrock pricing

Amazon Bedrock prices the model route selected in a supported region and inference mode. The official table separates Standard, Priority, Flex, Batch, cache operations and provisioned options where those modes are supported. A discount shown for one model family must not be copied to another model without an applicable row.

For the verified US East in-region example, Bedrock lists GPT-5.6 Terra at $2.75 per 1M input tokens and $16.50 per 1M output tokens. Its cache read and 30-minute cache-write prices are separate. Bedrock also lists Qwen3 Next 80B A3B at $0.15 input and $1.20 output in N. Virginia, Ohio and Oregon. Region availability and prices can differ elsewhere.

Check the current Amazon Bedrock pricing table.

Microsoft Foundry and Azure OpenAI pricing

Microsoft offers Standard on-demand token billing, provisioned throughput units and supported Batch routes. Deployment choices include Global, Data Zone and Regional. The Azure pricing page also warns that displayed prices are estimates and that actual charges can vary by agreement, purchase date and currency.

Microsoft’s current GPT-5.6 announcement lists Terra at $2.50 per 1M input tokens and $15.00 per 1M output tokens for Standard Global. The cited table does not publish a cached-input figure, so none is inferred here. Global Standard does not mean processing is restricted to the resource region; evaluate the required data-processing scope from Microsoft’s deployment documentation.

The verified Standard Global rate happens to match the direct OpenAI Standard rate for GPT-5.6 Terra on this date, but that does not make the routes interchangeable. Other deployment types, priority modes, contractual terms and currencies can differ.

Check Azure OpenAI pricing and the Azure calculator.

Google Vertex AI pricing

Google separates Standard, Priority and Flex/Batch pricing. Its current generative-AI pricing table also separates uncached and cached input. For Gemini 3.6 Flash, Standard pricing is $1.50 per 1M input tokens, $0.15 per 1M cached input tokens and $7.50 per 1M text-output tokens. Google specifies that text output includes the response and reasoning.

Location and context rules remain model-specific. Some Google rows show different Global and non-global prices, and some move the complete request to a long-context rate after the documented threshold. Read the exact model row rather than applying one Gemini price to the entire catalog.

Check the current Google Cloud generative-AI pricing table.

Alibaba Cloud Model Studio and Qwen pricing

Qwen is a model family; Alibaba Cloud Model Studio is a managed cloud service that offers Qwen models. Treating “Qwen” as a separate cloud beside Alibaba is therefore inaccurate. Model Studio pricing depends on the exact model ID, region, deployment scope, input tokens in each request and optional cache or Batch mode.

In the Singapore International deployment, the standard list price for qwen3.7-plus is $0.40 input and $1.60 output per 1M tokens when a request contains no more than 256K input tokens. A request above 256K and up to 1M uses the higher $1.20 input and $4.80 output tier for the whole request. The lower tier is not applied progressively to the first 256K tokens.

The same deployment lists qwen3.7-max at $2.50 input and $7.50 output per 1M tokens in its documented tier. Other Model Studio regions and deployment scopes can have different list prices. For the complete Qwen-by-region matrix, see the independent Qwen API pricing guide.

Check the current Alibaba Cloud Model Studio pricing table.

Direct API reference prices are not cloud marketplace quotes

Teams often compare a direct model-provider API with AWS, Azure, Google Cloud or Alibaba Cloud. That can be useful, but the route must remain visible. The following current direct or first-party managed API examples must not be copied into a different cloud’s calculator.

API route	Exact model	Scope	Input	Cached input	Output	Official source
OpenAI API	gpt-5.6-terra	Standard, short context	$2.50	$0.25	$15.00	OpenAI
Anthropic API	claude-sonnet-5	Introductory price through August 31, 2026	$2.00	$0.20 cache read	$10.00	Anthropic
Gemini Developer API	gemini-3.6-flash	Paid Standard tier	$1.50	$0.15	$7.50	Google
DeepSeek API	deepseek-v4-flash	Standard; input value is a cache miss	$0.14	$0.0028 cache hit	$0.28	DeepSeek
Mistral API	mistral-small-2603	Standard direct API	$0.15	$0.015	$0.60	Mistral AI / model catalog / cache docs
xAI API	grok-4.5	Standard, below 200K input tokens	$2.00	$0.30	$6.00	xAI

Additional constraints matter. OpenAI applies a higher long-context rate to qualifying GPT-5.6 requests above its documented threshold. Anthropic’s introductory Claude Sonnet 5 price is scheduled to become $3 input and $15 output per 1M tokens on September 1, 2026. xAI lists higher rates for Grok 4.5 requests at or above 200K input tokens. Check the linked table before using any row in a budget.

AI costs that a basic token table can miss

• Long-context tiers: a single large request can move all of its tokens to a higher rate.
• Reasoning tokens: some providers bill internal reasoning as output, increasing generated-token volume.
• Cache creation and storage: a cache write can cost more than standard input, while reads may be discounted. A cache saves money only when it receives enough valid hits.
• Batch, Flex and Priority: these modes can change price, latency and delivery guarantees. Support is model-specific.
• Built-in tools: web search, grounding, code execution or other tools can carry per-call or usage charges in addition to tokens.
• Provisioned capacity: reserved throughput can improve predictability but may create idle-capacity cost when utilization is low.
• Supporting systems: vector databases, embeddings, reranking, object storage, logging, observability, guardrails and network egress can be material.
• Retries and output control: repeated calls and unnecessarily long responses increase the bill even when the unit price is unchanged.
• Taxes and currency: list prices in USD are not necessarily the amount on an account’s final invoice.

Managed API pricing vs open-weight self-hosting

Open-weight Qwen or Llama models do not have one universal per-token price. A cloud marketplace, dedicated endpoint and self-hosted GPU cluster are different products with different cost structures. Downloadable weights may avoid a hosted per-token fee, but compute, storage, networking, engineering, monitoring, security and idle capacity still cost money.

Route	Primary billing unit	What must be measured	Best fit
Managed pay-as-you-go API	Input, cached input, output and tools	Token mix, context tier, retries and feature charges	Variable demand and low infrastructure overhead
Provisioned managed endpoint	Reserved throughput or capacity period	Utilization, peak demand, commitment and overage rules	Stable demand and predictable throughput requirements
Self-hosted open-weight model	GPU/accelerator time plus infrastructure	Measured throughput, utilization, replicas, storage, networking and operations	Teams with sufficient scale, expertise or control requirements

Self-hosted cost per 1M processed tokens =
(total infrastructure and operating cost for the period ÷
total processed tokens in the same period) × 1,000,000

A credible self-hosting estimate requires a benchmark on the exact checkpoint, quantization, accelerator, serving stack, prompt/output mix, concurrency and availability target. A generic GPU hourly price is not enough.

How to choose an AI cloud provider

Decision question	Evidence to collect
Does the model solve the task reliably?	A representative evaluation set, acceptance criteria and human review of failures
What does one successful task cost?	Actual input, cached-input, reasoning/output, tool and retry usage
Can the route meet latency and throughput targets?	P50/P95 latency, rate limits, concurrency and burst tests in the required region
Does it meet data and compliance requirements?	Deployment scope, processing location, retention, contract and service documentation
How much operational work is acceptable?	Monitoring, failover, upgrades, capacity planning, security and support requirements
What happens when demand changes?	On-demand, Batch, Flex, Priority and provisioned-capacity scenarios

Shortlist two or three routes, run the same evaluation and load test on each, and record the model version and settings. A cheaper unit rate is useful only when the route also meets the required quality, latency, availability and governance thresholds.

Why older GPT-4, Claude 2, Llama 2 and “Qwen-3 Turbo” rows were removed

A current comparison should not use legacy labels as defaults. Anthropic records Claude 2 and Claude 2.1 as retired from its API on July 21, 2025. “Qwen-3 Turbo” is not a current official model ID. Alibaba’s legacy qwen-turbo is no longer updated, and Alibaba recommends Qwen-Flash instead. GPT-4-era prices no longer represent current OpenAI or cloud-platform starting points.

Llama pricing needs a different correction: there is no single Meta-issued token price that applies to every hosted or self-hosted Llama deployment. The price belongs to the selected cloud, inference provider or infrastructure configuration. A Llama 2 rate from one provider cannot be presented as a universal “Llama price.”

Frequently asked questions

Which AI cloud is the cheapest?

There is no universal cheapest provider. The answer depends on the exact model, region, context band, input/output ratio, cache-hit rate, inference mode and required quality. Compare cost per successful task after an evaluation, not the lowest token rate in a catalog.

Is Qwen cheaper than GPT or Claude?

Some Qwen routes have lower list prices than selected GPT or Claude routes, but the models and services are not equivalent. Output length, reasoning behavior, retries, quality and regional requirements can change the effective cost. Use the same workload and acceptance test before deciding.

Is “Qwen-3 Turbo” a current official model?

No current official general-purpose model ID named Qwen-3 Turbo was found in the checked Alibaba Cloud catalog. The legacy ID qwen-turbo is no longer updated, and Alibaba recommends moving to Qwen-Flash.

Why is qwen3.8-max-preview not priced in the table?

The checked QwenCloud source documents it as a preview under Token Plan, and it was not listed in the Model Studio pay-as-you-go pricing table checked on July 26, 2026. Token Plan credits are not directly comparable with pay-as-you-go tokens, so no per-million-token price is invented here. This page also makes no claim that the preview is the default model in Model Studio, Qwen Chat, Fireworks or another service.

How much does 1 million AI tokens cost?

There is no single amount. One million input tokens and one million output tokens have different prices, and together they equal two million total tokens. State the input/output mix, multiply each volume by its own rate, then add any cache, tool, storage or capacity charges.

Why can direct API and cloud marketplace prices differ?

The routes can have different regions, processing scopes, commercial terms, inference modes, cache rules and platform services. A direct OpenAI, Anthropic or Gemini rate is not automatically an AWS Bedrock, Microsoft Foundry or Vertex AI quote.

Is self-hosting an open-weight model always cheaper?

No. Self-hosting can be economical at suitable scale and utilization, but the result depends on measured throughput, accelerator cost, replicas, idle capacity, engineering, monitoring, security and availability requirements. Benchmark the exact deployment before comparing it with a managed API.

Official pricing and lifecycle sources

• Amazon Bedrock pricing
• Azure OpenAI pricing and Microsoft Foundry GPT-5.6 pricing announcement
• Google Cloud generative-AI pricing
• Alibaba Cloud Model Studio pricing and Model Studio model catalog
• OpenAI API pricing
• Anthropic API pricing and Anthropic model deprecations
• Gemini Developer API pricing
• DeepSeek API pricing
• Mistral API pricing
• xAI API pricing
• QwenCloud Token Plan overview and QwenCloud text-model catalog
• Meta Llama cloud-deployment guide

Editorial policy: a numeric rate is published only when a first-party source identifies the model and billing unit and the comparison can state its scope. Promotional prices are separated from list prices. When a stable, reproducible public number is unavailable, the reader is directed to the official calculator instead of being shown an estimate.

---

Continue researching Qwen

Use the dedicated guides for a deeper Qwen-only comparison, implementation details and model catalog. If you are new to the model family, start with What Is Qwen AI?