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 routes are not presented as equivalent.
Last verified: August 23, 2026 for AWS Bedrock, Microsoft Foundry, the OpenAI direct API, Google Gemini Enterprise Agent Platform and Alibaba Cloud Model Studio. Prices are public list-price examples, not quotes. Providers can change prices without notice, so confirm the effective rate in the provider console before committing production spend.
AI cloud pricing: the quick answer
There is no universally cheapest AI cloud. A low input-token price can be offset by a higher output price, a long-context multiplier, reasoning-token volume, tool calls, retries or idle provisioned capacity. Models also differ in behavior, tokenization, latency and supported features. Treat the table below as a set of reproducible price examples, not as a performance ranking.
A defensible purchase process has two stages: first, run the same task-level evaluation on every candidate route; second, price the routes that meet the acceptance criteria using measured input, cached-input, output, reasoning and tool usage. Compare cost per successful task rather than the lowest catalog number.
Methodology and limits
This comparison includes standard or on-demand managed text-model inference from first-party cloud pricing pages. Every numeric row identifies a model ID and a deployment scope or region. It excludes consumer subscriptions, negotiated discounts, taxes, exchange rates, free credits, raw GPU instances, fine-tuning, embeddings, image and video generation, and provisioned capacity.
No live inference, latency or quality test was run for this update. We did not have a controlled multi-cloud test environment with identical prompts, versions, settings and regions. Consequently, this page makes no speed, quality or return-on-investment winner claim. The published numbers are documentation checks, not observed invoice totals.
| Included | Separated or excluded |
|---|---|
| Standard/on-demand text inference | Consumer chat subscriptions |
| Uncached input, cached input and output | Media, embedding and reranking charges |
| Exact model and documented route | Promotions, credits and negotiated contracts |
| USD list prices from first-party sources | Taxes and currency conversion |
| Declared token-mix calculations | Universal cheapest or best-model claims |
Current managed-cloud price examples
The following examples are USD per 1 million tokens. Cached input is shown separately because it is not interchangeable with standard input. “Not published in the cited row” means no figure was inferred.
| Cloud route | Exact model and scope | Input | Cached input | Output |
|---|---|---|---|---|
| Amazon Bedrock | openai.gpt-5.6-terra; in-region Standard On-Demand, US East, short context (≤272K) | $2.20 | $0.22 cache read; $2.75 30-minute cache write | $13.20 |
| Amazon Bedrock | qwen.qwen3-next-80b-a3b; On-Demand, listed US regions | $0.15 | Not published in the cited row | $1.20 |
| Microsoft Foundry | gpt-5.6-terra; Standard Global, short context | $2.00 | $0.20 cache read; $2.50 cache write | $12.00 |
| Google Gemini Enterprise Agent Platform | gemini-3.7-flash; introductory Standard pay-as-you-go through December 31, 2026 | $0.75 | $0.075 plus cache storage | $3.75 |
| Google Gemini Enterprise Agent Platform | gemini-3.7-flash; Standard pay-as-you-go from January 1, 2027 | $1.50 | $0.15 plus cache storage | $7.50 |
| Alibaba Cloud Model Studio | qwen3.7-plus; Singapore International, up to 256K input tokens per request | $0.40 | Depends on supported cache mode | $1.60 |
| Alibaba Cloud Model Studio | qwen3.7-max; Singapore International, up to 1M input tokens | $2.50 | Depends on supported cache mode | $7.50 |
Sources: Amazon Bedrock pricing, Microsoft Foundry GPT-5.6 pricing announcement, Google Gemini Enterprise Agent Platform generative AI pricing, and Alibaba Cloud Model Studio pricing.
OpenAI direct API and Microsoft Foundry are different routes
OpenAI’s direct API currently lists gpt-5.6-terra at $2.00 per 1M input tokens, $0.20 per 1M cached input tokens and $12.00 per 1M output tokens for the standard short-context tier. Prompts above 272K input tokens are billed at 2x input and 1.5x output for the full request, and cache writes are billed at 1.25x the uncached input rate.
Microsoft’s checked Standard Global short-context table now lists the same GPT-5.6 Terra token rates as the OpenAI direct API: $2.00 input, $0.20 cached input and $12.00 output per 1M tokens, with a $2.50 cache-write rate. Matching headline token rates do not make the routes interchangeable. Deployment scope, platform services, billing relationships, regional availability, contractual terms and future price changes can still differ. Keep each estimate tied to its cited route.
Token-cost formula and worked example
Total cost =
(uncached input tokens / 1,000,000 × input rate)
+ (cached input tokens / 1,000,000 × cached-input rate)
+ (output tokens / 1,000,000 × output rate)
+ tool, storage, cache-write, fine-tuning or capacity chargesFor a declared workload of 1M uncached input tokens and 200K output tokens, with no cache, tools, promotions, taxes or retries:
| Route | 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 × $2.20) + (0.2 × $13.20) | $4.84 |
| Microsoft Foundry — GPT-5.6 Terra | (1 × $2.00) + (0.2 × $12.00) | $4.40 |
| OpenAI direct — GPT-5.6 Terra | (1 × $2.00) + (0.2 × $12.00) | $4.40 |
| Google Gemini Enterprise Agent Platform — Gemini 3.7 Flash; introductory rate through December 31, 2026 | (1 × $0.75) + (0.2 × $3.75) | $1.50 |
| 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 |
The example aggregates many requests. Every individual Qwen3.7 Plus request remains at or below the 256K input threshold. A request above 256K and up to 1M uses Alibaba’s higher $1.20 input and $4.80 output tier for the whole request, so the lower-tier calculation would no longer apply.
Costs that token tables miss
- Long-context bands: crossing a threshold can reprice the entire request.
- Reasoning tokens: some providers include them in billed output.
- Cache economics: writes, reads and storage have separate rules.
- Tools: search, grounding, code execution and retrieval may add per-call charges.
- Retries and output control: failed calls and unnecessarily long answers increase total cost.
- Operations: monitoring, storage, vector search, networking, guardrails and support are outside basic token prices.
Managed API versus self-hosting
Open-weight models do not have one universal per-token price. A managed endpoint, a dedicated deployment and a model running on your own accelerators are different products. Downloadable weights may remove a hosted token fee, but compute, storage, engineering, security, monitoring, replicas and idle capacity still cost money.
Before estimating self-hosting, select an exact checkpoint in the Qwen model reference, confirm its license and files in the Qwen download guide, then benchmark the exact quantization, accelerator, serving stack, concurrency and prompt/output mix. For managed Qwen, use the region-specific Qwen API pricing guide and the implementation notes in the Qwen API guide.
How to choose a route
| Question | Evidence to collect |
|---|---|
| Does it solve the task? | A representative evaluation set, acceptance criteria and failure review |
| What does a successful task cost? | Actual uncached, cached, reasoning/output, tool and retry usage |
| Can it meet service targets? | P50/P95 latency, throughput, rate limits and regional load tests |
| Does it meet governance requirements? | Processing location, retention, contract, access and logging documentation |
| What changes with demand? | On-demand, batch, flex, priority and provisioned-capacity scenarios |
Frequently asked questions
Which AI cloud is cheapest?
There is no universal answer. It depends on the exact model, route, region, context band, input/output ratio, cache-hit rate, tool usage and the quality required. Compare cost per accepted task.
Why do OpenAI direct and Microsoft Foundry show the same Terra token rates in this table?
As of August 23, 2026, the checked short-context list rates match: $2 input, $0.20 cached input and $12 output per 1M tokens, with $2.50 cache writes. They remain separate commercial and deployment routes, so region, platform features, contracts and future prices can change independently.
Does the Gemini 3.7 Flash rate change in 2027?
Yes. On the cited Google Gemini Enterprise Agent Platform route, the introductory standard rates through December 31, 2026 are $0.75 input, $0.075 cached input plus storage, and $3.75 output per 1M tokens. From January 1, 2027, the published rates are $1.50, $0.15 plus storage, and $7.50 respectively. Recheck the official pricing page before budgeting.
Is self-hosting always cheaper?
No. It can be economical at suitable utilization and scale, but only a benchmark and a total-cost model for the exact deployment can establish that.
Official sources
- Amazon Bedrock pricing
- Microsoft Foundry GPT-5.6 pricing
- Google Gemini Enterprise Agent Platform pricing
- Alibaba Cloud Model Studio pricing
- OpenAI GPT-5.6 Terra model and pricing
Independent-site notice: qwen-ai.chat is an independent informational website. It is not Alibaba Cloud, Qwen, Amazon, Microsoft, Google or OpenAI.