For enterprise developers and Azure administrators, that means separating three decisions: which model delivers acceptable results, which deployment meets operational requirements, and which billing arrangement fits measured traffic. A cheap token is not much of a bargain if the application needs several attempts to produce usable work.
What the published token prices establish
The following examples are USD pricing snapshots, not equivalent-model benchmarks. AWS’s pricing page confirms the DeepSeek figures for the three named US regions, while Google’s table distinguishes Global pricing from non-global endpoints.
| Service and example | Deployment scope | Input per million tokens | Output per million tokens |
|---|---|---|---|
| Amazon Bedrock: DeepSeek v3.2 | Standard; N. Virginia, Ohio or Oregon | $0.62 | $1.85 |
| Azure OpenAI: GPT-4.1-mini | Global Standard in JournalArta’s comparison | Not established by the retrieved rendering | Not established by the retrieved rendering |
| Google Cloud: Gemini 3.5 Flash | Global Standard text | $1.50 | $9.00 |
| Google Cloud: Gemini 3.5 Flash-Lite | Global Standard text | $0.30 | $2.50 |
The AWS and Google entries are published model-specific prices, not provider-wide rates. Google’s Flash output meter includes both response and reasoning tokens.
The Azure gap needs careful interpretation. JournalArta reports that Microsoft’s page displayed $- placeholders rather than numerical token prices. That establishes a limitation of that page rendering—not that Azure has no obtainable rates, and certainly not that the service is free. Microsoft’s pricing explanation also says actual charges can depend on the customer agreement, purchase date and exchange rate.
Do not fill that blank with an OpenAI direct-API price. For procurement purposes, the missing figure should remain unquoted until the exact Azure model, deployment type, geography and commercial agreement are identified.
Output volume can dominate the estimate
Consider an illustrative Bedrock workload consuming 10 million input tokens and generating 2 million output tokens. At the listed DeepSeek v3.2 US Standard rates, its base token charge would be:
- Input: 10 × $0.62 = $6.20
- Output: 2 × $1.85 = $3.70
- Combined: $9.90
If output rises to 10 million tokens while input stays unchanged, the same calculation becomes $24.70. These are arithmetic illustrations, not benchmark results or complete application bills. They demonstrate why a budget needs both sides of the token ledger.
Batch discounts buy flexibility, not instant responses
AWS advertises batch inference at 50% below on-demand pricing for selected foundation models. The qualification matters: this is not a blanket reduction across every Bedrock model or interactive request.
Microsoft’s pricing description, reproduced in JournalArta’s reporting, advertises a 50% Batch API discount against Global Standard pricing, with completions returned within 24 hours for qualifying deployments.
The procurement recommendation is straightforward: split the evaluation into interactive requests and jobs that can wait. An overnight document-classification run should not automatically inherit the same deployment assumptions as a customer-facing assistant.
Nor should a spreadsheet automatically multiply every advertised saving together. AWS documents prompt caching for supported on-demand inference endpoints, but not its batch inference API. Consequently, a Bedrock estimate cannot simply stack the batch discount and prompt-cache savings on the same requests.
Bottom line: Establish eligibility and completion requirements before applying a batch discount.
Caching needs its own accounting
Caching changes the cost of reusable input. Its financial value depends on actual reuse, applicable write charges and retention—not merely on whether two prompts look similar to a developer.
AWS: inspect the usage fields
Amazon Bedrock supports implicit and explicit prompt caching, with availability varying by model and API. Cache hits are not guaranteed, and some models charge more for cache writes than for ordinary input. AWS recommends checking response usage fields to determine what was actually read from or written to cache.
For a pilot, record uncached input, cache reads, cache writes and output separately. Forecasting every repeated document as a successful cache read would turn a possible optimization into an assumed discount.
Azure: cache-write charges are model-dependent
Microsoft’s documentation distinguishes models before GPT-5.6, which have no additional cache-write charge, from GPT-5.6 and later families, where writes can incur charges alongside discounted reads. Charges for 12-hour prompt-cache writes apply to usage beginning November 1, 2026.
That makes a cache-read-only budget incomplete for affected deployments. Microsoft also documents separate cached_tokens and cache_write_tokens usage fields for supported Standard deployments. Those measurements are more useful than assuming that a longer retention setting must reduce the bill.
Google: retention can introduce storage costs
Google distinguishes automatic implicit caching from explicit caches created and referenced by the application. Its documentation says cache creation normally incurs standard input-token charges unless applicable cache-write pricing says otherwise. Explicit caching also adds storage charges based on retention; implicit caching has no separate storage charge.
For Gemini 3.5 Flash Global, the listed cached-input price is $0.15 per million tokens, compared with $1.50 for ordinary input. That discount applies to the cached portion, not the whole request or its output.
Bottom line: Evaluate cache savings after accounting for writes and storage, and review retained business information alongside the price calculation.
Provisioned capacity makes idle time a budget issue
Provisioned inference changes the purchasing unit from consumed tokens to allocated capacity. That can improve cost predictability, but only a workload measurement can establish whether the commitment is attractive.
| Provider | Capacity arrangement | Important billing boundary |
|---|---|---|
| AWS | Hourly Provisioned Throughput purchased in model units | Billing continues until deletion; committed resources cannot be deleted before their term ends |
| Azure | Deployed provisioned throughput units, with hourly billing or reservations | Deployments cannot be paused; reservation obligations are separate from deployment lifecycle |
| Google Cloud | Fixed-cost, fixed-term Provisioned Throughput subscription | Evaluate the supported model, location and subscription terms |
AWS’s documentation describes no-commitment, one-month and six-month options. Google describes its arrangement as a fixed-cost, fixed-term subscription rather than an ordinary per-token purchase.
Azure adds a particularly useful purchasing safeguard: create the deployment before buying the reservation. Microsoft warns that deployment capacity is dynamic, so a reservation could otherwise leave a buyer financially committed to capacity they cannot deploy. A reservation is a billing discount mechanism, not a guarantee that deployment capacity is available.
Deleting or shrinking an Azure deployment also does not automatically eliminate its associated reservation commitment. Model those obligations separately from hourly deployment charges.
Build a comparison the application can defend
A practical evaluation should produce one common worksheet, with a separate row for each eligible deployment:
- Define the task and acceptance criteria. Decide what constitutes usable output before comparing costs.
- Measure representative demand. Record request counts, input lengths, output lengths, peaks and quiet periods.
- Identify the purchased service precisely. Include model version, geography, deployment mode, currency and quoted rates.
- Separate billing meters. Track ordinary input, output, cache reads, cache writes and applicable storage.
- Test purchasing alternatives independently. Compare interactive, batch and provisioned scenarios only where their operating requirements fit.
- Include an exit calculation. Record deletion restrictions and remaining commitment costs.
These are procurement recommendations synthesized from the providers’ documented distinctions between token billing, caching and committed capacity—not evidence that one cloud is universally faster or cheaper.
The final comparison should be cost per accepted business result within the required operating conditions. Token prices belong in that calculation. They should not be allowed to impersonate the entire calculation.
References
- AI Inference Pricing: AWS, Azure and Google Cloud Costs - JournalArta JournalArta · 2026-11-15T15:59:10+00:00
- Context caching overview | Gemini Enterprise Agent Platform | Google Cloud Documentation docs.cloud.google.com
- Agent Platform Pricing | Google Cloud cloud.google.com