LLM API Pricing: OpenAI, Azure, AWS, Google Compare

The cheapest token price often does not lead to the lowest monthly bill. From what I see in this comparison, your total cost usually comes down to four things: prompt size, output volume, cache use, and extra cloud charges.

If I were sizing this up for a UK business today, I’d keep these points front and centre:

• OpenAI direct is simple to price at first glance, but long prompts, cache miss rates, file storage and fine-tuned usage can push spend up.
• Azure OpenAI Service adds Azure quotas, networking, monitoring and private access costs on top of token charges.
• Amazon Bedrock can work well for steady traffic, but provisioned throughput means you may pay by the hour even when demand dips.
• Google Vertex AI can look strong for large-context work, especially where the same content is reused through context caching, though logging, storage and egress still count.

A few figures stand out:

• OpenAI account limits can range from 2,000,000 TPM at one tier to 40,000,000 TPM at a higher tier.
• Vertex AI supports context windows of up to about 2 million tokens on some models.
• OpenAI file search storage is listed at $0.10 per GB per day after the first 1 GB.
• In the sample scenarios, estimated monthly costs run from roughly £250 to £3,000, depending on workload shape.

What matters most is simple: measure cost per task, not just cost per million tokens. A support bot, an internal copilot and long-document analysis can each land in very different price bands, even when the list pricing first looks close.

Quick Comparison

So if you’re comparing providers, I’d avoid choosing on headline rates alone. I’d price the actual workflow in £, include retries, logging, networking and reused context, and then compare the monthly total.

::: @figure {LLM API Cost Comparison: OpenAI vs Azure vs AWS vs Google (2024)} :::

Stop Overpaying for LLMs: One API, Every Provider, Lowest Cost

2. How OpenAI, Azure, AWS and Google charge for LLM usage

All four providers bill by tokens. But there’s a catch: larger context windows can push up input costs before the model gives you a single reply. So while token billing looks simple on the surface, the bill you get each month often depends on a few extra moving parts.

The main gaps in cost usually come down to context size, cache use, quota limits and add-on services.

OpenAI direct API: token prices, cache discounts and fine-tuned inference

OpenAI bills input and output tokens separately. That means you pay once for what you send in, and again for what the model sends back.

Prompt caching can cut the cost of repeated prompt prefixes. This matters if your workflow keeps reusing the same system instructions or shared setup text. In that case, caching can trim spend without changing how the app works.

Fine-tuned models also have their own inference pricing. So before you ship anything to production, it’s worth checking the rate for the exact model you plan to use rather than relying on the base model price.

Azure OpenAI Service: billing, quotas and enterprise overhead

Azure OpenAI Service keeps token billing, but layers it inside Azure’s deployment, governance and quota controls.

That can change the picture quite a bit. Regional deployment choices, rate-limit tiers and quota management can affect how much capacity you get for the same spend. On top of that, networking and monitoring charges sit outside the headline model price, so the listed token rate is only part of the story.

Amazon Bedrock and Google Vertex AI: on-demand tokens, provisioned capacity and add-on charges

Amazon Bedrock and Google Vertex AI both support on-demand token pricing. They also offer provisioned capacity for steadier workloads, which can make sense when usage is predictable.

The extra charges matter here too. Logging, networking and data-processing fees can make a clear difference to the monthly total, especially once usage starts to scale. The next section shows which of these costs tend to grow fastest in production.

3. Cost drivers that change your final bill

List prices are only the starting point. In production, spend climbs when prompts get longer, throughput slows down, or platform fees start piling up. The biggest swings usually come from long prompts, throughput limits and reused context.

Context windows, long prompts and repeated retrieval costs

Long prompts, chat history and retrieved documents can push input costs up fast. Every token you send - system instructions, conversation history and RAG documents - is billed on every turn of a multi-step conversation.

That means the same content can be charged again and again. A long policy file, a big product catalogue, or a stuffed chat history might not look expensive at first glance, but over many turns it starts to bite.

Some very long-context OpenAI sessions move into a higher price band once prompts cross the threshold. So one oversized request can reprice the whole session.

Gemini on Vertex AI handles this in a different way. Context caching lets you pay the full input rate once, then reuse that content at roughly 10–25% of the standard input cost, plus a time-based storage fee.[2][6] If your workload keeps querying the same large policy document or product catalogue, that can make a noticeable difference to monthly spend.

If prompt length is under control, throughput becomes the next bottleneck.

Rate limits, throughput models and scaling constraints

Rate limits can force batching, queues or multi-region deployment, and each fix brings extra infrastructure cost.

OpenAI uses tiered account limits measured in tokens per minute (TPM) and requests per minute (RPM). At Tier 3, limits sit at 5,000 RPM and 2,000,000 TPM; at Tier 5, those rise to 15,000 RPM and 40,000,000 TPM.[9] Azure OpenAI wraps similar controls inside Azure's quota system, where capacity is allocated per resource and subscription.

Amazon Bedrock takes a different route. Its provisioned throughput model bills per model unit, hourly, with optional one- or six-month commitments. In plain English, you are paying for reserved capacity whether you use it or not.

Vertex AI ties throughput to project and region quotas. For high-volume UK workloads, that can mean multi-region deployment just to stay within limits, which then adds cross-region networking costs. One cost problem solved, another one appears.

Once traffic settles down, the next place to save money is prompt reuse instead of resending the same content each time.

Caching, fine-tuning and hidden platform charges

Caching only pays off when large prompts stay stable and get reused a lot. OpenAI's prompt caching can cut the cost of a repeated 10,000-token system prompt by a fair margin - but only if the prefix is reused consistently. Cache writes on some newer OpenAI models are priced at around $0.25–$1.00 per million cached tokens, compared with higher uncached input rates.[9][11][12][13] If your prompt structure changes between calls, cache hit rates fall and the saving disappears.

Fine-tuning can improve unit economics when it lets you run a smaller, cheaper model for a narrow task. But it also brings training, hosting, versioning and drift-monitoring costs. On Amazon Bedrock, fine-tuned models inherit provisioned throughput pricing, so a customised model may need dedicated hourly capacity no matter how often it is called.[3][4][5]

Then there are the charges that sit outside tokens and compute. These often show up later, when a production bill lands and looks higher than expected.

OpenAI charges $0.10 per GB per day for file search storage after the first 1 GB free, and the same rate for file and image upload storage.[1][7] Azure Private Link comes with separate endpoint and data-processing fees, so private networking is not free at scale.[8] Across all four providers, log ingestion and retention fees, VPC endpoints and cross-region data transfer can add meaningful cost as usage grows.

4. Provider fit for common UK workloads: budget scenarios

These examples show a simple truth: the same model can end up in very different cost bands depending on prompt length, output volume, and how often you reuse prompts. The scenarios below use public list prices to show how the shape of a workload changes spend.

Customer support, internal copilot and document analysis: cost scenarios

Customer support assistant - output-heavy. For 10,000 requests a day, with 500 input tokens and 300 output tokens per request, this setup is driven mostly by output volume. That also makes it sensitive to retries and caching. GPT-4.1 and Gemini 1.5 Pro sit at different public list prices, and if repeated system prompts can be cached, the input-side bill drops further. If traffic is steady and prompts repeat, caching helps. But when traffic is consistent, throughput can matter more.

Internal engineering copilot - concurrency-heavy. Using 2,000 active users, 5 requests per user per day, 1,500 input tokens and 400 output tokens per request, Amazon Bedrock's Claude Sonnet lands at a higher monthly spend on list price. Provisioned throughput is worth a look when usage is predictable. Azure OpenAI fits steady, governance-heavy workloads where quota management and enterprise controls matter more than raw token price. If prompts are long rather than frequent, input size and prompt reuse become the main cost levers.

Long-document analysis - input-dominated, lower volume. Using 500 documents per day, 20,000 input tokens and 1,000 output tokens per document, input cost does most of the damage. At public list prices, Google Vertex AI is the lowest-cost option for this profile, while very large prompts still need careful chunking and controls around reprocessing.

Ranges are illustrative. Actual spend varies with model choice, cache hit rate, retries and platform overhead.

How to build a cost model before procurement

To compare providers properly, build the model from your own workload, not the list price. Start with the things that shape the bill in practice:

• tokens per request
• requests per day
• context size
• cache hit rate
• retries
• monitoring costs
• networking costs

Those inputs give you a monthly estimate that is much closer to reality than list prices on their own.[17][18]

It also helps to compare providers on cost per 1,000 tasks, not just per-million-token rates. A provider may look cheaper on paper per million tokens, then end up costing more overall if its throughput model needs provisioned capacity or if your workload leans on long prompts or large context windows.

And this is where many teams get caught out: pilot prompts often look neat and tidy, while production prompts are messier, longer, and less predictable. Test a sample of real production prompts before signing anything, because that difference can change the bill fast.[17][18]

5. Conclusion: what UK teams should check before choosing an LLM API

The lowest token price almost never leads to the lowest production cost once you factor in throughput, support and platform overhead. The right way to compare providers is to run the same workload across each one, rather than getting pulled in by headline pricing alone.

Input-heavy work, like document analysis, tends to be shaped by prompt length. Output-heavy work, like support tools and copilots, is more affected by response volume and repeated context. That difference matters. Two providers can look close on paper, then drift apart fast once your actual usage pattern kicks in.

Support tiers and enterprise commitments can also change total spend in a big way at scale. A price that looks fine at small volume can become far less attractive once support plans, platform fees and delivery limits enter the picture.

Hokstad Consulting can help teams build provider-neutral cost models and cut AI, cloud and DevOps spend.

Key criteria for decision-makers

Before procurement, check these five points.

Pick the provider that fits your workload, governance needs and day-to-day operating limits best.

FAQs