LLM API pricing comparison: OpenAI, Anthropic and Google in November 2026

Most teams pick a language model by reputation and only look at the price when the first invoice arrives. That is backwards, because the gap between the cheapest and the dearest option on a single provider's list is enormous. On Anthropic's pricing page, a million input tokens costs $0.10 on the smallest current Claude model and $10 on the largest.

This guide compares LLM API pricing across OpenAI, Anthropic and Google using each provider's own published rates, checked on 3 November 2026. It covers flagship and small models, input and output prices, and what a typical business task costs once you do the arithmetic. The aim is to help you budget for real workloads in the US, UK, EU and beyond, not to crown a winner, because the cheapest model per token is often not the cheapest model per finished task.

Quick answer

In November 2026, small models from all three providers cost between $0.10 and $0.30 per million input tokens, mid-tier models sit around $0.75 to $2, and flagships run from $2 to $10. Output tokens cost five to eight times more than input. A typical support reply costs well under one cent on a mid-tier model.

How we compiled this

Every price in this post comes from a provider's own pricing page, opened on 3 November 2026: Anthropic, OpenAI and Google. We did not use third-party price trackers, because they lag and often mix up model versions. All prices are in US dollars per million tokens, standard tier, before tax, and exclude discounts such as batch processing unless we say so.

The context for how prices move comes from published research: Epoch AI on historical price declines and Gartner on its forecast to 2030. None of the figures here are Klevere's own data. We also did not benchmark quality, so a lower price does not mean a model is equivalent. Prices change often, and some pages state scheduled changes, so check the page before you sign off a budget.

Flagship models: what the top tier costs

Flagship models are the ones you reach for when the task is hard, such as long multi-step reasoning, complex code or difficult documents. They are also the most expensive, and the three providers price their top tier very differently.

The Gemini 4 Argon price is the one to watch. A budget built on today's rate could double within a month if the scheduled change applies as written. This is a good example of why a pricing page should be re-read at the point of commitment rather than copied into a spreadsheet once.

Mid-tier models: where most business work sits

Mid-tier models handle the bulk of real business tasks: drafting replies, extracting fields from documents, summarising calls, routing requests and answering questions from a knowledge base. They are the sensible default, and the prices have converged.

Three of these share the same $2 input price. That is useful, because it means you can choose on quality, latency, data handling and integration fit rather than on price. Claude Opus 5.5 at $4 and $20 is the step up if you need more capability without paying flagship rates.

Small models: the cheapest tokens available

Small models are built for volume. They are good at classification, extraction, tagging, short answers and routing, and weaker at long reasoning chains. Their prices are low enough that cost stops being the main concern for many workloads.

Anthropic's and OpenAI's smallest models are priced identically on the face of it. Google's lite models cost more per token at the low end but offer different context and modality options. As with every comparison here, the right question is what the model does on your task, not what it costs per token in isolation.

What a typical business task costs

Per-million-token prices are hard to picture, so here are two worked examples. They use the standard-tier prices above and ignore caching and batch discounts. The token counts are our own illustrative assumptions, not measured data from any client.

The first example is a customer support reply. Assume 2,000 input tokens, covering the customer message, the policy snippets retrieved and the instructions, and 400 output tokens for the reply. Run through 10,000 of these a month, the arithmetic gives the following monthly totals.

The second example is document processing. Assume 20,000 input tokens, roughly a long contract or a bundle of invoices, and 1,000 output tokens for the structured summary, repeated 1,000 times a month. A flagship at $10 and $50 costs $250. A mid-tier model at $2 and $10 costs $50. Claude Haiku 5.5 or GPT-6 Luna costs $2.50, and Gemini 3.1 Flash-Lite costs $6.50.

Two conclusions follow. First, even a flagship is cheap in absolute terms at these volumes, so for a small business the model bill is rarely the main cost of an AI project. Second, the multiples between tiers are large, which makes routing the right task to the right model the biggest saving available. Our guide to what an AI agent costs a small business covers the other cost lines that sit around the token bill.

Why the sticker price misleads

Comparing the headline numbers is a start, but several details change what you actually pay. None of them is exotic, and each can move a bill by tens of per cent.

Tokenisers are not the same

A token is not a fixed unit of text. Anthropic's page states that Claude 4.7 and later models use a newer tokeniser that produces approximately 30% more tokens for the same text, and that the exact increase depends on the content. It also says the number of tokens the same text produces depends on the language and type of content. So two models with identical per-token prices can cost different amounts for the same document. Count tokens on your own data with each provider's tooling before you compare.

Caching and batch processing

Both discounts are large. Anthropic says its batch prices are 50% of base input and output prices for every model, and cache hits cost 0.1x base input on most models and less on the newest ones. OpenAI lists batch rates that are exactly half of standard for its flagship models, and cached input at a twentieth of the standard input price on GPT-6.1 Sol. Google lists batch prices at half of standard and a context caching rate of $0.075 against $0.75 input for Gemini 3.8 Flash.

If your workload repeats a long system prompt or a shared knowledge base, caching can cut the input side dramatically. If the work is not time-critical, such as overnight document processing, batch halves the bill. Cache writes cost more than normal input, though: Anthropic prices a five-minute write at 1.25x base input and a one-hour write at 2x, so caching only pays when prompts are reused enough to recover that premium.

Regional processing and data residency

If you need data processed in a specific region, expect a surcharge. OpenAI's page states a 10% uplift on regional processing endpoints for eligible models released on or after 5 March 2026. Anthropic states that setting a US inference region adds a 1.1x multiplier on Claude 4.6 and later models. For European and UK businesses with data-handling obligations, that uplift belongs in the budget from the start rather than as a surprise later.

Tools and long prompts

Tool use adds tokens, because tool names, descriptions and schemas are sent with each request. Server-side tools carry their own charges: Anthropic lists web search at $10 per 1,000 searches on top of token costs. Long prompts can also change the rate. Claude Haiku 5.5 moves to the higher tier above 100,000 prompt tokens, and Gemini 3.1 Pro Preview does the same above 200,000. A retrieval setup that stuffs too much context into each call can quietly push you into a more expensive bracket.

Where prices are heading

The long-run direction is down, but the picture is less simple than a single trend line. Epoch AI measured the lowest price needed to reach fixed performance thresholds on six benchmarks and found that prices fell between 9x and 900x per year depending on the milestone. One example it gives is a 40x annual fall for the price of matching GPT-4's performance on PhD-level science questions. Its authors note that the fastest drops occurred in the most recent year and that it is less clear those will persist.

That finding is about the price of a fixed capability, not the price of the newest model. The frontier does not get cheaper; yesterday's capability does.

Gartner forecasts that by 2030, running inference on a one-trillion-parameter model will cost providers over 90% less than in 2025. It also warns that falling provider costs will not be fully passed on to enterprise customers, and that agentic models need 5 to 30 times more tokens per task than a standard chatbot. In other words, unit prices can fall while your total bill rises. Budget for volume, not just rates.

A practical rule follows from this. Do not plan around a price cut you cannot see on a pricing page, and do not lock in a design that only works at today's cost. Design so you can swap models without rewriting the system, then take the savings as they arrive.

How to choose a model without overspending

Start from the task, not the provider. Write down the volume, the average input and output length, how fast the answer is needed and what a wrong answer costs. Then work down the tiers.

Build a small evaluation set from real examples, 50 to 100 is a reasonable start, and run each tier against it. The cheapest model that meets your accuracy bar wins. Quality testing matters here because a cheap model that needs a human to fix a fifth of its outputs can cost more than a dearer one that does not. For a broader view of the choice, see our guide to which AI models suit small businesses, and for turning token spend into a business case, see how to measure AI ROI.

How Klevere approaches model costs

Klevere designs AI agents around the task, so the model is a component rather than the product. Across 500+ AI agents deployed on 50+ projects in 12 industries, the pattern we see is consistent: the token bill is a small part of the total cost of ownership, and the real savings come from sensible design. That means routing simple steps to small models, keeping prompts short, caching shared context and reserving the expensive model for the few steps that need it.

We also build agents so the model can be changed. Providers revise prices, retire models and schedule increases, as the Gemini 4 Argon entry above shows. An agent tied to one model inherits every one of those changes. One that sits behind a thin routing layer can move to a better-priced option after a test run. If you are scoping a build, our custom AI agent development service starts from your task volumes and accuracy needs, and our AI ROI calculator helps you size the return before any commitment.

Cost is only one failure risk in production. Runaway loops, oversized prompts and unmonitored retries can multiply a bill overnight, which we cover in what can go wrong with AI agents in production. Set per-agent spend limits and alerts from day one.

Frequently asked questions

How much does the OpenAI API cost?

On OpenAI's pricing page, checked on 3 November 2026, GPT-6 Luna costs $0.10 input and $0.50 output per million tokens, GPT-6.1 Sol costs $2 and $10, and GPT-6 Astra costs $10 and $50. Cached input and batch processing lower those rates, with batch at half price on the listed flagship models.

How much does the Claude API cost?

Per Anthropic's pricing page, Claude Haiku 5.5 is $0.10 input and $0.50 output per million tokens, Sonnet 5.5 is $2 and $10, Opus 5.5 is $4 and $20, and Fable 5.1 is $10 and $50. Batch processing takes 50% off and cache hits cost a fraction of base input.

How much does the Gemini API cost?

Google's pricing page lists Gemini 3.1 Flash-Lite at $0.25 input and $1.50 output, Gemini 3.8 Flash at $0.75 and $3.75, and Gemini 4 Argon at $2 and $10 per million tokens. The page also lists a scheduled increase for Argon on 1 December 2026, so recheck before budgeting.

Why are output tokens more expensive than input tokens?

Output is generated one token at a time, while input can be processed in parallel, so output costs the provider more compute. In the prices above, output is five to eight times the input rate for most models. Tasks that produce long answers therefore cost far more than tasks that read a lot and answer briefly.

Will LLM API prices keep falling?

The price of a fixed level of capability has fallen sharply. Epoch AI found declines of 9x to 900x per year depending on the benchmark, though it is unclear the fastest rates will last. Gartner expects provider costs to drop by over 90% by 2030, but warns that savings will not be fully passed on and agentic workloads use far more tokens.

What is the cheapest way to run an AI agent?

Route each step to the smallest model that passes your tests, keep prompts short, cache repeated context and use batch processing for work that can wait. For most small business workloads the token bill is modest, so also weigh build, integration and monitoring costs before choosing on price alone.

If you want a clear view of what your own workloads would cost to run and where a cheaper model would do the job, book a free AI audit and we will map your tasks to the right model tier.