Claude, GPT and Gemini in comparison

Compare models of Anthropic, OpenAI and Google It's harder than it looks, and not for technical reasons: because the three suppliers don't expose prices in the same way, and a quick comparison almost always ends up comparing different things.

This guide puts the official price lists next to each other, explains where the comparison holds up and where it doesn't, and proposes a method of deciding. Verified data on the official documentation of the three suppliers: August 19, 2026 for Anthropic and OpenAI, 3 September 2026 for Google.

The price lists in comparison

Prices in dollars per million tokens. For Google the plan is indicated Standard.

For Gemini 3.1 Pro the two digits are the scales: the first applies to prompts of up to 200,000 tokens, the second beyond that threshold. The model is referred to as in preview on the official page.

Four things that are eye-catching

The top band has the same price. Anthropic's most expensive model and OpenAI's both cost $10 and $50 per million tokens. It's a coincidence that it says something about the market: on that range price is no longer a competitive lever, and the choice shifts entirely on quality, speed and behavior over real cases.

In the mid-range, prices converge. Three models of three different suppliers start at $2 input. The differences are in output —$10, $12, and $12 —and boundary conditions.

In the economic belt, the gap is huge. Between 0.20 and 1.50 dollars of input there is a factor seven. On high volumes this is the band where the choice of supplier weighs the most on the final account.

The context no longer differentiates at the top. Top models of the three suppliers declare windows in the order of one million tokens. Those who chose according to that parameter today must look elsewhere, for the reasons discussed in context window: how much it really matters.

Why comparisons you read around don't hold up

Beyond the price lists, the most serious problem is comparing the performance. There are three structural reasons why a comparison between suppliers is almost always spoiled, and does not require bad faith to occur.

• Each supplier choose which tests to quote: the ones where it goes well
• Everyone knows Optimizing your model Better than others'
• The comparison comes out when the own model has just come out
• The levels of reasoning often used are not declared
• The scaffolding around the model changes the result

The fourth point deserves a note, because it is the most underrated. All three suppliers expose a parameter that regulates how much the model processes before answering: the OpenAI documentation describes a scale that goes from no reasoning up to the maximum, Anthropic indicates a high default level on top models, Google counts the reasoning tokens in the output. Comparing two models to different levels does not measure the models: it measures the configurations.

The complete reasoning is in because vendor benchmarks are not comparable.

The Limit Dates of Knowledge

The dates are relatively close on the high end and more distant on the way down. But the practical conclusion is the same for everyone: in well-designed systems the limit date matters little, because the facts come from the documents you provide and not from the internal knowledge of the model.

Those who build an assistant who answers on procedures, contracts or company price lists do not have to worry about the cutoff. Those who rely on internal knowledge to answer on legislation or on issues that change over time have a problem that the most recent cutoff does not solve, only postpones. The reasoning is in knowledge cutoff: what it means in practice.

How to Really Choose

None of the tables on this page will tell you which supplier to choose. It tells you how much they cost and what they state. The decision is made on something else.

The last row of the left column is the one that hardly anyone measures and that most often decides the fate of a project. A model that is obviously wrong costs a correction; one that produces plausible and wrong answers costs a mistake that enters the company data and is discovered months later.

The method, in five hours

• Write the task in a sentence precise
• Collect ten real cases, at least three difficult
• Choose two or three candidates of comparable band
• Run it same identical test on all
• Have them assess who Knows the Work
• Also Measure cost and time, not just quality

The third point is where the tables on this page really serve: to choose candidates that stand on the same price range. Comparing a top model with an inexpensive one does not produce useful information, because you already know how it will end.

The complete method, with the step-by-step procedure, is in how to read AI benchmarks.

Google's service plans, explained

It is the most important structural difference between the three price lists, and the one that makes most online comparisons unusable.

Anthropic and OpenAI display a single tariff per model: that price, for that model, period. Google instead articulates the price on several service plans, which respond to different needs of priorities and processing times.

On the official page, in addition to the Standard plan, appear significantly cheaper options —in the order of half —designed for jobs that are not urgent, and a higher priority option that costs significantly more. The tables we verified for the flagship model show exactly this structure: the Standard at $2 and $12, the economic options around half, the priority one around eighty percent more.

For a company, however, this articulation is an opportunity, not just a complication. If your work is in bulk and has no urgency — overnight processing, periodic analysis, report generation. the economic plan significantly reduces spending without any effect on the result.

The second detail: reasoning tokens

Google documentation specifies that the output price includes thought tokens, i.e. the intermediate processing that the model produces before responding.

This makes the comparison delicate in an unintuitive way. A much reasoned model can consume much more in output than the visible response suggests, and on a complex task the difference between perceived and billed lengths can be substantial.

The only way to know is to measure the tokens actually consumed on real cases, instead of estimating them by the length of responses. It is the same principle described in cost per task or price per token, here made even more necessary by the price list structure.

What changes as you fall off the band

The last row has the widest internal gap between suppliers, and it is also the one covering most traffic in actual business systems. Anyone processing tens of thousands of repetitive cases each month has much to gain from seriously comparing this tier, where differences in list prices translate directly into thousands of euros a year.

However, the rule that applies to all bands remains the same: the lowest unit price does not guarantee the lowest expenditure. If one economic model requires three attempts where another requires one, the final bill may tip over.

Practical examples

SMEs that need to choose a supplier for the first AI project. The temptation is to compare the rankings and choose the winner. The useful path is opposite: define the task in a sentence, understand in which band falls and try two comparable price candidates on their ten real cases. In most cases the difference between suppliers is smaller than that between bands, and the decision shifts to cost and speed.

Company with night processing in bulk. Here a structural difference matters more than any benchmark: Google offers plans for non-urgent loads at about half the standard tariff, while the other two suppliers display a single tariff. On processing that can wait, that detail is worth more than any score.

Team processing audio. Some Gemini models charge more for audio than text, images, and videos. A comparison made on the prices of the text, in this case, leads to a wrong estimate. It is the kind of detail that is discovered by reading the exact line of the list, not a summary article.

Mistakes to Avoid

• Comparing Gemini prices without declaring the service plan.
• Ignoring the scales by prompt size where they exist.
• Forgetting reasoning tokens, included in the Gemini output.
• Choosing by context window, which on the high end no longer differentiates.
• Trusting published comparisons from one of the suppliers on the other.
• Deciding without trying on their own documents.

How to apply it in the company

The procedure that makes the incomparable comparable takes one afternoon and produces only one number.

Take a hundred real cases and execute them on each candidate with the same prompt and, as far as possible, the same level of reasoning. Read the tokens actually consumed instead of estimating them: it is the only way to neutralize in a single stroke scales, service plans, reasoning tokens and verbosity differences.

Then divide the total expenditure by the number of cases concluded correctly. That number —the cost per task completed —is comparable between suppliers who structure the lists differently, while the price per token is not. The method is in cost per task or price per token.

Finally, the criterion remains that no table contains: How They Mistaken. A model that claims not to know costs a correction; one that produces a plausible and wrong answer costs a mistake that no one will verify. It is observed only by looking at the difficult cases one by one, according to the method of how to read AI benchmarks.

If you want to figure out what solution makes sense for your company and bring it into production, Giallo Studio can help you turn the problem into a real workflow.

To complete the reasoning. The method for building the test bench is in when an economic model is enough. The calculation of cost per completed task, which is the number on which you really decide, is in Mixed model strategy: light and powerful. To understand when it is advisable to question the configuration, the criterion is in when it is best to change the AI model.

The three questions to ask before choosing the supplier

Before comparing models it is better to answer three questions about your system. They change the result more than any benchmark.

What parts of the cargo really need to be answered immediately? It is a half-hour exercise that almost always reveals a substantial share of work without urgency. On that share the economic plans, where the supplier offers them, significantly reduce the expenditure without any effect on the result.

How long are typical prompts? Some models apply echelons by prompt size: after a threshold is exceeded, the rate changes. If your system sends entire documents, you may be in the most expensive echelon without knowing it — and the first step is not to change providers but to improve context selection.

Does the system process audio? On some models the audio rate is twice that of text. A comparison made on the wrong line of the list leads to a wrong estimate.

Using different providers in the same system

It is technically possible and sometimes sensible, but adds complexity that needs to be honestly assessed.

The right column is justified when a specific task is clearly best served by a supplier —for example, a block processing that benefits from the economic plans, alongside an interactive flow that is better off elsewhere.

It does not justify itself as a default choice. In most cases, staying on a single supplier reduces management work more than costs in failure to optimize, especially for an SME that does not have a dedicated team to follow it.

Conclusion

On the official price lists the picture is clear: in the Anthropic and OpenAI flagships line up on the same price, in the intermediate range the three suppliers converge around 2 dollars of input, in the economic range the gap opens up to a factor of seven.

But prices are the easy part of the comparison. The part that decides is how each model behaves on your documents, and that is nowhere to be read: it is measured in half a day out of ten real cases.

If you want to figure out what solution makes sense for your company and bring it into production, Giallo Studio designs AI agents and automations tailored to SMEs.