The top AI models are converging on quality. Their prices are not.

If you picked an AI provider in the last year or two you probably picked one of the two obvious ones, and that was the right call at the time, because the gap between the best models and everything else was wide enough that it wasn't really a decision. That gap has closed a lot faster than most people running a business have had reason to notice, and the prices have not moved with it, so a growing number of companies are now paying frontier rates for work that does not need a frontier model any more.

This piece is the numbers behind that, read on 5 August 2026 from the providers' own price lists and the main independent benchmark, plus what we think you should actually do about it. Everything here is a list price, so if you have negotiated rates your own figures will differ, but the shape of it will not.

What the quality numbers actually say

The Artificial Analysis Intelligence Index is the closest thing the industry has to a neutral scoreboard, because it aggregates a spread of benchmarks rather than letting any one lab pick the test it happens to win. It runs nine public evaluations and averages them into a single number, weighted towards agent tasks at 34%, then coding and scientific reasoning at 24% each, with long context and general knowledge making up the rest. So a score on it is not a measure of one clever trick, it is roughly how well a model does a spread of the work you would actually give it. Here is the top of the table, and the thing to look at is not who is first, it is how little separates the first four.

• Claude Opus 5, from Anthropic, scores 60.7. Closed weights.
• Claude Fable 5, from Anthropic, scores 59.9. Closed weights.
• GPT-5.6 Sol, from OpenAI, scores 58.9. Closed weights.
• Kimi K3, from Moonshot in China, scores 57.1. Open weights, and you can download it.
• Further down, and skipping several models in between, Claude Sonnet 5 scores 53.4, GLM-5.2 from Z.AI scores 51.1, and DeepSeek V4 Flash scores 49.9.

3.6 points separate the best model in the world from the fourth best, which is under 6% of the leader's score, and the fourth best is one you can download and run yourself. That gap has been closing steadily rather than suddenly, and there is no particular reason to expect it to stop, because two things are pushing it along. The first is distillation, where a large expensive model is used to train a much smaller one that answers almost as well, so the expensive part happens once and everybody downstream buys the cheap copy. The second is that several labs now publish the model itself rather than just selling access to it, and once the weights are public the price stops being any one company's decision to make.

What the price numbers say

AI is billed per token, which is roughly three quarters of a word, and it is billed separately for what you send in and what the model writes back. Output is the expensive half, so that is the column to watch. These are list prices per million tokens as at 5 August 2026.

• Claude Fable 5. $10 per million in, $50 per million out.
• GPT-5.6 Sol. $5 in, $30 out.
• Claude Opus 5. $5 in, $25 out.
• Kimi K3. $3 in, $15 out.
• Claude Sonnet 5. $3 in, $15 out from 1 September 2026, currently discounted to $2 and $10 on an introductory rate.
• GLM-5.2. $1.40 in, $4.40 out.
• DeepSeek V4 Pro. $0.44 in, $0.87 out.
• DeepSeek V4 Flash. $0.14 in, $0.28 out.

Put the two lists next to each other and the picture is odd. Fable 5 scores 59.9 and charges $50 per million output tokens. Kimi K3 scores 57.1 and charges $15. So 2.8 points of measured intelligence, which is under five percent, is the difference between paying full price and paying seventy percent less. Go further down and it gets stranger. DeepSeek V4 Flash still delivers 83% of Fable's score, and it charges 28 cents per million output tokens against Fable's $50, which is a 179 times difference on the same meter.

Why one model costs more than another

It is worth understanding this, because it tells you which way the prices are likely to go. The cost is not really a branding decision, it is a fairly direct reflection of how much computing work the hardware has to do to answer you, and two things drive that.

Only part of the model runs

Most current models are what is called a mixture of experts, which means the full model is enormous but only a slice of it is used for any given word. Kimi K3 holds 2.8 trillion parameters in total and uses about 104 billion of them per token, and the bill tracks the part that actually runs, because that is the part the hardware has to load and read. This is why a bigger model and a more expensive model stopped meaning the same thing, and it is a large part of why the cheaper models have been able to close the quality gap without the price coming with them.

Writing costs more than reading

When you send a prompt the whole thing gets read in one pass, which is fast and efficient. When the model replies it writes one token at a time, and every single one of those tokens makes the hardware read the active part of the model out of memory again. That is why output costs somewhere between two and five times input on every provider's price list, whether they are American or Chinese, and it is not a margin choice, it is what the machines do. The practical consequence for you is that a system which makes the AI write long answers is far more expensive than one which makes it read a lot and answer briefly.

So why can't you just switch?

Because when a business adopts an AI provider, the adoption does not stay in one place. It spreads into the work, and after a year or so of that there are four things holding you where you are.

1. 1.Your prompts. Whoever wrote them tuned them over months against one model's particular habits, and a prompt tuned that way does not fail loudly on a different model, it just gets quietly worse in ways nobody notices for a fortnight.
2. 2.Your tools. If you have connected AI to your CRM, your diary or your accounts, those connections were described in one provider's format, and nothing else knows how to read them.
3. 3.Your files and memory. Whatever the system has learned and stored about your business is sitting in a shape that one platform understands, and often behind that platform's login.
4. 4.Your team. They learned one product properly, and moving costs them a month of being slower at their jobs, which is a real cost that nobody puts in the business case.

None of those is technically hard to fix on its own. Together they mean switching provider stops being a decision and becomes a project, and a project needs a budget and an owner and a slot in a roadmap that is already full, so it does not happen. That is the actual trap, and it is worth being clear that it is not a trap anybody set deliberately. It is just what happens when you build on top of a product instead of building something the product plugs into.

What a control plane is

A control plane is a filing system for your AI work that no single vendor owns. That is genuinely most of the idea, and the unglamorous nature of it is probably why it does not get sold very hard. The principle is that everything which is actually yours lives in your own files in formats anything can read, and the model becomes a component you plug in at one clearly marked point, rather than the foundation the whole thing is built on.

In practice there are four parts to it, and none of them belong to a model provider.

1. 1.Instructions in plain files. Your standards, your tone of voice, your rules about what the AI may and may not do, written once in plain text files that sit in your own storage. Every major provider can now read a plain instruction file, so the same instructions work whichever model is running today.
2. 2.Memory and documents in your folders. Named, dated and readable without logging into anybody's platform to get at them. If the only copy of what your AI knows about your business is inside somebody's product, you do not really own it.
3. 3.Tools described once. One open description of what your systems can do and how to call them, written in a format the whole industry has settled on, so a new model can pick up your CRM connection without anybody rewriting it.
4. 4.One seam for the model. There should be exactly one place in the whole setup where a provider's name appears. Change it there and you have changed it everywhere.

The reason this works now, and would not have worked two years ago, is that the conventions have arrived. Every serious provider reads plain instruction files, most of them answer to the same shape of tool call, and there is now a widely adopted open protocol for connecting AI to business systems. Nobody has to invent a standard any more. The work now is deciding what belongs on your side of that seam and what does not, and getting that call wrong is what costs you the second time round.

What it is worth, on one real job

Take a job a lot of professional services firms actually run, which is producing a client report. The system pulls the figures from wherever they live, drafts the commentary that goes around them, and then something checks the whole thing before it goes out. Call it roughly forty thousand tokens going in and six thousand coming out per report, which is a fair size for that kind of work. Here is the same job priced three ways, per thousand reports, at the list prices above.

We would rather show the working than ask you to take the number on trust, so here is the routed case in full. Of the forty million input tokens per thousand reports, thirty million go to DeepSeek V4 Flash for pulling the figures along with two million tokens of output, which costs $4.76. Eight million input and three and a half million output go to Kimi K3 for the drafting, which is the expensive part at $76.50 because drafting is where the writing happens. The last two million input and half a million output go to Claude Opus 5 for the final check, at $22.50. That is $103.76 all in, against $700 for running the whole thing on the most expensive model, which is a saving of just over eighty five percent.

That is an eighty five percent reduction against the first option and seventy percent against the second, on identical output, and we would gently point out that the saving is not really the interesting part. The interesting part is that the third option is only available to a business that can move work between models without a rebuild, and the first option is what everybody else is stuck with.

The honest catch

We are not going to pretend this is free of complications, because there are two worth knowing about before you get excited.

The first is about data, and it is the objection we hear most. A lot of the models closing the gap fastest come out of Chinese labs, and plenty of business owners are not comfortable with their client data going to a Chinese company, which is a perfectly reasonable position to hold. The thing to understand is that who built the model and where it runs are two completely separate questions. When a model has open weights you can run it on Western infrastructure, or on hardware you own outright, and your data never goes near the lab that trained it. The weights are just a file. Where you choose to run that file is your decision, and it is the decision that actually governs where your data goes. It is worth being straight that these are two different bills, though, because the $3 and $15 we quoted for Kimi K3 is Moonshot's own hosted API, and hosting the weights yourself or with a Western provider costs more than that, so price the route you would actually take rather than the cheapest number on the list.

The second is licensing, and it needs reading properly rather than assuming. Open weights does not automatically mean free for any commercial use, and the terms vary model by model. Kimi K3's licence, for instance, is generous but caps free commercial use at twenty million dollars of revenue, which covers essentially every firm we work with, though it would not cover everybody. Somebody should read the licence for whichever model you land on before it goes near production, and that somebody should not be the person who is excited about the price.

What we would actually do

If you are running AI in your business today, the useful exercise is not to go shopping for a cheaper model. It is to find out what it would cost you to move, because that number tells you how exposed you are. Ask whoever built or bought your setup a simple question, which is what would have to change if we switched provider next month, and listen carefully to the answer. If it is a config change then you are in good shape and you can go and price up your options whenever you like. If it turns into a list of things that would need rewriting, then the model price is not really your problem, the architecture is.

The way we build at Pulsar is to put the filing system in first and treat the model as swappable from day one, so that a price change or a better model arriving is a config change rather than a project. It is not the exciting part of an AI implementation and nobody has ever asked us for it by name, but it is the part that decides whether the next two years of price movement works for you or against you.