Cheaper AI Tokens. A Bigger Bill.

Three dollars looks cheaper than four. Fifteen looks cheaper than twenty.

Put those numbers in an AI pricing table and the buying decision seems easy.

Then count how many tokens each model actually uses.

Kimi K3 has lower published API token rates than GPT-5.6 Sol. In Artificial Analysis's Intelligence Index v4.2, however, Kimi K3 at maximum reasoning effort has a higher reported cost per task: $1.58 against $1.25 for Sol at maximum effort. That is about 26% more, despite token rates that are 25% lower.

I wanted a concrete follow-up to my post on the real cost of an AI pilot. This is a useful example because you can check the numbers yourself.

Kimi K3 vs GPT-5.6 Sol: the official token prices

These are direct-provider API prices in US dollars per million tokens, checked on 7 September 2026. They are not consumer subscription prices or a reseller's offer.

Sources: Moonshot's Kimi K3 announcement, Availability section, OpenAI's Sol model documentation and OpenAI API pricing.

The Sol figures are for standard processing with up to 272,000 input tokens. OpenAI lists higher rates above that threshold and separate cache-write charges. Its Sol pricing is promotional, available at least through 21 November 2026. Batch, fast processing, regional charges and taxes can also change a bill; they are outside this headline comparison.

For each of the three categories in the table, Kimi's listed rate is 25% lower. If the billed quantities were identical and no other charges applied, Kimi would cost less.

The quantities are where the comparison gets interesting.

What the independent benchmark measured

Artificial Analysis publishes its own measurements for both models. This is an independent evaluator's evidence, not an OpenAI or Moonshot claim about the competitor.

The figures below were checked on 7 September 2026, on pages identifying the suite as Intelligence Index v4.2. Both entries use maximum reasoning effort.

Sources: Artificial Analysis: Kimi K3, Artificial Analysis: GPT-5.6 Sol and its benchmark methodology.

Using those displayed figures, Kimi produced about 84% more output tokens across the suite. Its reported weighted cost per task was 26.4% higher: $1.58 divided by $1.25, minus one. These percentages are calculated from rounded public figures.

That cost measure includes input, cache reads, cache writes, reasoning and answer tokens, with each evaluation weighted according to the index. It is not obtained by multiplying the output-token total by a price and dividing by an assumed task count. Artificial Analysis also uses live measurements of typical cache-hit rates in its cost reporting. Treat this as the evaluator's published cost estimate, not a copy of a customer's invoice.

This is evidence that the cheaper token rate can produce the more expensive benchmark run. It does not establish that Kimi costs more for every individual task, or that the two models deliver identical quality. Similar aggregate scores do not settle whether either model can do your particular job.

Why a short answer can still use a lot of tokens

The paragraph you read at the end is only part of what you may be paying for.

A reasoning model can generate tokens before producing that answer. OpenAI's reasoning documentation explicitly says that reasoning tokens are billed as output tokens. Its API usage details report them within output usage. Counting them again on top of the output total would overstate the bill.

Moonshot's K3 documentation says the model always reasons, with low, high and max settings; max is the default. Sol's documented default is medium. The benchmark comparison above uses max for both, but an application that leaves settings unspecified may be comparing different operating choices.

Tool loops add another variable. The agent can search, read files, call a tool, inspect the result and try again. More context can come back as input on subsequent requests. A retry still consumes resources even if you throw its answer away.

I pay for all the tokens it takes to finish the job.

A task-level calculation you can check

Suppose the task is to extract a set of fields from a document and return source references. For illustration, assume both models receive 10,000 uncached input tokens in one request and both outputs pass the same checks.

The usage counts below are hypothetical. They explain the arithmetic; they are not measured Kimi or Sol results. Output means all billed output, including reasoning where applicable. Assume no cache writes, tool fees or other charges.

The model with the lower rate ends up 37.5% more expensive in this example. Change the usage counts and the result can reverse.

For the output component alone, the break-even ratio is 20 / 15 = 1.333…. Kimi can generate about one-third more output tokens before its output cost equals Sol's. Beyond that, its output-rate advantage is gone. Input usage, caching and other charges determine where the whole task breaks even.

What I would measure before choosing your model

I would give each candidate the same authorised task set and acceptance criteria, then record the exact model, reasoning setting, provider and prices used. Each provider's own billing counts matter: tokenisers do not necessarily count identical text the same way.

For every completed task, keep the uncached input, cache reads and writes, total output, tool charges, retries and review time. Across the test, include the cost of failures when calculating cost per accepted result.

Then ask whether a different reasoning setting or a simpler model preserves the quality you need. Maximum effort is a configuration choice. It needs to earn its cost.

The data boundary still comes first. A cheaper result does not authorise sending a restricted document to an external provider. Any comparison has to use routes you have actually permitted.

I scope and build useful AI assistance and automation through CPLT. If you are choosing models and the proposal only shows a price per million tokens, bring the task to a scoping call. Describe the work and its constraints; keep confidential documents out of the first message.

I want to know what an accepted result costs before recommending what you should buy.