AI’s Bill Is Due — Intelligence or Brand?

The model answered in two seconds. The invoice will outlive the conversation. Welcome to Cold Logic. Tonight: are you buying intelligence, or paying for the brand? Cobalt brings the models. Amber brings the bill. Neither is leaving until the numbers survive contact with a completed job.

Then let us expand the map. DeepSeek, Qwen, Kimi, G L M, and MiniMax belong at this table. So does France's Mistral. American models remain useful comparison points. But a purchasing decision that begins and ends with two American logos is a very expensive way to avoid doing an evaluation.

First, define the workload before you crown a champion. Our price example is one thousand jobs. Each consumes ten thousand uncached input tokens and two thousand billable output tokens. That means ten million input, two million output. No tools, taxes, retries, or human repairs. A price calculation. Not a performance test.

And not identical amounts of useful work. Different models tokenize text differently. Reasoning can consume billable output before the answer appears. If your agent reads the same repository twelve times, you do not get to describe that as one inexpensive question. Amber has already started sharpening a spreadsheet.

At published October thirtieth rates, G L M five point three Flash prices that example at two dollars fifty. MiniMax M three point one Flash, two dollars seventy. DeepSeek four point one Flash also reaches two dollars seventy, off peak. Those are serious candidates for evaluation. Cheap does not automatically mean incapable.

Nor does a low price prove competence. DeepSeek doubles those Flash rates during its specified weekday peak windows. Our example becomes five dollars forty. The model did not suddenly grow another brain. The clock changed the invoice. Record the provider, model version, service tier, and billing conditions.

What about the larger offerings? And please resist explaining every price difference with a flag. Our viewers want working systems, not an international sporting event with a token meter.

The larger DeepSeek comes to ten dollars and fifty six cents, off peak, in our example. The full G L M model costs twenty two dollars and eighty cents. Alibaba's Qwen Max costs thirty two dollars on the international rate card. Those are different offerings, not proof of equal quality. The mainland rate card is a separate quote.

Kimi belongs on the shortlist too. But its direct pricing table did not expose the numbers during our check. We found a reseller reference. We are not quietly relabeling that as Moonshot's own confirmed rate. A beautifully formatted number is still the wrong number if it belongs to another seller.

Now France. Mistral Large four is listed at seventy cents per million input tokens and two dollars ten per million output. Eleven dollars twenty for our example. Its model page labels it public preview. That matters when you are promising somebody a stable production dependency.

And Mistral's catalogue also lists Chinese models. The lab that builds the weights and the company serving your request can be different. Origin, hosting location, contract, and license are separate questions. A European endpoint does not turn Chinese weights into French weights. Nor does a model's birthplace tell you where your data is processed.

While we are separating things: an available interface, downloadable weights, and permission for your commercial use are different facts. Then comes memory. A mixture of experts can activate a fraction of its parameters while still needing the rest stored somewhere. Active parameter count is not a free pass around the memory budget.

Before this becomes a discount victory lap: include inexpensive American tiers. The Haiku five point five rate card from Anthropic puts our short context example at two dollars. Geography is not a pricing tier. Comparing everyone else's efficient model only with America's most expensive model would rig the argument.

Agreed. G P T six Astra and Claude Fable five point one each price our uncached example at two hundred dollars. That premium may be justified on a difficult task if it avoids failures and expensive review. We have not measured that here. Capability is a hypothesis to test, not a loyalty program.

Google's Gemini four Argon prices the example at forty dollars under the current promotion. The published December rates would make it eighty. Budgeting a year of production against an introductory price is how a presentation becomes an incident report. Your cost model needs an effective date.

So the cheapest model wins the bill and the expensive model claims it saves the engineer. We can settle this without asking either marketing department. Run the same held out tasks. Same tools. Same time limit. Same acceptance checks. Show what passed, what failed, and what the provider actually billed.

And use more than one task. Extract facts from a technical document. Repair a repository and run meaningful tests. Explain an engineering decision accurately in English and Arabic. A model that wins one can lose another. Route the work to the smallest model that clears your quality and latency requirements.

Now include the human. At an assumed thirty dollars an hour, five minutes of checking costs two dollars fifty for one task. That is an illustration, not a measured failure rate for any model. But it tells you why saving half a cent on tokens can be irrelevant if the answer requires surgery.

Then comes the comment underneath every cloud bill: just run it locally. Amber, this is normally where you stop being grumpy and start being evangelical. Can we have the version with depreciation included?

Gladly. Runpod lists an eighty gigabyte H one hundred P C I Express at two dollars eighty nine an hour. Keep it provisioned for seven hundred thirty hours and compute alone is two thousand one hundred nine dollars seventy. Before storage and other charges. And that does not prove your chosen model fits on one card.

Or that it serves enough concurrent users at an acceptable latency. You need the weights, runtime overhead, and attention cache to fit. Quantization can help, but its quality tradeoff needs testing. An idle accelerator produces no useful answers while enthusiastically maintaining its hourly relationship with your credit card.

Owned hardware changes the calculation. Purchase price less resale value, spread over its service life. Measured wall power at your actual electricity tariff. Maintenance and operator time. And opportunity cost. If the graphics card is busy answering questions, it may not be rendering the show that pays for the graphics card.

One more invoice trap: a monthly chat subscription or coding plan is not automatically an unlimited general purpose interface. Quotas, endpoint rules, and overage charges belong in the calculation. A subscription screenshot is not an application budget. Read what the plan actually includes.

Our conclusion is not that one country won artificial intelligence. Build a shortlist across labs. Pin the model and host. Measure accepted work, latency, and the complete bill. Sometimes a cheap model is enough. Sometimes the premium model earns its premium. Sometimes local control is worth paying for directly.

Divide everything you spent across all attempts by the number of accepted tasks. Include failed attempts, tools, infrastructure, and review. Then tell me the cost. Tokens per second is throughput. It is not a business outcome. Neither is a benchmark trophy taped to an invoice.

Your turn. Lowest price after it passes your tasks? A premium model that cuts rework? Local control? Or a mix routed by task? Vote in AudiencePulse at two T interactive dot com slash watch. This is Cold Logic, Episode One. Bring the receipt. Then bring the accepted result.