NexBDM Blog

LLM API Pricing in 2026: What South African Businesses Pay to Run AI, and the Increase Already Set for January

By NexBDM Team · 2026-08-30

Key takeaways

LLM API pricing read straight off Google's and OpenAI's own rate cards on 30 August 2026. The rate everyone is calling an introductory launch discount is nothing of the sort, the newest model is not the cheapest one, and the Gemini Flash tier doubles on 1 January 2027.

LLM API pricing read straight off Google's and OpenAI's own rate cards on 30 August 2026. The rate everyone is calling an introductory launch discount is nothing of the sort, the newest model is not the cheapest one, and the Gemini Flash tier doubles on 1 January 2027.

LLM API pricing is charged per million tokens of text in and out, not per user seat. Read on 30 August 2026, Google lists Gemini 3.7 Flash at $0.75 input and $3.75 output per million tokens, and OpenAI lists gpt-5-nano at $0.05 and $0.40. Batch processing halves both. The Gemini rate doubles on 1 January 2027.

That last sentence is the part worth your attention, and it is the part the coverage of this month's model launch mostly got wrong. This post works off the two published rate cards themselves, read on 30 August 2026, rather than off the announcement posts about them.

What the published rates actually are

Google released Gemini 3.7 Flash on 13 August 2026. Here is what its own pricing page lists today, alongside the models most South African businesses are actually running underneath their chatbots, their document readers and their lead qualifiers. All figures are US dollars per million tokens, standard tier.

Model	Input	Output
Gemini 3.7 Flash (to 31 Dec 2026)	$0.75	$3.75
Gemini 3.7 Flash (from 1 Jan 2027)	$1.50	$7.50
Gemini 3.6 Flash (to 31 Dec 2026)	$0.75	$3.75
Gemini 3.5 Flash	$1.50	$9.00
Gemini 2.5 Flash	$0.30	$2.50
Gemini 2.5 Flash-Lite	$0.10	$0.40
gpt-5.6-luna	$0.20	$1.20
gpt-5-mini	$0.25	$2.00
gpt-5-nano	$0.05	$0.40
gpt-4o-mini	$0.15	$0.60

Sources: Google's Gemini API pricing page and OpenAI's API pricing page, both read on 30 August 2026. We have deliberately left Anthropic out of the table rather than quote a figure we did not read off the vendor's own page on the same day.

The introductory rate is not a launch discount

Almost every write-up of the 3.7 Flash launch described $0.75 and $3.75 as introductory pricing for the new model, expiring 31 December 2026. The first half of that is accurate. The framing is not.

Look at the row below it. Gemini 3.6 Flash, released three weeks earlier, carries the identical $0.75 and $3.75, with the identical 31 December 2026 end date and the identical doubling to $1.50 and $7.50 on 1 January. The price is not attached to the new model. It is attached to the Flash tier.

That changes what the date means. A launch discount is something you lose by not adopting early. A tier reprice is something that reaches you wherever you already are, on a model you may have been running since July, and it does not care whether you upgraded. If you are building on Flash today, on either version, your input cost doubles on 1 January 2027 and your output cost doubles with it.

You have roughly four months of notice, in writing, on a public page. That is considerably more warning than most infrastructure gives you, and it is worth using.

The newest model is not the cheapest one

The second thing the table shows is that model generation and model price have come apart, at both vendors.

Gemini 2.5 Flash sits at $0.30 and $2.50. That is less than half the input cost of 3.7 Flash even at the promotional rate, and a fifth of it once January arrives. Gemini 2.5 Flash-Lite sits at $0.10 and $0.40. At OpenAI the same inversion appears: gpt-5.6-luna is the newer small model at $0.20 and $1.20, while gpt-5-nano, an older one, is $0.05 and $0.40. On input that older model is four times cheaper.

This matters because of how buying decisions actually get made. An owner asks which model is best, a vendor names the newest one, and the newest one goes into everything: the classifier that decides whether a message is a complaint, the extractor that pulls a date off an invoice, the router that picks a department. None of those jobs need a frontier model. Most of them were solved adequately two generations ago, at a fraction of the price, and they are the jobs that run thousands of times a month.

The expensive model earns its rate on the hard, low-volume work. The cheap model should be carrying the high-volume work. Getting that backwards is a common way a South African business ends up with an AI bill it cannot explain, and it is a configuration problem rather than a pricing problem.

What actually moves your bill

Per-seat software pricing trained everyone to think in headcount. Token pricing does not work that way, and the difference catches people out.

Three things drive the number. The first is how much text you send in, which includes every instruction, every example and every document you attach to the question, on every single call. The second is how much text comes back, which is priced between three and eight times higher than input at every rate in the table above. The third is how many times the whole thing runs.

The practical consequence is that a verbose system prompt is not a one-off cost. It is a cost you pay on every message, permanently. A business sending a 2,000-token instruction block with each of ten thousand customer queries a month is paying for twenty million input tokens of the same repeated text.

Where the cost actually comes out

Four mechanisms, all available on the published rate cards, none of them requiring you to accept a worse answer.

Move anything that is not a conversation to batch. Both vendors price batch processing at half the standard rate, and Google's batch tier stays half after the January increase, at $0.375 and $1.875 through December and $0.75 and $3.75 from January. Overnight report generation, bulk document classification, catching up a backlog: none of that needs an answer in two seconds. If a job can wait a few hours, it costs half.

Route by difficulty, not by reputation. Send the classification, the extraction and the routing to the cheapest model that passes your accuracy test, and reserve the expensive one for the work that genuinely fails on the cheap one. This is a configuration decision that gets made once and pays every month.

Stop re-sending the same context. Instructions, product lists and policy documents that go out unchanged on every call are the largest avoidable line in most bills. Prompt caching exists at both vendors specifically for this. The work is to identify what is genuinely static and separate it from what changes per request.

Cap the output. Output is the expensive side of every rate in the table. A model asked for a one-line answer that returns four paragraphs of preamble is charging you for the preamble. Setting an explicit length limit is a single configuration value and it lands directly on the most expensive number you pay.

None of these are clever. They are the four places the money sits, and they are the four things a proper look at your setup will find. If you want to know what your own configuration is costing you before January arrives, that is exactly the kind of question a business autopsy is built to answer.

What we are not going to do

We are not converting these figures to rands. The rate cards are published in dollars, the exchange rate moves daily, and a converted number in a blog post is wrong within a week while looking authoritative for a year. Take the dollar figures, apply whatever rate your bank gives you on the day you are budgeting, and you will have a number you can defend.

We are also not going to quote what it costs to have this sorted out. That conversation happens once we have seen your setup, because the answer depends entirely on what is in it.

Frequently Asked Questions

What is a token in LLM API pricing?

A token is roughly three-quarters of an English word, so a thousand tokens is about 750 words. Both input and output are counted, and vendors publish rates per million tokens rather than per word or per request.

Is LLM API pricing cheaper than a monthly AI subscription?

It depends entirely on volume. Consumer subscriptions are a fixed monthly amount per person. API pricing charges for what you use, which is cheaper for light or automated workloads and more expensive for constant heavy use by many people.

Why does output cost more than input?

Generating text takes considerably more computation than reading it. Across the rates published by Google and OpenAI on 30 August 2026, output runs between three and eight times the input price on the same model.

Does the January 2027 Gemini increase affect models other than 3.7 Flash?

Yes. Gemini 3.6 Flash carries the identical rate and the identical 1 January 2027 doubling on Google's published pricing page, which is why the change is better read as a tier reprice than as a launch promotion ending.

Do these prices change what an AI project should cost to run in South Africa?

The token rates are global and identical wherever you are. What differs locally is the messaging layer on top, particularly WhatsApp, which carries its own per-message fees separate from anything in this table.