Nearly nine in 10 organisations have limited or redesigned AI use cases because rising token consumption is driving up the cost of running the tools, according to a new global survey.
The State of Enterprise AI 2026 report by data platform company Cloudera is based on a September 2026 poll of 1 500 IT leaders at director level or above. Only 19% of them continuously monitor what they spend on tokens and inference, the computing work of running a trained model.
South Africa was not among the nine markets surveyed by Wakefield Research for Cloudera. Its Europe, Middle East and Africa (Emea) sample was drawn from Italy, Germany and the United Arab Emirates. But a TechCentral comparison of published prices shows how the cost of leading AI models has shifted for South African users over the past two years – and why bills can climb even where prices fall.
Almost all respondents (99%) have acted in response to higher or more volatile AI costs. Some 45% have moved a workload to another cloud, a private environment or an on-premises platform, and 24% have paused or cancelled an AI use case. Emea respondents were the most likely to say token usage and inference costs grew faster than expected (44%, against 42% in Asia-Pacific and 39% in the Americas).
The problems are not only financial: 98% have had to pull back a deployment at least once because of failures, hallucinations or unexpected behaviour.
Per-token prices tell a mixed story. The latest “Instant” model used in ChatGPT, offered to developers as chat-latest, lists at US$30 per million output tokens, double the $15 that OpenAI’s documentation gives for its ChatGPT-4o predecessor. GPT-6.1 Sol, the second-priciest of the models OpenAI lists as flagships, costs $10 per million output tokens, the same as the $10 OpenAI set in August 2024 for GPT-4o, then its flagship.
| October 2024 model | Price (input/output) | October 2026 model | Price (input/output) | Change in rand (output) |
|---|---|---|---|---|
| ChatGPT-4o | $5/$15 (R87.92/R263.77) | chat-latest | $5/$30 (R82.75/R496.50) | +88.2% |
| GPT-4o | $2.50/$10 (R43.96/R175.85) | GPT-6.1 Sol | $2/$10 (R33.10/R165.50) | –5.9% |
| o1-preview | $15/$60 (R263.77/R1 055.09) | GPT-6 Astra | $10/$50 (R165.50/R827.49) | –21.6% |
| Claude 3 Opus | $15/$75 (R263.77/R1 318.86) | Claude Fable 5.1 | $10/$50 (R165.50/R827.49) | –37.3% |
| Claude 3.5 Sonnet | $3/$15 (R52.75/R263.77) | Claude Sonnet 5.5 | $2/$10 (R33.10/R165.50) | –37.3% |
Standard API list prices per million tokens. Rand at R17.58/$ (October 2024 average) and R16.55/$ (1-23 October 2026 average), from US Federal Reserve data. ChatGPT-4o and o1-preview prices are from OpenAI’s documentation for those retired models. Gemini is left out because Google’s September 2024 price-cut announcement gave percentages only
Some 2024 models cost more than today’s top tier. OpenAI’s o1-preview reasoning model listed at 60 per million output tokens and [Claude 3 Opus](https://www.anthropic.com/news/claude-3-family) at $75. GPT-6 Astra, the priciest on OpenAI’s flagship list, costs $50, as does Claude Fable 5.1, which shares the top of Anthropic’s price list with Claude Mythos 5.1. Anthropic’s Sonnet line is 37% cheaper in rand. A firmer rand – an average of R16.55/ so far this month, against an average of R17.58/$ in October 2024 – has taken almost 6% off dollar-priced bills.
Why bills climb anyway
List prices tell only part of the story. OpenAI, Anthropic and Google all bill the “reasoning” or “thinking” tokens their models generate as output. And Anthropic says the tokeniser – the component that splits text into tokens – in Claude 4.7 and later models produces about 30% more tokens for the same text. Allowing for that, the rand saving on Sonnet narrows to about 18%.
Some prices are rising outright. Google’s price list shows its flagship, Gemini 4 Argon, doubling to $4/$20 from 1 December.
The bigger driver is volume. Gartner has warned that AI coding costs could approach or exceed developer salaries within two years as token consumption surges and vendors switch to usage-based billing.
Standard Bank told TechCentral that its cloud costs rose 37% in the six months to June. The bank listed expanding its AI capabilities among several drivers of higher software and cloud costs but did not quantify AI’s share, and its banking technology costs rose only 2% overall. – © 2026 NewsCentral Media
