<\/p>\n
When ChatGPT-5 launched, I was ready to double down on my loyalty to OpenAI. Tokens looked cheaper on paper, and, as a publisher-facing tech business, NoahWire consumes AI at scale to create, curate and distribute content. Every improvement in the models means an improvement in what we can deliver. But within weeks that optimism turned into hard questions about cost, speed and reliability.<\/p>\n
On launch day, Sam Altman, OpenAI\u2019s ceo, proudly said API usage had \u201cdoubled overnight\u201d. That puzzled me. Publishers, for example, don\u2019t suddenly double the number of calls they make to an API just because a model gets smarter \u2013 their daily requirements are tied to editorial workflows.\u00a0<\/p>\n
What I soon realised was that while usage hadn\u2019t changed, costs had. Token burn shot up. Even Altman, who I admire and often think of as the closest thing we have to a Messiah in this industry, seemed to have missed this. He believed people were simply using the system more, when in fact the model itself was consuming more tokens to do the same work.<\/p>\n
Paying more for the same job<\/strong><\/p>\n GPT-5 tokens may be cheaper per unit, but the model is verbose and \u201cover thinks\u201d even simple editorial tasks. A summarisation or headline job that GPT-4 completed neatly now consumes far more tokens. Worse, it runs slower.\u00a0<\/p>\n For publishers who integrate AI into live pipelines, where feeds, newsletters, CMS integrations and alerts demand split-second accuracy, unpredictability is as damaging as under-performance. Even \u201cover-performance\u201d can cause problems when you\u2019re building precisely tuned stacks.<\/p>\n Then there are the online costs. Models like Gemini tempt publishers with a free allowance, but once you\u2019re running tens of thousands of queries a day \u2013\u00a0the norm for live feeds \u2013\u00a0the bills explode. OpenAI\u2019s trick is subtler: GPT-5 looks cheaper per unit, but by bloating responses and tokens it drives up the true cost in production. Hidden charges like these can quietly wreck a publishing P&L.<\/p>\n Rethinking model choices<\/strong><\/p>\n This has pushed us, and the publishers we work with, to rethink how tasks are assigned across models:<\/p>\n * ChatGPT-5 \u2013 Still the best for polished, multilingual long-form drafts. Excellent for features or in-depth explainers, but too slow and costly for daily pipelines\n<\/p>\n * GPT-4 \u2013 Cheaper, faster and more predictable. Less \u201cbrilliant\u201d than GPT-5, but a reliable workhorse for summaries, headlines and bulletins\n<\/p>\n What this means for publishers<\/strong><\/p>\n ChatGPT-5 was supposed to simplify publishing operations. Instead, it has pushed publishers \u2013\u00a0and many others beyond publishing \u2013\u00a0into becoming experts in model economics, workflow design and cost optimisation. Entire weeks have been lost re-engineering systems simply to stabilise them.<\/p>\n Perhaps this is the new reality of AI publishing. The winners won\u2019t be those who pledge loyalty to one model, but those who treat AI as a toolkit, choosing the right model for each editorial job and constantly auditing cost against output.<\/p>\n For publishers, that means high-quality features may still be written with ChatGPT-5, but live newsrooms, newsletters and bulletins demand the speed and economy of simpler models. The true cost of AI publishing isn\u2019t just about tokens \u2013\u00a0it\u2019s about online calls, latency and efficiency. Those who grasp this will publish faster, cheaper and better. Those who don\u2019t will keep getting caught out.<\/p>\n\n
\n<\/li>\n<\/ul>\n