GPT-4o Halves the Price and Lands on the Free Tier
The demo was the talking; the commercially interesting part was the bill.
On 13 May 2024 OpenAI announced GPT-4o, presenting it as a model that works across audio, vision and text together in real time. The "o" stands for "omni", and the change underneath the name was architectural: one network now handles every input and every output, instead of a chain of separate models passing text to each other.
The launch was staged as a livestream of demonstrations — two instances of the model interacting and singing, real-time translation, a customer-service proof of concept. The latency figures were the substance behind them. OpenAI reported audio replies in as little as 232 milliseconds and 320 milliseconds on average, and compared that to how quickly people answer one another in conversation. The earlier Voice Mode averaged 2.8 seconds on GPT-3.5 and 5.4 seconds on GPT-4, because it ran three models in sequence: one transcribing, one reasoning, one speaking. OpenAI was direct about what that design threw away — the reasoning model never heard the audio, so tone, multiple speakers and background noise were gone before it saw anything, and it could not laugh, sing or express emotion in return. OpenAI's chief technology officer, Mira Murati, framed the launch around interaction rather than raw capability, saying the model "reasons across voice, text and vision".
For anyone who commissions software, two lines further down mattered more. In the API, GPT-4o was twice as fast as GPT-4 Turbo, half the price, and carried rate limits five times higher. And in ChatGPT, it went to the free tier.
Why the price line outweighed the demonstrations
A cheaper model is only interesting if it is not a worse one. OpenAI's claim was that GPT-4o matches GPT-4 Turbo on English text and code while improving significantly on other languages, and the published evaluation chart supports the first half: 88.7% on MMLU against 86.5% for GPT-4 Turbo. OpenAI also said the model set new high marks on multilingual, audio and vision benchmarks.
The multilingual part carries a commercial edge that is easy to miss. OpenAI published token counts for the same sentence in twenty languages under the new tokenizer. Russian needed 1.7 times fewer tokens, 39 down to 23. Arabic halved, 53 down to 26. Hindi fell by 2.9 times, Gujarati by 4.4. English gained only 1.1 times, from 27 tokens to 24. Since API bills are counted in tokens, that table reads closer to a per-language price list than to a research footnote. Romanian was not among the twenty, so nobody in Chisinau could read their own number off it.
Then there is reach. OpenAI said more than a hundred million people use ChatGPT every week, and put GPT-4o in front of free users together with GPT-4 level intelligence, answers drawn from the model and the web, data analysis and charts, photo discussion, file upload, the GPT Store and Memory. Plus subscribers got a message limit up to five times the free one; when a free user reaches the ceiling, ChatGPT switches back to GPT-3.5 so the conversation can continue.
What this means if you are building something
Work out the cost per interaction, not the monthly licence. A halved token price does not halve a bill if the cheaper, faster model invites longer conversations — and it will. Estimate tokens in and out for one realistic exchange, multiply by the traffic the page actually gets, and only then decide whether the feature belongs on every visitor's screen or behind a login. This is the calculation that changed in May 2024: an assistant on a public page stopped being a separate line item and started to fit inside a normal build budget.
The model still knows nothing about your business. GPT-4o could translate a photographed menu and explain where the dish comes from, in OpenAI's own example. It could not know your stock levels, your delivery terms, your prices or last week's policy change. Every useful assistant we have scoped spends most of its effort on getting the client's own data into the prompt reliably. Budget for that work, because it is the part nobody demonstrates on stage.
Voice was an announcement, not a delivery. At launch OpenAI released text and image inputs with text outputs only. Audio outputs were to be limited to a set of preset voices, the new Voice Mode was promised in alpha to Plus users in the following weeks, and API audio and video were to reach a small group of trusted partners first, with OpenAI citing the novel risks of the audio modalities. A proposal dated May 2024 that promised a talking website was promising something that did not yet exist.
Assume your visitors have already used it. With free access at that scale, the reference point for what a chat box should do is set by ChatGPT, not by whatever widget was on the site before. Half-working assistants now compare badly against something people use for nothing.
Our reading: this was the month an AI feature became affordable enough to discuss on ordinary projects, and the month the hard part moved from the model to the data behind it.
Sources
