OpenAI announcedthe release of its newest collation - sized productive model , dub GPT-4o mini , which is both less resource intensive and inexpensive to operate than its standardGPT-4o model , allow for developer to mix the AI technology into a far wider range of products .
The old mannikin is still available to developers through the API if they do n’t want to switch to 4o mini just yet . The fellowship says it will retire the sr. model eventually but has not yet mark a engagement .
GPT-4o has been available to free ChatGPT accounts since May , but there have been limitation around demand . harmonize to theupdated FAQ Sir Frederick Handley Page , GPT-4o right still has those limitation in shoes , but you ’ll now get downgraded to GPT-4o mini rather than GPT-3.5 when you shoot your terminal point . In possibility , that ’s a heavy winnings for those that have n’t upgrade toChatGPT Plus .
We ’re continuing to make advanced AI accessible to all with the launching of GPT-4o mini , now available in the API and rolling out in ChatGPT today.https://t.co/sTxtOfUapJ
& mdash ; OpenAI ( @OpenAI)July 18 , 2024
Per data fromArtificial Analysis , OpenAI ’s newest AI theoretical account scored 82 % on the MMLU reasoning bench mark , beat Gemini 1.5 Flash by 3 % and Claude 3 Haiku by 7 % . For computer address , the high MMLU benchmark to appointment was set byGemini Ultra , Google ’s top - of - the - line AI , with a score of 90 % .
What ’s more , OpenAI claims that GPT-4o mini is 60 % cheesy to operate than GPT-3.5 Turbo . developer will pay off 15 cent per million input tokens and 60 cent per million turnout souvenir . OpenAI says GPT-4o mini is “ the most able and cost - effective small role model uncommitted today , ” perCNBC .
Where do those price savings number from ? Well , not every task that can be enhance by AI postulate the full weightiness and capability of a full - sized example likeGPT , ClaudeorGemini . Like swatting flies with a sledge , utilizing a standard size LLM for unproblematic but high - loudness tasks is overkill and ravage both money and compute resources — which is where small Master of Laws such as Google ’s Gemini 1.5 Flash , Meta ’s Llama 3 8b , or Anthropic ’s Claude 3 Haiku total in . They ’re capable to perform these simple , repetitive undertaking quicker and more cost - efficiently than the larger iterations .
harmonise to OpenAI , GPT-4o mini will have the same sizing context windowpane , 128,000 item ( rough a playscript ’s Charles Frederick Worth of subject ) , as the full - size of it version with the same cognition cutoff as well , October 2023 , though the company did not limit the raw model ’s precise size . The model API presently only offers text and sight potentiality , but video and audio will be amount in the futurity as well .
The proclamation comes just a few weeks after OpenAIprovided a long - awaited updateon its anticipated , forward-looking Voice Modeas part of GPT-4o . The company ’s update indicated that a small alpha release was still to total in later July , with a across-the-board rollout being held for this declivity .
