Fine Tuning Cost Calculator

Fine-tuning updates the weights of an already pre-trained language model on a task-specific dataset, so the behaviour you want is baked into the parameters instead of being requested at inference time. The cost model is close to linear: every example is read once per epoch, so billable training tokens equal examples multiplied by average tokens per example multiplied by epochs. That is why a 1,000-example set at 500 tokens run for 3 epochs bills 1.5 million tokens, and why moving from 3 epochs to 5 raises the invoice by two thirds with no change to the data. Hosted rates differ sharply by base model - roughly 0.003 US dollars per 1,000 training tokens for a small model versus about 0.025 for a frontier one - and inference on the resulting custom model usually carries a 1.5x to 6x premium over the base endpoint. Practitioners treat 10 examples as a floor, 50 to 100 as the point where formatting and tone changes become visible, and 500 to 5,000 as the production range; returns flatten well before 50,000 for most narrow tasks. Two to four epochs is the standard band, with smaller sets pushed higher and sets above 5,000 examples often converging in one or two. Hold out 10 to 20 percent as validation: when validation loss turns upward while training loss keeps falling, the run has started memorising. The main limitations are overfitting on small or inconsistent data, erosion of the base model safety behaviour, no inheritance of later base-model improvements, and platform lock-in, since hosted fine-tuned weights generally cannot be exported.

The pre-train then fine-tune pattern was established for language in 2018. Jeremy Howard and Sebastian Ruder described ULMFiT in early 2018, showing that a language model trained on general text could be adapted to a classification task with far less labelled data, and Alec Radford and colleagues at OpenAI published the first GPT paper the same year with the same two-stage recipe. Google BERT, released in October 2018 by Jacob Devlin and co-authors, made task-specific fine-tuning the default workflow across NLP. Instruction tuning and reinforcement learning from human feedback, described in the InstructGPT paper by Long Ouyang and colleagues in 2022, extended the idea from narrow tasks to general assistant behaviour. Commercial fine-tuning as a metered service arrived with the OpenAI fine-tuning endpoint for GPT-3 in 2021, followed by GPT-3.5 Turbo fine-tuning in August 2023 and GPT-4 class support afterwards, each priced on training tokens plus a higher inference rate. On the open-weight side, Edward Hu and colleagues at Microsoft published LoRA, Low-Rank Adaptation of Large Language Models, in 2021, freezing the base weights and training small rank-decomposition matrices instead; it typically trains under 1 percent of parameters at comparable quality. Tim Dettmers and co-authors added QLoRA in 2023, combining 4-bit base weights with LoRA adapters so multi-billion-parameter models could be tuned on a single GPU, which is the main reason self-hosted fine-tuning became economically competitive with API training.