GPT-4o Mini Guide: The Cheap Default and When It Is Enough

For a long stretch, GPT-4o mini was the model people reached for without thinking about it. It was cheap enough that nobody needed approval to use it and good enough that most simple tasks just worked. That kind of default is useful, and it is also how a model ends up handling jobs it was never good at.

Mini is still available, but not for much longer. OpenAI announced on 14 October 2026 that gpt-4o-mini-2024-07-18, its only snapshot, shuts down on 14 January 2027, with gpt-6-luna named as the replacement. That gives you about three months to work out which of its jobs were safe and which were luck.

What it replaced

OpenAI released GPT-4o mini on 18 July 2024 as the small sibling of GPT-4o. Its job was to retire GPT-3.5 Turbo as the cheap model, which it did in ChatGPT, and it became the obvious API replacement as well.

Microsoft's launch post for Azure gave the comparison that mattered: 82% on MMLU for mini against 70% for GPT-3.5 Turbo, at a price more than 60% lower. A cheaper model that scores twelve points higher on a broad knowledge test is not a difficult decision, and the migration happened fast.

OpenAI's current model page describes mini as a fast, affordable small model for focused tasks. It lists a 128,000-token context window, up to 16,384 output tokens, text and image input, and a knowledge cutoff of October 2023. Function calling, structured outputs and fine-tuning are all supported.

Price and the token maths

OpenAI lists mini at $0.15 per million input tokens and $0.60 per million output tokens, with cached input at $0.075. The Batch API halves that to $0.075 and $0.30. OpenRouter shows the same figures for openai/gpt-4o-mini and its :batch endpoint at the time of writing; the batch variant is the same model on an asynchronous queue, and the dated gpt-4o-mini-2024-07-18 ID is the same weights pinned to a fixed version.

The output price is four times the input price, which shapes how to use it. Mini is cheapest when it reads a lot and writes a little: labels, yes or no answers, short extractions, routing decisions. Ask it for long prose and the output side dominates the bill.

For reference, GPT-4.1 nano undercut mini at $0.10 and $0.40 with a far larger window, but nano shuts down on 23 October 2026. Mini outlives it by nearly three months.

Where a small model holds up

A model like mini is enough when three things are true: the task is narrow, the output is checkable, and a wrong answer is cheap. In practice that covers a lot of work:

• Classification and tagging against a fixed label set, especially with structured outputs enforcing the vocabulary.
• Extraction of named fields from short documents, where a schema catches malformed results.
• Routing, deciding which larger model or tool should handle a request.
• Rewriting short text into a different tone or format.

It holds up less well when the task needs several reasoning steps, when instructions conflict, or when the input is ambiguous enough that a human would ask a question first. Those failures tend to look plausible, which is the dangerous kind. When a cheap model is enough covers how to draw that line with data rather than instinct.

Mini as the first tier of a router

The most robust way to use mini was never alone. Put it first, give it a way to say it is unsure, and escalate those cases to GPT-4o or a stronger model. You pay small-model prices on the easy majority and large-model prices only where they earn it.

The tuning work is in the escalation rule. Too eager and you pay for the big model anyway; too reluctant and mini's mistakes reach users. Model routing and fallbacks covers how to set and monitor that threshold.

When choosing between mini and full GPT-4o for a given task, run both over a hundred real inputs and look at where they disagree. If disagreements are rare and the mini answers in those cases are acceptable, mini wins on cost. If they cluster in one category of input, route that category to the larger model and keep the rest on mini.

Fine-tuned mini models

Mini was the first GPT-4o-class model OpenAI opened to fine-tuning, before full GPT-4o followed in August 2024. Fine-tuned mini inference was priced at $0.30 input and $1.20 output per million tokens, double the base rate.

Fine-tuned versions follow the same 14 January 2027 shutdown as the base model. If your production system depends on a fine-tuned mini, the migration is a training job, not a configuration change. Start by checking whether a larger base model with good few-shot examples now matches your fine-tune; it often does, and it removes a training pipeline from your maintenance list.

Before January

1. List every call site using gpt-4o-mini, including judge prompts and offline scripts.
2. Group them by task type: classification, extraction, routing, generation.
3. For each group, build a small labelled set and test the replacement against it.
4. Check token usage per answer as well as accuracy, since newer models may write more.