o4-mini Guide: OpenAI's Cheap Vision Reasoner and Its Successor

For about a year, o4-mini was the default answer to a common question: which OpenAI reasoning model should I use when I do not want to pay for o3? It was cheap, it read images, it called tools, and on many coding and maths tasks it was close enough to o3 that the price gap decided things.

That year is over. OpenAI's deprecations page lists o4-mini and its snapshot o4-mini-2025-04-16 for API shutdown on 23 October 2026, six days after this article was published, with gpt-5.6-terra as the recommended replacement. Fine-tuned models built on it, ft-o4-mini-2025-04-16, go on the same date. If you are reading this with o4-mini in production, the last section is the one to start with.

Same price as o3-mini, more model

OpenAI released o4-mini on 16 April 2025, the same day as o3. The model page describes it as a fast, cost-efficient reasoning model optimised for coding and visual tasks.

The comparison with its predecessor is lopsided. o4-mini kept o3-mini's list price of $1.10 per million input tokens and $4.40 output, and added:

• Image input. o3-mini was text-only; o4-mini reads screenshots, diagrams and photos.
• A newer knowledge cutoff. 1 June 2024 against o3-mini's 1 October 2023.
• Cheaper caching. Cached input at $0.275 per million, half o3-mini's $0.55.
• Tool use inside reasoning. Like o3, it could call tools such as web search and Python during its chain of thought, per the shared system card.
• Fine-tuning. Of o3, o3-mini, o3-pro and o4-mini, it is the only one with a fine-tuning endpoint on OpenAI's model pages.

The rest of the specification matches its siblings: 200K context, up to 100K output tokens, available on Chat Completions, Responses and Batch. OpenRouter lists it at $1.10 and $4.40 at the time of writing, with an openai/o4-mini:batch endpoint at $0.55 and $2.20. That is the same model on OpenAI's asynchronous Batch API, discounted in exchange for results that arrive within hours rather than seconds.

Strong at reasoning, weak at recall

The o3 and o4-mini system card is unusually candid about the small model's weak spot. On SimpleQA, a set of short factual questions, o4-mini scored 0.20 accuracy against 0.49 for o3. On PersonQA, questions about people, its hallucination rate was 0.48 against 0.33 for o3 and 0.16 for o1. OpenAI's explanation is the obvious one: smaller models hold less world knowledge and hallucinate more.

Its reasoning held up much better. On OpenAI's internal pull-request replication eval it scored 39% against o3's 44%, a far narrower gap than the factual-recall numbers. That pattern, close on reasoning and far behind on recall, is the most useful thing to know about o4-mini.

The practical consequence is simple. Give o4-mini the facts and let it reason over them: documents, code, retrieved context, tool results. Do not ask it to remember things. If a workflow depends on the model knowing who someone is, what a library's API looked like, or what happened in 2023, either ground it or use a larger model.

o4-mini or o3: the decision

At list prices, o3 costs about 1.8 times as much per token ($2 and $8 against $1.10 and $4.40). For most bounded tasks that gap was not worth paying:

• Pick o4-mini for high-volume coding help, structured extraction from documents and images, and agent steps where the context supplies the facts.
• Pick o3 when the task leans on world knowledge, when one wrong answer is expensive, or when you measured a real pass-rate gap on your own data.

The o3 guide covers the larger model, which stays in the API until 11 December. If you ran o4-mini at high reasoning effort to close the gap, the o4-mini-high guide covers how that changes the cost comparison.

Vision at mini prices

The image input mattered more than the benchmark tables suggest. Reading a screenshot of a failing UI, a chart in a report or a photographed form at $1.10 per million input tokens made vision a routine step rather than a premium one. Images are tokenised, so a large screenshot is not free, but it is cheap enough to include by default. Choosing a model for vision tasks covers what to test before you rely on it.

Moving to the successor this week

With days rather than weeks left, prioritise:

1. Fine-tuned models first. A fine-tune does not migrate. If you depend on ft-o4-mini-2025-04-16, you need a plan for the replacement's own fine-tuning or a prompt-based substitute now.
2. Save a baseline. Run a representative sample through o4-mini today and keep the outputs and token counts.
3. Set reasoning effort explicitly on the replacement instead of inheriting its default.
4. Check image-heavy paths. Different models tokenise images differently, which shifts both cost and context use.
5. Add a fallback. If the new model misbehaves on a request type, route it elsewhere rather than failing.

The GPT-5.6 Terra guide covers OpenAI's recommended replacement.