o3-mini Guide: OpenAI's Text-Only Reasoner and Its Shutdown

If o3-mini is still in your stack, the most important fact about it is not a benchmark. OpenAI's deprecations page lists o3-mini and its only snapshot, o3-mini-2025-01-31, for API shutdown on 23 October 2026. That is six days after this article was published. The recommended replacement is gpt-5.6-sol.

So this guide does two jobs. It explains what o3-mini was, because that tells you what your prompts were tuned for. And it lays out the shortest sensible path off it, because a week is not long enough for a careful migration.

What o3-mini was for

OpenAI released o3-mini on 31 January 2025, and the launch post set its target precisely: more capability while keeping the low cost and reduced latency of o1-mini. It was a small reasoning model for science, maths and coding, priced to be used routinely rather than reserved for hard problems.

Its real improvement over o1-mini was the developer surface. o3-mini supported function calling, structured outputs, developer messages and streaming, and it accepted a reasoning effort setting. o1-mini had none of the effort control. That combination made it the first small reasoning model you could wire into an ordinary tool-calling application without workarounds.

The specification: 200K-token context, up to 100K output tokens, knowledge cutoff of 1 October 2023, available on Chat Completions, Responses, Assistants and Batch. OpenAI lists it at $1.10 per million input tokens, $0.55 cached and $4.40 output. OpenRouter lists the same rates at the time of writing, plus an openai/o3-mini:batch endpoint at $0.55 and $2.20, which is the same model through the asynchronous Batch API at half price.

One model, three effort levels

o3-mini accepted low, medium and high reasoning effort. In ChatGPT the free tier ran it at medium, and paid users got a separate picker entry, o3-mini-high, which was the same weights at high effort. On OpenRouter that setting is exposed as its own slug, covered in the o3-mini-high guide.

The setting is a trade between depth and spend. Lower effort returns faster and uses fewer reasoning tokens; higher effort thinks longer and generally does better on multi-step problems, at the cost of more hidden reasoning tokens, all billed at the output rate. If your integration never set the parameter, it was running at the default, and your replacement's default may not match.

What it could not do

o3-mini is text-only. The model page lists text input and text output, with no image support, which is the clearest functional gap against o4-mini and o3. If a workflow needed to read a screenshot, a chart or a scanned page, o3-mini was never the right tool.

Its knowledge is also older than its successors'. A cutoff of October 2023 means it does not know about libraries, APIs or events after that point, and it will confidently use deprecated interfaces unless you put current documentation in the prompt.

Two smaller quirks worth knowing if you are debugging old behaviour:

• No Markdown by default. Microsoft's Azure documentation notes that o3-mini avoids Markdown formatting unless the developer message begins with Formatting re-enabled. If your UI expected code fences, that string may be in your prompt for a reason.
• Sampling parameters are rejected. Like other o-series models, it does not accept temperature, top_p or max_tokens. Use max_completion_tokens on Chat Completions.

What the system card says about its coding

The o3-mini system card is a safety document, but it contains one useful capability number. On SWE-bench Verified (477 tasks), o3-mini with OpenAI's internal tool scaffold scored 61%, against 48% for o1. The launch candidate run through the Agentless scaffold scored 39%. The difference between those two figures is a reminder that the harness can matter as much as the model.

The card also notes o3-mini was the first OpenAI model rated Medium risk on the Model Autonomy category of its Preparedness Framework, because of its coding and research-engineering performance. That is a safety classification rather than a product claim, but it shows how far a small model had come by early 2025.

o3-mini or o4-mini: the choice was already made

When o4-mini arrived in April 2025 it took the same $1.10 and $4.40 list price, added image input, a newer knowledge cutoff and a cheaper cached-input rate of $0.275. For most new work there was little reason to pick o3-mini after that. The o4-mini guide covers it, but note that o4-mini shuts down on the same 23 October date, so it is not a migration target either.

What to do this week

1. Find every reference. Search configs and environment variables for o3-mini, including aliases a gateway may map to it.
2. Capture a baseline today. Run twenty representative requests and save the outputs while the model still answers.
3. Switch to the replacement behind a flag. Set reasoning effort explicitly rather than inheriting a default.
4. Diff the outputs and the token counts. Check format, tool-call shape and total output tokens per task.
5. Keep a fallback. Routing with a fallback model means a surprise on the new model degrades a request rather than failing it.