o4-mini-high: When High Effort Beats Moving Up to o3

The interesting question about o4-mini-high is not what it is. That part is short: OpenRouter describes openai/o4-mini-high as the same model as o4-mini with reasoning_effort set to high, and in ChatGPT it was a picker entry for paid users. The interesting question is whether turning effort up on a small model is a better deal than switching to a bigger one.

That question outlives the model itself. OpenAI shuts down o4-mini in the API on 23 October 2026, and o4-mini-high goes with it. The recommended replacement is gpt-5.6-terra. But the decision between escalating effort and escalating model comes up with every family, and o4-mini against o3 is a clean case to reason about.

What changes when effort goes to high

Nothing about the model's weights, context or price per token. OpenRouter lists o4-mini-high with the same 200K context, 100K maximum output and $1.10 input / $4.40 output per million tokens as o4-mini at the time of writing. Image input, tool calling and structured outputs all behave the same.

What changes is how many hidden reasoning tokens the model produces before answering. Those are billed as output tokens and count against the output cap. At high effort, a request that would have produced a short reasoning trace at medium can produce a much longer one, and you pay the output rate on all of it.

The practical consequences are longer time to first token, a larger bill per request, and a real risk of hitting max_completion_tokens before any visible answer. OpenAI's reasoning guide suggests reserving at least 25,000 tokens while experimenting. The o3-mini-high guide goes deeper on that truncation trap, which works the same way here.

o4-mini-high or o3: where the costs overlap

On a price sheet, o3 at $2 and $8 per million tokens looks like 1.8 times the cost of o4-mini. With effort in the picture, that ratio is not fixed.

Suppose a task costs o4-mini at medium effort 3,000 output tokens including reasoning, and o4-mini-high 7,000. Suppose o3 at medium effort solves it in 3,500. Those figures are illustrative, not measured, but they show the shape. Counting output alone, 7,000 tokens at $4.40 per million costs about 3.1 cents, and 3,500 tokens at $8 per million costs 2.8 cents. The bigger model at normal effort can be cheaper than the small model straining.

Whether that happens depends on your tasks, which is why the comparison has to be measured. Read the reasoning token count from each response's usage details, run the same inputs through both configurations, and compare cost per correct answer. The system card gives one hint about where the gap is widest: o4-mini trails o3 sharply on factual recall (0.20 against 0.49 accuracy on SimpleQA) but only modestly on OpenAI's pull-request replication eval (39% against 44%). Extra thinking does not add knowledge the model never had.

Effort inside agent loops

The effort multiplier compounds in an agent. A loop that makes forty model calls at high effort pays the reasoning premium forty times, and the latency adds up across every step. A loop that felt responsive at medium can become one that takes minutes per task.

Most steps in an agent loop do not need deep reasoning. Reading a file, choosing the next tool, formatting a result: these are routine. The steps that benefit are planning, diagnosing a failing test and deciding a fix is correct. Setting high effort on the whole loop spends the premium on all of them. Agent token budgets covers how to cap that spend without capping capability.

A routing rule: escalate effort, then model

A pattern that works across model families:

1. Start every request at medium effort on the cheaper model.
2. Escalate to high effort when a cheap check fails: a test does not pass, a schema does not validate, a self-check disagrees.
3. Escalate to the larger model when high effort also fails, or immediately for request types that need world knowledge.
4. Log which rung solved each request. After a week you know your real escalation rate, and whether the middle rung earns its place.

If most escalations skip straight past high effort to the larger model, drop the middle rung. If high effort resolves most of them, you have avoided paying large-model rates on the bulk of your traffic. Model routing and fallbacks covers the implementation.

Before 23 October

If you send traffic to o4-mini-high, collect a sample of those requests now, with outputs and reasoning token counts, so you have something to compare against on the replacement. Then rebuild the routing ladder on the new family rather than assuming the old effort settings carry over; newer models accept a wider range of effort values and set their own defaults. The o4-mini guide has the rest of the migration checklist.