GPT-5.2 Pro Guide: Paying $168 per Million Output Tokens

GPT-5.2 Pro is not a faster or broader GPT-5.2. It is the same generation given more reasoning time and more compute per request, sold at a price that only makes sense if you send it very few, very hard questions. Most of the mistakes people make with it come from treating it like a default model.

OpenAI released the GPT-5.2 family in December 2025 as three tiers: Instant, Thinking and Pro. Pro sits at the top. OpenRouter records its listing on 10 December, and OpenAI's API snapshot is dated a day later, gpt-5.2-pro-2025-12-11. Either way it is now ten months old, and that shapes the decision below.

What the Pro tier actually is

Thinking and Pro are both reasoning models: they spend hidden output tokens working through a problem before they answer. Pro spends more of them. OpenAI's API page for the model exposes three reasoning effort levels, medium, high and xhigh. There is no low or none setting, which tells you what it is for. If you want a quick answer, this is the wrong model at every setting.

The published limits match the rest of the GPT-5.2 generation: a 400,000-token context window, up to 128,000 output tokens, text and image input, text output, and a knowledge cutoff of 31 August 2025. Tool support covers function calling, web search, file search, image generation and MCP.

OpenAI has not published a parameter count or architecture for any GPT-5.2 tier, and there are no open weights. Anything you read about its internals is speculation.

The price, in units you can feel

OpenAI lists GPT-5.2 Pro at $21 per million input tokens and $168 per million output tokens, and OpenRouter shows the same figures at the time of writing. Prices change, so check before you budget.

Per-token rates hide the real cost of a reasoning model, because the expensive part is output you never see. Suppose a hard question carries 20,000 tokens of context, and the model reasons for 30,000 tokens before writing a 2,000-token answer. That is about $0.42 of input and $5.38 of output: close to six dollars for one reply. Put the same shape inside an agent loop that makes forty calls and a single task costs more than two hundred dollars.

That arithmetic is why Pro belongs at the end of a pipeline, not inside its loop. The token cost of reasoning models covers how hidden reasoning tokens show up on the invoice and how to measure them per request.

Responses API only, and requests that take minutes

Two operational details catch people out. First, the model is served on the Responses API only. OpenAI's model page marks Chat Completions, Realtime and its own Batch endpoint as unsupported, so a client or coding tool hard-wired to /v1/chat/completions will not reach it directly from OpenAI.

Second, OpenAI warns that some requests may take several minutes to finish and recommends background mode to avoid timeouts. Background mode means you submit the request, get an id back, and poll for the result when it is ready. It is a job queue, not a chat call, and your code should be shaped accordingly.

If you route through a proxy or gateway, check its idle timeout before anything else. A 60-second or 100-second limit somewhere between you and the provider will kill exactly the requests you are paying the most for. Timeout tuning for LLM calls covers picking numbers that hold for long reasoning requests.

It has been superseded, twice

The honest context for any GPT-5.2 Pro decision in October 2026 is that OpenAI itself points elsewhere. Its model page now recommends GPT-5.5 Pro as the newer pro model, and GPT-5.4 Pro shipped in between.

Third-party trackers list GPT-5.2 Pro at about 54% on ARC-AGI-2, and coverage of the GPT-5.4 launch reports a considerably higher figure for GPT-5.4 Pro on the same test. These are vendor-reported numbers, but the direction is not ambiguous: newer Pro tiers score higher on exactly the kind of hard reasoning that Pro exists to sell.

So the real reasons to stay on 5.2 Pro are narrow:

• You have a validated pipeline. An evaluation set, prompts tuned to this model's behaviour, and downstream code that depends on its output format. Swapping models is a regression-testing job, not a configuration change.
• You need comparability with past runs. Research or audit work that must line up with results produced in early 2026.
• Your access path only offers this tier. Some providers and gateways lag behind OpenAI's own catalogue.

If none of those apply, start at the newest Pro tier, covered in the GPT-5.5 Pro guide, and work backwards only if cost forces you to.

When paying for Pro makes sense at all

The test is simple: is a wrong answer more expensive than the reply? Pro pays off on questions where a person would otherwise spend hours, and where checking an answer is much cheaper than producing one.

• A concurrency bug that two cheaper models have already failed on, with the full trace attached.
• A proof sketch, an algorithm design or a numerical method you can verify independently afterwards.
• A one-off review of a dense, regulation-heavy document set where a missed clause is costly.

It does not pay off on bulk code generation, chat, classification, summarisation, or anything inside an agent loop. Use the base GPT-5.2 or a newer generation there, and escalate to Pro only when the cheaper model fails a check you can automate.

A note on the listed variants

OpenRouter lists two endpoints: openai/gpt-5.2-pro and openai/gpt-5.2-pro:batch. A :batch listing is the same model on an asynchronous endpoint at a discount, here $10.50 input and $84 output per million tokens, half the standard rate. Because OpenAI's own model page lists its Batch endpoint as unsupported for this model, confirm the batch route really accepts your requests before you plan a budget around it.

For new work the rule is short: escalate to a Pro tier rarely and deliberately, prefer the newest one, run it in background mode, and log reasoning tokens on every call so you know what each answer actually cost.