GPT-5 Pro Guide: When $120 Per Million Output Tokens Pays Off

GPT-5 Pro is the most expensive model in the GPT-5 generation by a wide margin, and most of the confusion around it comes from treating it like a normal chat model with a higher price. It is not. It has one effort setting, one interactive endpoint, requests that can take minutes, and a bill that only makes sense for a narrow class of problems.

OpenAI brought it to the API with the snapshot gpt-5-pro-2025-10-06, and describes it as a version of GPT-5 that uses more compute to think harder and produce smarter, more precise responses. This guide is about the practical consequences of that sentence.

What you are paying for

OpenAI lists GPT-5 Pro at $15 per million input tokens and $120 per million output tokens, with no cached-input discount. For comparison, standard GPT-5 lists at $1.25 and $10. Pro is twelve times the price on both sides.

The output rate is the one that hurts, because Pro is a reasoning model and its hidden reasoning tokens bill as output. A model designed to think harder produces more of them by construction. A single hard request that burns 30,000 reasoning tokens costs $3.60 in output alone, before the visible answer.

The other specifications, from OpenAI's model page:

• Context window — 400,000 tokens
• Maximum output — 272,000 tokens. Some catalogues list 128,000; OpenAI's own figure is the one to plan around.
• Modalities — text and image in, text out
• Knowledge cutoff — 30 September 2024

It only runs one way

Three constraints catch people who swap Pro into an existing integration.

Effort is fixed at high. OpenAI lists high as both the default and the only supported reasoning effort. Some catalogues show "reasoning effort control" as a capability; for this model there is nothing to control. If your code sends low to save money, it will not get a cheaper response, so check how your client handles an unsupported value before relying on it.

Responses API only. For interactive use, Pro is served through the Responses API. Chat Completions is not supported, so a client built on the older endpoint needs changing before it can call Pro at all. The Batch endpoint is supported.

No code interpreter. Function calling, structured outputs, web search, file search, image input and MCP are supported. The hosted code interpreter is not, so a workflow that relies on the model running Python needs another route.

Calling it without timeouts

OpenAI warns that some Pro requests take several minutes and recommends background mode to avoid timeouts. A synchronous HTTP call held open for four minutes will be cut by a proxy, a load balancer or a serverless function limit long before the model finishes, and you still pay for the work.

Background mode returns immediately with a response ID, and you poll for the result:

from openai import OpenAI
import time

client = OpenAI()

job = client.responses.create(
    model="gpt-5-pro",
    input="Review this migration plan for data-loss risks: ...",
    background=True,
)

while job.status in ("queued", "in_progress"):
    time.sleep(10)
    job = client.responses.retrieve(job.id)

print(job.output_text)

Set a generous max_output_tokens as well. The cap covers reasoning plus visible output, and a request that hits it returns as incomplete: you pay for the reasoning and get no answer. On a model at this price, that is the most expensive failure mode available. max_tokens and truncation explains the mechanics.

A Pro model is not the same as pro mode

Newer OpenAI models handle this differently. GPT-5.6 and GPT-6 models support a reasoning.mode setting with standard and pro values, where pro mode does more work on the same model and bills the tokens at that model's normal rates. OpenAI's documentation says existing Pro model IDs keep their current behaviour and pricing.

So GPT-5 Pro is a separate model ID with its own premium rate card, while something like GPT-6.1 Sol Pro is the base model run in pro mode. If you are choosing today rather than maintaining an existing integration, compare against the newer approach, because the per-token price gap is large.

When it is still worth it

Pro makes sense when three things are true at once: the question is genuinely hard, a wrong answer is expensive, and nobody is waiting on the response in real time. Reviewing a database migration for data-loss paths, checking a proof or a pricing model, or producing a careful second opinion on an architecture decision all fit.

It is the wrong choice for anything interactive, anything high-volume, and anything where standard GPT-5 at high effort already gets the right answer. That last case is more common than people expect. Before routing work to Pro, run the same ten hard tasks through GPT-5 at high effort and through Pro, and count how many answers actually change. If the answer is one or two, you are paying twelve times the price for a small improvement on a handful of tasks.

Batch, and how to keep the bill bounded

OpenRouter lists openai/gpt-5-pro:batch, the same model on the asynchronous batch endpoint at $7.50 input and $60 output per million tokens at the time of writing. Since Pro is already slow, batch costs you little in experience and halves the bill. For overnight review jobs it should be the default route.

A practical checklist before turning Pro on:

1. Confirm your client uses the Responses API and background mode.
2. Set a per-request max_output_tokens and a daily spend alert.
3. Route to Pro only after a cheaper model has failed or flagged low confidence.
4. Log reasoning token counts per request, so you can see what a hard question actually costs.