ModelsGPT-5.2 Guide: xhigh Effort, GDPval and a 40% Price Rise
GPT-5.2 added an xhigh effort level, a 2025 knowledge cutoff and a 70.9% GDPval claim, and raised the price 40% over GPT-5.1. How to judge whether it earns that.
GPT-5.2 is the first model in OpenAI's GPT-5 line that cost more than the one before it. GPT-5 and GPT-5.1 shared a price; GPT-5.2 raised both input and output rates by 40%. That makes it a useful case study in a question every team faces on each upgrade: is the new model better enough, on your work, to pay for itself?
OpenAI pitched it at professional knowledge work and long-running agents. This guide covers what it added, how to read the headline benchmark, and how to decide between it and its neighbours.
What GPT-5.2 added
OpenAI released GPT-5.2 on 11 December 2025, as Instant, Thinking and Pro variants in ChatGPT and as gpt-5.2 in the API, with the dated snapshot gpt-5.2-2025-12-11. The API changelog lists three additions over GPT-5.1:
- An xhigh reasoning effort level, above high, for tasks where latency is not the constraint.
- Concise reasoning summaries, so you can see the shape of the model's thinking without paying to read all of it.
- Compaction-based context management, the technique the Codex models introduced for keeping long sessions coherent.
The knowledge cutoff moved to 31 August 2025, nearly a year later than GPT-5.1's 30 September 2024. The envelope stayed the same: a 400,000-token context window, up to 128,000 output tokens, text and image in, text out, and support for Chat Completions, Responses and Batch. The default reasoning effort remains none, as in GPT-5.1, with low, medium, high and xhigh above it.
Reading the GDPval number
The headline claim was GDPval. OpenAI reported that GPT-5.2 Thinking beat or tied industry professionals in 70.9% of comparisons on well-specified knowledge-work tasks across 44 occupations, as judged by expert graders. The figure it gave for GPT-5 on the same evaluation was 38.8%, and reports of the launch put GPT-5.2 Pro at 74.1%.
Three things to keep in mind. It is a vendor-run evaluation of the vendor's own model. It measures well-specified tasks, which is the easy half of most real jobs; the hard half is working out what the task is. And beating or tying a professional on a deliverable is not the same as being trusted to ship that deliverable without review. GDPval explained covers what the benchmark does and does not capture.
OpenAI also reported a roughly 30% reduction in responses containing errors compared with GPT-5.1 Thinking. That is the kind of improvement that matters more in production than a leaderboard position, and it is also the hardest to verify without your own test set.
Coding results in context
Launch reports put GPT-5.2 Thinking at 55.6% on SWE-Bench Pro and 80.0% on SWE-bench Verified. The coding-tuned GPT-5.2-Codex that followed scored 56.4% on SWE-Bench Pro, under a point higher, at the same price per token.
That small gap is the useful finding. For most coding work the general model was nearly as capable as the specialist, and it had things the specialist lacked: Chat Completions support, Batch support and a none effort setting for quick turns. The Codex model earned its place in long agent harnesses, not in general use.
Paying 40% more than GPT-5.1
OpenAI lists GPT-5.2 at $1.75 per million input tokens, $0.175 cached input and $14 per million output. GPT-5.1 is $1.25, $0.125 and $10. The catalogue we sync from OpenRouter showed the same GPT-5.2 rates at the time of writing.
Whether that 40% is worth it depends on where your spend goes. If most calls run at none effort on short prompts, the quality difference is hard to see and the price difference is not. If your spend is dominated by long agent runs, a model that finishes in fewer turns or fails less often can cost less per completed task despite the higher rate. The only way to know is to measure cost per accepted result on your own tasks, not cost per token.
OpenRouter also lists a gpt-5.2:batch variant: the same model on a discounted asynchronous endpoint, shown at $0.875 input and $7 output per million tokens at the time of writing. For evaluation runs and bulk jobs that do not need an answer straight away, that puts it below GPT-5.1 at standard rates.
Where GPT-5.2 stands in October 2026
On 3 February 2026 OpenAI reported making GPT-5.2 about 40% faster through inference optimisation, with the weights unchanged. That is a reminder that latency figures from launch week can go stale even on a pinned snapshot.
Unlike most of its siblings, gpt-5.2 is not on OpenAI's deprecations list at the time of writing. Its model page calls it the previous flagship and recommends GPT-6 Astra for new work. The related GPT-5.2 Chat alias and GPT-5.2-Codex have both been retired.
If you run GPT-5.2 today, the practical checklist is:
- Pin the dated snapshot rather than the bare id, so a future alias change cannot alter behaviour silently.
- Set reasoning effort per call type; reserve
xhighfor tasks where you have measured that it helps. - Move anything non-interactive to the batch endpoint.
- Keep a regression set ready, because a model described as the previous flagship is a model whose deprecation notice is a matter of time.
Common questions
How much more does GPT-5.2 cost than GPT-5.1?
OpenAI lists GPT-5.2 at $1.75 input and $14 output per million tokens, against $1.25 and $10 for GPT-5.1: 40% more on both. Whether that costs more per finished task depends on how many turns each model needs on your work.
What does the xhigh reasoning effort level do?
It lets the model spend more reasoning tokens than the high setting, for tasks where depth matters more than latency. Those tokens are billed as output, so use it where you have measured a quality gain rather than as a default.
Is GPT-5.2 deprecated?
Not at the time of writing. OpenAI's model page calls it the previous flagship and recommends GPT-6 Astra for new work, but gpt-5.2 does not appear on the deprecations list, unlike GPT-5.2 Chat and GPT-5.2-Codex.
