GPT-5.2-Codex: What Changed From Codex-Max, and What It Cost

When a vendor ships a coding-specialised version of its general model, the question worth asking is how much the specialisation buys over the general model at the same price. For GPT-5.2-Codex the answer, on OpenAI's own numbers, was less than a point on SWE-Bench Pro. The interesting changes were elsewhere.

This guide covers what GPT-5.2-Codex changed relative to GPT-5.1-Codex-Max, how it compared with plain GPT-5.2, and where it stands now that OpenAI has retired it.

Two release dates, a month apart

OpenAI announced GPT-5.2-Codex on 18 December 2025, and it became available to paid ChatGPT users through the Codex CLI, IDE extensions, cloud tasks and code review from 19 December. API access came later: the changelog records gpt-5.2-codex arriving in the Responses API on 14 January 2026. Some catalogues use that January date as the release date, which is why sources disagree.

The gap was deliberate. OpenAI said it was preparing to enable API access securely, which ties into the cybersecurity framing below.

What changed versus GPT-5.1-Codex-Max

OpenAI positioned it as GPT-5.2 tuned for agentic coding, and listed four areas of improvement over the previous Codex model:

• Native compaction and stronger long-context understanding, so task state survives long sessions.
• Large code changes — more reliable tool calling on refactors, upgrades and migrations.
• Windows — better behaviour in native Windows environments.
• Vision — better reading of screenshots, technical diagrams, charts and UI mock-ups shared during a task.

The knowledge cutoff also moved forward with the GPT-5.2 base, to 31 August 2025, against 30 September 2024 for the GPT-5.1 Codex models. For a coding model that matters more than it sounds: almost a year of library releases and API changes that the older model simply did not know about.

The benchmark margin over GPT-5.2 is small

OpenAI reported these results:

• SWE-Bench Pro — 56.4%, against 55.6% for GPT-5.2 and 50.8% for GPT-5.1
• Terminal-Bench 2.0 — 64.0%, against 58.1% that OpenAI had reported for GPT-5.1-Codex-Max at launch

Read the first line carefully. Against the previous generation the gain is real. Against the general GPT-5.2 model, at the same per-token price, it is 0.8 of a point on a vendor-run evaluation, which is within the range that scaffold and sampling differences can move. SWE-bench explained covers why a percentage on a patch benchmark is a narrower claim than it looks.

The Terminal-Bench gap over Codex-Max is larger and more relevant to agents that drive a shell. Even so, the honest summary is that GPT-5.2-Codex was a better agent harness partner rather than a much smarter coder than GPT-5.2.

The cybersecurity framing

OpenAI described GPT-5.2-Codex as its strongest cybersecurity model to date, while saying it did not reach the High cybersecurity capability level in its Preparedness Framework. It cited gains on professional capture-the-flag tasks, and pointed to a researcher who, using GPT-5.1-Codex-Max and the Codex CLI while studying a critical React Server Components flaw, found and disclosed further vulnerabilities.

Alongside the launch OpenAI announced an invite-only trusted access pilot for vetted security professionals doing defensive work. For an ordinary engineering team the practical takeaway is narrower: a model that is good at finding vulnerabilities is also good at writing code that touches security boundaries, and its changes there deserve the same review a human contributor's would.

Price, effort levels and the input limit

OpenAI listed the model at $1.75 per million input tokens, $0.175 cached input and $14 per million output, matching GPT-5.2 and 40% above the $1.25 and $10 of the GPT-5.1 Codex models. The catalogue we sync from OpenRouter showed the same input and output rates at the time of writing.

Reasoning effort ran from low through medium and high to xhigh; there was no none setting, unlike GPT-5.2. The context window was 400,000 tokens, but the model page caps input at 272,000 with up to 128,000 for output. If your harness packs the prompt to 400K, requests will fail before compaction ever runs. Like the other Codex models it was Responses-only, with no Chat Completions or Batch support. On 3 February 2026 OpenAI reported making GPT-5.2 and GPT-5.2-Codex about 40% faster through inference optimisation, with the weights unchanged.

Retired: what to do with a 5.2-Codex workflow

OpenAI announced the deprecation of gpt-5.2-codex on 22 April 2026, with removal from the API on 23 July 2026 and gpt-5.6-sol as the named replacement. If a catalogue still lists it, confirm with a test request before depending on it.

When you move, budget for more than the per-token difference. What a model upgrade costs covers the hidden parts: re-tuning prompts, re-running evaluations, and the change in turns per task. For an agentic coding model, turns per accepted change is the number that decides the bill, and it is the one to measure first.