OpenAI ships GPT‑5.1‑Codex‑Max, its first model built for million-token coding sessions

OpenAI just dropped GPT‑5.1‑Codex‑Max, a frontier agentic coding model that does something none of its predecessors could: work coherently across millions of tokens in a single task. The secret is a training technique the company calls compaction, which lets the model natively operate across multiple context windows without losing the plot. That matters if you’ve ever watched an AI coding assistant forget what it was doing halfway through a large refactor.

The model was trained on real-world software engineering workflows — pull request creation, code review, frontend coding, and Q&A — and sits on top of an updated reasoning backbone that also covers math, research, medicine, and general computer use. It’s the first OpenAI model purpose-built for the kind of long-horizon, multi-step development tasks that eat up a senior engineer’s afternoon.

On the safety side, the system card reveals a layered defense strategy. Model-level mitigations include specialized training to resist harmful tasks and prompt injections, while product-level protections lean on agent sandboxing and configurable network access. OpenAI evaluated GPT‑5.1‑Codex‑Max under its Preparedness Framework and found it “very capable” in cybersecurity but still below the High threshold — though the company warns that current capability trends suggest models will cross that line soon. It is being treated as High capability on biology and deployed with the same safeguards used for GPT‑5. AI self-improvement remains below High.

What’s striking is the honesty about trajectory. OpenAI isn’t pretending the safety picture is static. They’re telling us the cybersecurity threshold is going to get crossed, probably soon, and they’re shipping this model knowing it pushes right up against that line. Whether the sandboxing and compaction controls hold up under real adversarial pressure is the question nobody can answer yet.