GPT-5.2 Codex is the December 2025 Codex model from OpenAI: GPT-5.2 with post training aimed specifically at long running software engineering work. Inside Atlas it is a strong default for the build agent, because it is tuned to stay on task across many tool calls rather than answering in one shot. GPT-5.2 Codex carries the same 400K context and 128K max output as base GPT-5.2 and costs the same $1.75 per Mtok input and $14 per Mtok output, so the coding specialization costs nothing extra.
What does Codex post training change in GPT-5.2 Codex?
Codex post training in GPT-5.2 Codex targets agentic coding: staying on task across many tool calls rather than answering in one shot. Released in December 2025, GPT-5.2 Codex is the variant of GPT-5.2 tuned for long running software engineering work rather than general chat or analysis.
The distinction shows up on turn 30, not turn 1. Atlas fans out work to subagents that can run in the foreground or in parallel background sessions, and a single Atlas task routinely involves reading a dozen files, running a build, reading the failure, and patching. A general model drifts across that sequence, restating the problem or abandoning the original goal. GPT-5.2 Codex is post trained precisely against that shape of work, which is why it is the sensible pin for the Atlas build agent rather than an occasional escalation.
How much does GPT-5.2 Codex cost compared to GPT-5.2?
GPT-5.2 Codex costs $1.75 per Mtok input and $14 per Mtok output, which is identical pricing to base GPT-5.2. The Codex specialization carries no price premium at all in 2026, so choosing GPT-5.2 Codex over base GPT-5.2 for coding work is a free decision on cost.
Identical pricing between GPT-5.2 Codex and base GPT-5.2 simplifies the decision to a pure capability question: is this task coding, or is it not. If the work is code, GPT-5.2 Codex is the better bet at $1.75 per Mtok input and $14 per Mtok output. If the work is general reasoning, writing, or analysis, base GPT-5.2 is the better default at the same price. There is no cost argument in either direction, which is unusual and worth exploiting.
Can GPT-5.2 Codex read a service and rewrite it in one turn?
GPT-5.2 Codex has a 400K token context with 128K max output, which is enough to read a service and rewrite it in the same turn. Atlas indexes code by AST declarations using tree-sitter, not blind line windows, so the 400K window fills with whole functions and types rather than truncated fragments.
The 400K in, 128K out shape is what makes single turn refactors realistic with GPT-5.2 Codex. You can put an entire service module, its tests, and its call sites into context and ask for the rewritten file, not a diff hint. Atlas searches code with hybrid semantic and keyword retrieval fused by reciprocal rank fusion, so the retrieval that fills that window pulls in the call sites you forgot about. The result is that GPT-5.2 Codex rarely has to guess at a signature it never saw.
When is base GPT-5.2 a better pick than GPT-5.2 Codex?
Base GPT-5.2 is the better pick over GPT-5.2 Codex for non coding reasoning. GPT-5.2 Codex is narrower by design: its December 2025 post training targets software engineering, so for architecture essays, data analysis, or general writing inside Atlas, base GPT-5.2 at the same $1.75 per Mtok is the better default.
The other honest reason to look elsewhere is age. GPT-5.2 Codex was superseded by GPT-5.3 Codex in February 2026 at the same price, so a team starting fresh in 2026 has a newer sibling available for the same money. GPT-5.2 Codex remains a solid, well understood choice, particularly for teams already pinned to it, but there is no cost penalty in moving to the newer Codex generation. Atlas lets you switch the active model and provider on the fly with favorites and recents, so testing both against the same prompt is a matter of one keystroke.
How do you review what GPT-5.2 Codex writes in Atlas?
Review GPT-5.2 Codex output in Atlas by keeping permissions on ask for bash and edit, so every Codex generated write is surfaced as a unified diff before it lands. Every Atlas tool call is permission-gated against allow, ask, and deny rules before it runs, which is exactly the guardrail a long Codex session needs.
GPT-5.2 Codex is post trained to keep going across many tool calls, and that persistence is a double edged property: a confidently wrong Codex run will keep confidently editing. Atlas computes a unified diff for every file edit and surfaces it for approval before writing, so the persistence stays bounded by your review. Atlas also snapshots file changes as git patches so edits can be diffed and rolled back, which means a bad GPT-5.2 Codex run is a revert, not an archaeology project.
Setup
- 01Set OPENAI_API_KEY, or add apiKey under the openai provider in atlas.json.
- 02Run `atlas models openai` and confirm gpt-5.2-codex appears in the resolved model list.
- 03Pin "model": "openai/gpt-5.2-codex" so the main build agent uses the Codex post trained weights.
- 04Keep permissions on ask for bash and edit so every Codex generated write is reviewed as a unified diff.
- 05Let Atlas checkpoint edits as git patches so you can roll back a bad Codex run.
Frequently asked questions
- How much does GPT-5.2 Codex cost per million tokens?
- GPT-5.2 Codex costs $1.75 per Mtok input and $14 per Mtok output, which is exactly the same price as base GPT-5.2.
- What is the difference between GPT-5.2 and GPT-5.2 Codex?
- GPT-5.2 Codex is GPT-5.2 with post training aimed at long running software engineering work: staying on task across many tool calls rather than answering in one shot. Both share a 400K context, a 128K max output, and $1.75 / $14 pricing.
- How do I set GPT-5.2 Codex as the Atlas model?
- Set OPENAI_API_KEY or add apiKey under the openai provider in atlas.json, run `atlas models openai` to confirm gpt-5.2-codex, then pin "model": "openai/gpt-5.2-codex" so the main build agent uses the Codex post trained weights.
- Is GPT-5.2 Codex still worth using in 2026?
- GPT-5.2 Codex is one generation old: GPT-5.3 Codex superseded it in February 2026 at the same price. It remains a capable agentic coding model at $1.75 per Mtok input, but a fresh setup has a newer sibling available for the same money.
- What is the context window of GPT-5.2 Codex?
- GPT-5.2 Codex has a 400K token context window with a 128K max output, the same limits as base GPT-5.2.
- Should I use GPT-5.2 Codex for non coding tasks?
- No. GPT-5.2 Codex is narrower than a general model by design. For non coding reasoning inside Atlas, base GPT-5.2 is the better default, and it costs the same $1.75 per Mtok input and $14 per Mtok output.
- How do I roll back a bad GPT-5.2 Codex run in Atlas?
- Atlas snapshots file changes as git patches so edits can be diffed and rolled back. Keep permissions on ask for bash and edit, and every GPT-5.2 Codex write is surfaced as a unified diff before it touches the working tree.
Try Atlas in your terminal
The terminal-native AI coding agent. Free core, single binary.
Install AtlasRelated guides
Atlas for Erlang in 2026
Atlas is a terminal-native AI coding agent for Erlang/OTP in 2026. Run it in an app with a rebar.config, map supervisors and gen_server modules, review every diff.
Run the Test Suite and Triage the Failures with Atlas in 2026
How to triage a failing test suite with Atlas in 2026: bash truncates at 2000 lines or 50 KB and saves the full log, then grep groups failures by root cause.
Atlas vs Zed: Terminal AI Coding Agents in 2026
Comparing Atlas, the terminal-native AI coding agent, with Zed, the GPU-accelerated collaborative editor, for developers in 2026. Evaluate their AI autonomy, privacy, and workflow.
Atlas for Electron: Terminal-Native AI Coding for Main, Preload, and Renderer in 2026
Atlas is a terminal-native AI coding agent for Electron in 2026, where the main and renderer split, contextIsolation, and preload bridges are the security model.
Atlas for Rust in 2026
Adopt Atlas, the terminal-native AI coding agent, for Rust development in 2026. Tackle borrow checker errors and clippy lints with Atlas's secure, approval-gated assistance.
Atlas vs PearAI: Choosing Your AI Coding Agent in 2026
Compare Atlas, the terminal-native AI coding agent, with PearAI, a VS Code fork. Evaluate features, pricing, and community support for developers in 2026.
Upgrade a Dependency and Fix the Breakage with Atlas (2026 Workflow)
How to upgrade a dependency and fix the breakage with Atlas in 2026: bash drives the package manager, webfetch pulls the release notes, edit fixes each compiler error.
Automate GitHub Issue and Pull Request Triage with Atlas (2026 Workflow)
How to automate GitHub issue and pull request triage with Atlas in 2026: the atlas github command checks the actor has admin or write permission before it does anything.