GPT-5.2 Pro is the high effort Pro variant of GPT-5.2 from OpenAI, released 2025-12-11. Inside Atlas it is a deliberate, on demand escalation, not a default. GPT-5.2 Pro keeps the same 400K token context as base GPT-5.2 but is priced at $21 per Mtok input and $168 per Mtok output, which is 12x the input price of base GPT-5.2 at $1.75. That buys substantially more reasoning compute per request. Use it for the single hard question, then switch back.
What is GPT-5.2 Pro best at inside Atlas?
GPT-5.2 Pro, released 2025-12-11, is the high effort reasoning tier of GPT-5.2, and inside Atlas it is best at one hard question at a time: an architecture decision, a subtle concurrency bug, or a migration design that has to be right on the first attempt.
Atlas searches code with hybrid semantic and keyword retrieval fused by reciprocal rank fusion, so by the time a question reaches GPT-5.2 Pro, the relevant declarations are already in the window rather than a pile of guessed line ranges. That pairing matters at Pro rates. GPT-5.2 Pro spends a much larger reasoning budget per turn than base GPT-5.2, and the value of that budget is proportional to how well posed the question is. A well retrieved, well scoped prompt against GPT-5.2 Pro is where the $21 per Mtok input goes furthest. An open ended "look around the repo and see what is wrong" prompt is where it is wasted.
How much does GPT-5.2 Pro cost per Mtok?
GPT-5.2 Pro costs $21 per Mtok input and $168 per Mtok output. That is 12x the input price of base GPT-5.2, which sits at $1.75, and it is above the older GPT-5 Pro on both sides: $21 versus $15 on input and $168 versus $120 on output, but on a newer base model.
The output side is the number that should shape how you drive GPT-5.2 Pro in Atlas. At $168 per Mtok output, a model that thinks at length and answers at length is a model that adds up fast across a session. The input side, $21 per Mtok, means that stuffing a 400K token window full of speculative context is also a real cost, not a rounding error. Track spend in the TUI session cost footer while you work. GPT-5.2 Pro is the one model in the OpenAI lineup where watching that footer during the run, rather than after it, changes decisions.
Does GPT-5.2 Pro's 400K context fit a whole subsystem?
GPT-5.2 Pro holds the same 400K context as base GPT-5.2, plus a 128K max output, so it can hold a full subsystem and still emit a large design doc or patch in one response. Atlas indexes code by AST declarations using tree-sitter, not blind line windows, so that 400K is filled with whole symbols.
The combination of a 400K token context and a 128K output ceiling is what makes GPT-5.2 Pro usable for a genuine one shot deliverable. You can put an entire service in front of GPT-5.2 Pro and ask for a complete rewritten module or a full migration plan, rather than a fragment you then have to stitch together across follow up turns. Each of those follow up turns would re-bill the input at $21 per Mtok, so the ability to answer once and answer completely is not just convenient, it is the cost model working in your favor.
When should you pick a different model than GPT-5.2 Pro?
Pick a different model than GPT-5.2 Pro for any long agent loop. At $168 per Mtok output, dozens of sequential tool calls become genuinely expensive, and GPT-5.2 Pro is slow by construction, which is a bad fit for the interactive Atlas TUI edit loop where you want a response in seconds.
GPT-5.2 Pro is a consultant, not an employee. The Atlas edit loop, where the agent reads a file, proposes a unified diff, waits for approval, runs a command, and reads the output, is a loop that rewards speed and discipline over raw reasoning depth. Base GPT-5.2 at $1.75 per Mtok input is the sane default for that loop, and the Codex post trained variants are built for it. Reserve GPT-5.2 Pro for the moment the loop stalls: nobody on the team knows why the invariant breaks, and one very expensive, very deep answer is cheaper than another day of guessing.
How do you keep GPT-5.2 Pro spend under control in Atlas?
Keeping GPT-5.2 Pro spend under control in Atlas means never leaving it pinned. Leave "model" on a cheaper tier, switch to gpt-5.2-pro on demand via /models for one shot deep reasoning, then switch back. Atlas lets you switch the active model and provider on the fly with favorites and recents.
The favorites list in the Atlas /models picker is the practical mechanism here. Keep base GPT-5.2 as the pinned model in atlas.json and add gpt-5.2-pro as a favorite so the escalation is one keystroke and the return trip is one keystroke. Because Atlas keeps the session and its context across a model switch, you do not lose your place: you hand the already assembled context to GPT-5.2 Pro, take the answer, and drop back to the cheap tier to actually apply it. The $21 per Mtok input meter then runs for one turn instead of a whole afternoon.
Can the Atlas plan agent make GPT-5.2 Pro cheaper?
The Atlas plan agent makes GPT-5.2 Pro cheaper by keeping exploration off the $21 per Mtok meter. Atlas drafts a plan in a read-only plan agent and asks before switching to a build agent, so you can reach GPT-5.2 Pro with a tight prompt that is already scoped and already grounded.
Use the Atlas plan agent to produce a tight prompt before invoking GPT-5.2 Pro, so you are not paying Pro rates for exploration. Exploration is cheap work: read files, grep, list symbols, form a hypothesis. Run it on a cheap model. Then hand GPT-5.2 Pro exactly the question you could not answer, with exactly the code that bears on it. Atlas computes a unified diff for every file edit and surfaces it for approval before writing, so whatever GPT-5.2 Pro produces still lands under review rather than straight into the working tree.
Setup
- 01Export OPENAI_API_KEY in your shell so the Atlas openai provider can authenticate.
- 02Confirm resolution with `atlas models openai` and check that gpt-5.2-pro appears in the list.
- 03Leave "model" on a cheaper tier in atlas.json and switch to gpt-5.2-pro on demand via /models for one shot deep reasoning.
- 04Use the Atlas plan agent to produce a tight prompt before invoking GPT-5.2 Pro, so you are not paying Pro rates for exploration.
- 05Track spend in the TUI session cost footer during the run, because $168 per Mtok output moves the number visibly.
Frequently asked questions
- How much does GPT-5.2 Pro cost?
- GPT-5.2 Pro costs $21 per Mtok input and $168 per Mtok output. That is 12x the input price of base GPT-5.2, which is $1.75 per Mtok input.
- What is the context window of GPT-5.2 Pro?
- GPT-5.2 Pro has a 400K token context window, the same as base GPT-5.2, along with a 128K max output ceiling.
- Is GPT-5.2 Pro worth it over GPT-5.2 for coding?
- For the everyday Atlas edit loop, no: base GPT-5.2 at $1.75 per Mtok input does the same work far cheaper. GPT-5.2 Pro earns its $168 per Mtok output only on single hard questions where a larger reasoning budget per request changes the answer.
- How do I switch Atlas to GPT-5.2 Pro?
- Export OPENAI_API_KEY, run `atlas models openai` to confirm gpt-5.2-pro resolves, then open /models in the Atlas TUI and switch to it on demand. Atlas lets you switch the active model and provider on the fly with favorites and recents.
- Is GPT-5.2 Pro cheaper than GPT-5 Pro?
- No. GPT-5.2 Pro is priced above the older GPT-5 Pro on both sides, $21 versus $15 on input and $168 versus $120 on output, but it runs on a newer base model.
- Can I run GPT-5.2 Pro as the default Atlas model?
- You can, but it is a bad idea. GPT-5.2 Pro is slow by construction and $168 per Mtok output makes long agent loops genuinely expensive. Leave "model" on a cheaper tier and escalate to gpt-5.2-pro only when a question needs it.
- When was GPT-5.2 Pro released?
- GPT-5.2 Pro was released 2025-12-11 as the high effort Pro variant of GPT-5.2, keeping the 400K context of the base model while spending substantially more reasoning compute per request.
- How do I track GPT-5.2 Pro spend in Atlas?
- Watch the TUI session cost footer while the run is in progress. At $21 per Mtok input and $168 per Mtok output, GPT-5.2 Pro is the one model where the footer changes fast enough to influence what you do next.
Try Atlas in your terminal
The terminal-native AI coding agent. Free core, single binary.
Install AtlasRelated guides
Atlas for Astro: Islands, Content Collections, and Zero JS by Default in 2026
Atlas is a terminal-native AI coding agent for Astro in 2026. It reads astro.config.mjs, src/pages, and content collection schemas, drops needless client:load directives, and runs astro check.
Onboard to an Unfamiliar Codebase with Atlas in 2026
How to onboard to an unfamiliar codebase with Atlas in 2026: use codebase_search, glob, read, lsp, task, and todowrite to build a mental model fast.
Atlas for Perl: A Terminal-Native AI Coding Agent for CPAN Distributions in 2026
Atlas is a terminal-native AI coding agent for Perl in 2026. It reads cpanfile deps and @EXPORT lists, writes Test2::V0 cases, runs prove -lr t/, and runs perltidy on the diff.
Atlas for Haskell in 2026
Atlas is a terminal-native AI coding agent for Haskell in 2026. Run it in a project with a .cabal file or a stack.yaml, let cabal build drive it, and review the diff.
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.
Atlas vs Codebuff: Terminal AI Coding Agents in 2026
Atlas and Codebuff are terminal AI coding agents for 2026. Compare Atlas's terminal-native TUI, permission-gated tools, and diff review with Codebuff's multi-agent system and flexible pricing.
Atlas vs Magic.dev: Terminal AI Coding Agents in 2026
Compare Atlas, the terminal-native AI coding agent with permission-gated tools and diff review, against Magic.dev's research claims of 100 million token context models in 2026.
Atlas vs Warp: Choosing Your AI Coding Agent in 2026
Compare Atlas, the terminal-native AI coding agent, with Warp, a smart terminal with AI Agent Mode, for developers in 2026. Evaluate features, privacy, and workflow.