GPT-5.4 nano is OpenAI's cheapest current reasoning-capable model at $0.20 / $1.25 per Mtok (input / output) with a 400K token context window. It is built for the high-volume, low-stakes calls Atlas makes constantly: titles, summaries, classification. GPT-5.4 nano is explicitly not a build-agent model. It drops invariants on multi-file work and will churn through Atlas permission prompts, so pin it to small_model and leave a real model on model.
What are the high-volume background calls GPT-5.4 nano handles in Atlas?
GPT-5.4 nano handles the high-volume, low-stakes calls Atlas makes constantly: titles, summaries, classification. At $0.20 / $1.25 per Mtok, those per-message background calls stop being a line item you think about, which is the entire design goal of the nano tier.
Every Atlas session generates model traffic you never see. Naming a session. Summarizing a diff after Atlas reads git branches, status, and diffs. Deciding whether a file is relevant. None of it needs judgment, all of it needs to be fast and nearly free. GPT-5.4 nano at $0.20 per Mtok input is 25 times cheaper than GPT-5.6's $5, which is what makes it viable for per-message background calls rather than something you batch and ration.
Why does GPT-5.4 nano keep a 400K context window?
GPT-5.4 nano retains a 400K context window despite the nano tier, so it can still read a large file it was handed. Most cheap tiers pair a low price with a small window, which forces you to spend a model call summarizing before you can spend the cheap one.
The 400K window is what makes GPT-5.4 nano usable rather than merely cheap. A background call that summarizes a large diff has to hold that diff. Because Atlas indexes code by AST declarations using tree-sitter, a handed-off chunk is often a full declaration rather than a truncated fragment, and 400K comfortably takes it. GPT-5.4 nano also supports reasoning and tool calling, so a background subagent running on it can search and read rather than only rephrase what it was given.
Why should GPT-5.4 nano never be your Atlas build agent?
GPT-5.4 nano is not a build-agent model. It drops invariants on multi-file work and will churn through Atlas permission prompts, so a multi-file refactor on GPT-5.4 nano costs you attention at every approval and still lands the wrong diff.
The failure is structural, not stylistic. Every Atlas tool call is permission-gated against allow, ask, and deny rules before it runs, so a model that proposes bad calls does not fail silently: it fails at you, repeatedly, at the approval prompt. GPT-5.4 nano's price is irrelevant in that scenario, because the scarce resource is your review bandwidth, not $0.20 per Mtok. Never pin GPT-5.4 nano as model unless you are only doing read-only exploration.
How much cheaper is GPT-5.4 nano than a frontier model?
GPT-5.4 nano charges $0.20 per Mtok input, 25 times cheaper than GPT-5.6's $5. Output runs $1.25 per Mtok. At that ratio, a thousand background calls on GPT-5.4 nano cost less than a handful of frontier turns, which changes what background work is worth doing at all.
A 25x price gap is not an optimization, it is a category change. Work that would be indefensible at $5 per Mtok input becomes routine at $0.20: classify every file touched, summarize every subagent result, title every session. Atlas fans out work to subagents that can run in the foreground or in parallel background sessions, and GPT-5.4 nano is what makes generous fan-out affordable. The frontier model then only sees the turns that actually need it.
How do you assign GPT-5.4 nano to the cheap slot in Atlas?
Assign GPT-5.4 nano to the cheap slot in atlas.json with "small_model": "openai/gpt-5.4-nano". Set OPENAI_API_KEY or run atlas login first, then confirm with atlas models openai that the $0.20 / $1.25 per Mtok tier actually resolves against your key.
The two-slot split is the whole configuration. small_model takes GPT-5.4 nano and absorbs the titles, summaries, and classification traffic. model stays on something that can actually build, because GPT-5.4 nano drops invariants on multi-file work. Never pin GPT-5.4 nano as model unless you are only doing read-only exploration, where nothing it proposes will be written and Atlas is not asking you to approve edits.
Setup
- 01Set `OPENAI_API_KEY`, or run `atlas login`.
- 02Confirm with `atlas models openai`.
- 03Assign it to the cheap slot in atlas.json: `"small_model": "openai/gpt-5.4-nano"`.
- 04Never pin it as `model` unless you are only doing read-only exploration.
- 05Keep a frontier model in the `model` slot so multi-file edits are not driven by the nano tier.
Frequently asked questions
- how much does gpt-5.4 nano cost per million tokens
- GPT-5.4 nano costs $0.20 / $1.25 per Mtok (input / output). Its $0.20 input rate is 25 times cheaper than GPT-5.6's $5.
- what is the cheapest openai model for atlas
- GPT-5.4 nano is OpenAI's cheapest current reasoning-capable model at $0.20 / $1.25 per Mtok, and it still carries a 400K context window.
- what is gpt-5.4 nano's context window
- GPT-5.4 nano retains a 400K context window despite the nano tier, so it can still read a large file it was handed.
- can i use gpt-5.4 nano as my main coding model
- No. GPT-5.4 nano is not a build-agent model. It drops invariants on multi-file work and will churn through Atlas permission prompts. Only pin it as `model` if you are doing read-only exploration.
- how do i set gpt-5.4 nano as atlas's small model
- Add `"small_model": "openai/gpt-5.4-nano"` to atlas.json. Set `OPENAI_API_KEY` or run `atlas login` first, then confirm with `atlas models openai`.
- does gpt-5.4 nano support tool calling
- Yes. GPT-5.4 nano supports both reasoning and tool calling, which is what lets it run background subagent work in Atlas rather than only rephrasing text.
- what does atlas use a cheap model like gpt-5.4 nano for
- Atlas makes high-volume, low-stakes calls constantly: session titles, summaries, and classification. GPT-5.4 nano at $0.20 / $1.25 per Mtok is built for exactly that traffic.
Try Atlas in your terminal
The terminal-native AI coding agent. Free core, single binary.
Install AtlasRelated guides
Atlas for Gin in 2026
Atlas is a terminal-native AI coding agent for Gin in 2026. It reads router groups and binding tags, then runs go test ./... -race behind a permission prompt.
Atlas for Python in 2026
Atlas is a terminal-native AI coding agent for Python in 2026. Run it in a repo with a pyproject.toml or requirements.txt and review every diff before it lands.
Atlas for PowerShell: A Terminal-Native AI Coding Agent for Modules and Pester in 2026
Atlas is a terminal-native AI coding agent for PowerShell in 2026. It reads your .psm1 module and .psd1 manifest, adds SupportsShouldProcess, and runs Invoke-Pester behind a prompt.
Atlas vs Qodo: Choosing Your AI Coding Agent in 2026
Comparing Atlas, the terminal-native AI coding agent, with Qodo 2.0, the multi-agent PR reviewer, for developers in 2026. Evaluate features, pricing, and workflow.
Diagnose a Hanging or Long-Running Command with Atlas in 2026
How to diagnose a hanging command with Atlas in 2026: the bash tool races every command against a timeout and tells you whether it is slow or blocked on input.
Migrate a Deprecated API Across Every Callsite with Atlas (2026 Workflow)
How to migrate a deprecated API across every callsite with Atlas in 2026: the lsp tool's findReferences enumerates callers, todowrite tracks them, apply_patch migrates each one.
Debug a Single Failing Test with Atlas in 2026
How to debug one failing test with Atlas in 2026: run it in isolation with bash, walk the call graph with the lsp tool, and fix the code, not the assertion.
Plan a Multi-File Change Before Editing with Atlas in 2026
How to plan a multi-file change with Atlas in 2026: the plan agent denies all edit tools, you research with codebase_search and lsp, then plan_exit hands off.