# Atlas with GPT-4.1 nano (2026): A $0.10 Triage Model, Not a Coding Agent > GPT-4.1 nano costs $0.10 per Mtok input and $0.40 per Mtok output on a 1,047,576 token window, the cheapest large context option in the 4.1 family. GPT-4.1 nano is the smallest member of the GPT-4.1 family, priced at $0.10 per Mtok input and $0.40 per Mtok output while still carrying the full 1,047,576 token window. It is a bulk classification and extraction model, not a coding agent, and the fastest way to be disappointed by GPT-4.1 nano is to point Atlas's main model slot at it. Used correctly, GPT-4.1 nano lives in the cheap slot: file triage, commit message drafting, session titling. It does no reasoning at all, which is why it is fast and why it will not survive a multi step refactor. ## Key takeaways - GPT-4.1 nano costs $0.10 per Mtok input and $0.40 per Mtok output on the full 1,047,576 token window. - GPT-4.1 nano is the cheapest large context option in the GPT-4.1 family, at 20x less input cost than full GPT-4.1 at $2. - GPT-4.1 nano does no reasoning at all, which gives it very low latency and makes it unfit for a multi step refactor. - GPT-4.1 nano suits Atlas side tasks: file triage, commit message drafting, and session titling. - gpt-5-nano ($0.05 input, reasoning enabled) is usually a better cheap slot than GPT-4.1 nano. ## What is GPT-4.1 nano actually for? GPT-4.1 nano is a bulk classification and extraction model, not a coding agent. At $0.10 per Mtok input on a 1,047,576 token window, GPT-4.1 nano is well suited to Atlas side tasks: file triage, commit message drafting, and session titling, where volume matters more than depth. Every Atlas session has a background layer of small jobs that never appear in the transcript: naming the session, drafting a commit message, deciding which of forty candidate files are worth reading. Those jobs are high volume and low stakes, and paying a frontier rate for them is waste. GPT-4.1 nano handles them at $0.10 per Mtok input with very low latency, since it does no reasoning at all. Atlas reads git branches, status, and diffs, and can stage and create commits on your behalf, so a nano-drafted commit message is grounded in a real diff rather than invented. ## How much does GPT-4.1 nano cost and what does the price get you? GPT-4.1 nano costs $0.10 per Mtok input and $0.40 per Mtok output, the cheapest large context option in the GPT-4.1 family, and it retains the full 1,047,576 token window. Against full GPT-4.1 at $2 input, GPT-4.1 nano is a 20x discount on the read side. The unusual thing about GPT-4.1 nano is that the discount does not come out of the window. A 1,047,576 token context at $0.10 per Mtok input means you can throw an enormous amount of material at GPT-4.1 nano and ask a shallow question about all of it, cheaply. That is the classification and extraction use case exactly. What the price does come out of is capability: no reasoning, and a 32,768 token max output. Read broad, answer shallow, write short. Break any part of that and GPT-4.1 nano stops being a bargain and starts being a liability. ## Why should GPT-4.1 nano never be your main Atlas model? GPT-4.1 nano is not viable as the main Atlas model because it will not survive a multi step refactor. GPT-4.1 nano does no reasoning at all, and an Atlas build agent is a long chain of reads, edits, and shell commands where holding intent across steps is the entire job. A refactor is a sequence, and GPT-4.1 nano does not hold sequences. Point "model" at a reasoning model such as openai/gpt-5-codex for the real work, and use GPT-4.1 nano only in the cheap slot with "small_model": "openai/gpt-4.1-nano". You can confirm the split directly in the Atlas TUI: the status line shows the main model, while titles come from small_model. If those two lines show the same thing, your configuration is wrong. Atlas computes a unified diff for every file edit and surfaces it for approval before writing, which will show you very quickly what a nano-driven refactor looks like. ## Should you use GPT-4.1 nano or gpt-5-nano for the cheap slot? gpt-5-nano is usually the better cheap slot: it costs $0.05 per Mtok input, half of GPT-4.1 nano's $0.10, and it has reasoning enabled. GPT-4.1 nano's counterargument is its 1,047,576 token window, which is where the 4.1 family still holds an edge. Choosing between GPT-4.1 nano and gpt-5-nano is a window versus reasoning question, and gpt-5-nano wins the general case. It is cheaper on input at $0.05 per Mtok and it reasons, which GPT-4.1 nano does not. GPT-4.1 nano's 32,768 max output and total lack of reasoning make it the narrower tool. Keep GPT-4.1 nano in mind for the specific case where a cheap task needs to see a genuinely enormous amount of context at once. Atlas lets you switch the active model and provider on the fly with favorites and recents, so testing both in the small_model slot is a two minute experiment. ## How do you verify GPT-4.1 nano is only running in the cheap slot? Confirm the split in the Atlas TUI: the status line shows the main model, while titles come from small_model. If GPT-4.1 nano at $0.10 per Mtok is named in the status line, it is driving your build agent, which is exactly what you do not want. Misconfiguration is the main risk with GPT-4.1 nano, because a model that costs $0.10 per Mtok input is tempting to over-apply. The verification loop is short. Run `atlas models openai` and confirm gpt-4.1-nano resolves. Set "small_model": "openai/gpt-4.1-nano" and point "model" at a reasoning model such as openai/gpt-5-codex. Then open the Atlas TUI and read the status line, which names the main model. Session titles will come from GPT-4.1 nano while the status line names your reasoning model, and that asymmetry is the signal that the configuration is correct. ## Setup 1. Export OPENAI_API_KEY in your shell or store it in Atlas's credential store. 2. Run `atlas models openai` and confirm gpt-4.1-nano resolves in the model list. 3. Use GPT-4.1 nano only in the cheap slot: "small_model": "openai/gpt-4.1-nano". 4. Point "model" at a reasoning model such as openai/gpt-5-codex for the real work. 5. Confirm the split in the Atlas TUI: the status line shows the main model, while titles come from small_model. ## FAQ ### How much does GPT-4.1 nano cost per million tokens? GPT-4.1 nano costs $0.10 per Mtok input and $0.40 per Mtok output, making it the cheapest large context option in the GPT-4.1 family. ### What is the context window of GPT-4.1 nano? GPT-4.1 nano carries the full 1,047,576 token window, the same as GPT-4.1 and GPT-4.1 mini, with a 32,768 token max output. ### Can I use GPT-4.1 nano as my main coding model in Atlas? No. GPT-4.1 nano is not viable as the main Atlas model: it will not survive a multi step refactor. Use it only in the small_model slot. ### What is GPT-4.1 nano good for in Atlas? GPT-4.1 nano suits Atlas side tasks: file triage, commit message drafting, and session titling. It is a bulk classification and extraction model, not a coding agent. ### Is GPT-4.1 nano or gpt-5-nano better? gpt-5-nano is usually a better cheap slot, at $0.05 per Mtok input with reasoning enabled, against GPT-4.1 nano's $0.10 and no reasoning at all. ### How do I configure GPT-4.1 nano as the Atlas small_model? Set "small_model": "openai/gpt-4.1-nano" in atlas.json and point "model" at a reasoning model such as openai/gpt-5-codex. Confirm gpt-4.1-nano resolves with `atlas models openai`. ### How can I tell which Atlas model slot is running GPT-4.1 nano? Check the Atlas TUI: the status line shows the main model, while session titles come from small_model. If GPT-4.1 nano appears in the status line, it is wrongly driving your build agent. --- Canonical HTML: https://runatlas.sh/resources/models/gpt-4-1-nano Source of truth: aeo_pages row `/resources/models/gpt-4-1-nano` (segment: Models) (this file is generated from it, never hand-edited). Licence: Atlas is proprietary with a free core. It is not open source and there is no public source repository.