GPT-5 Nano is the cheapest model in the entire OpenAI registry: $0.05 per Mtok input and $0.40 per Mtok output, yet it still carries a 400K context and a 128K max output and is a reasoning model. Inside Atlas, GPT-5 Nano belongs in the small_model slot, where it drives session titles, summaries, and cheap subagent fan out. GPT-5 Nano is 25x cheaper than GPT-5 and 5x cheaper than gpt-5-mini on input. It is not a main agent.
How cheap is GPT-5 Nano compared to GPT-5?
GPT-5 Nano costs $0.05 per Mtok input, which is 25x cheaper than GPT-5 and 5x cheaper than gpt-5-mini. Output runs $0.40 per Mtok. GPT-5 Nano is the cheapest model in the entire OpenAI registry, and nothing else in the 5.x line comes close on price.
A 25x input discount changes what is worth automating. Work that would never justify $1.25 per Mtok, such as summarizing every session, generating a title for every conversation, or scanning a hundred files for a pattern, becomes essentially free at $0.05 per Mtok input. GPT-5 Nano is not competing with GPT-5 on quality and was never meant to. GPT-5 Nano competes on the question of whether an operation happens at all, and at that price the answer is usually yes.
Does GPT-5 Nano really have a 400K context window?
GPT-5 Nano carries a 400K token window, which is remarkable at $0.05 per Mtok input and makes it usable for bulk file scanning. Most models this cheap cut context to 32K or less. GPT-5 Nano keeps the full 400K, along with a 128K max output on paper.
The 400K window on GPT-5 Nano is what makes bulk scanning practical inside Atlas. Atlas indexes code by AST declarations using tree-sitter, not blind line windows, so a wide sweep arrives as whole declarations, and GPT-5 Nano can accept an enormous slice of that at $0.05 per Mtok input. Be honest about the limit, though: GPT-5 Nano has a 128K max output on paper, but quality degrades long before you reach it. Use the window to read widely, not to write at length.
Why is GPT-5 Nano the right Atlas small_model?
GPT-5 Nano is an ideal fit for Atlas's small_model slot: session titles, summaries, and cheap subagent fan out. Add "small_model": "openai/gpt-5-nano" to atlas.json and leave "model" on a stronger tier, and incidental calls run at $0.05 per Mtok input instead of full price.
Every Atlas session generates background work you never asked for explicitly. Titles get written, long transcripts get compacted, and subagents get spawned. Atlas fans out work to subagents that can run in the foreground or in parallel background sessions, and paying frontier rates for all of that is pure waste. GPT-5 Nano is a reasoning model at $0.05 per Mtok input, which means the background tier is cheap without being incompetent. Use GPT-5 Nano as the summarizer for long Atlas sessions so compaction costs almost nothing.
Can GPT-5 Nano be the main Atlas build agent?
GPT-5 Nano is not a main agent: it will lose the thread on multi step refactors. Do not route the build agent to GPT-5 Nano, no matter how attractive $0.05 per Mtok input looks. An Atlas build session involves dozens of sequential tool calls, and GPT-5 Nano does not hold a goal across them.
The failure mode is specific and worth naming. GPT-5 Nano will start a refactor, complete two files, and then quietly forget the original constraint on the third. Atlas computes a unified diff for every file edit and surfaces it for approval before writing, so you will catch it, but catching a drifting agent over and over is worse than paying for a competent one. Check the TUI status line to confirm which slot GPT-5 Nano is filling. It should read small_model, never the build agent.
How much does session compaction cost on GPT-5 Nano?
Session compaction on GPT-5 Nano costs almost nothing, because summarization is input heavy and GPT-5 Nano bills $0.05 per Mtok input with $0.40 per Mtok output. Long Atlas sessions produce large transcripts, and compacting them on a frontier model at $1.25 per Mtok input is money burned on housekeeping.
Compaction is the archetypal GPT-5 Nano task: read a lot, write a little, no deep reasoning required. Routing it to GPT-5 Nano through the small_model slot means an Atlas session can run all day, compact repeatedly, and never accumulate a meaningful compaction bill. Atlas snapshots file changes as git patches so edits can be diffed and rolled back, so the real state of your work lives in git, not in the transcript, which is exactly why a cheap summarizer is a safe choice here.
Setup
- 01Set OPENAI_API_KEY so the Atlas openai provider can authenticate.
- 02Run `atlas models openai` and confirm gpt-5-nano resolves.
- 03Add "small_model": "openai/gpt-5-nano" to atlas.json; leave "model" on a stronger tier.
- 04Use it as the summarizer for long Atlas sessions so compaction costs almost nothing.
- 05Do not route the build agent here; check the TUI status line to confirm which slot it is filling.
Frequently asked questions
- What is the cheapest OpenAI model?
- GPT-5 Nano is the cheapest model in the entire OpenAI registry at $0.05 per Mtok input and $0.40 per Mtok output, and it still carries a 400K context window.
- Does GPT-5 Nano have a small context window?
- No. GPT-5 Nano carries a 400K token window, which is remarkable at $0.05 per Mtok input and makes it usable for bulk file scanning.
- Can I use GPT-5 Nano as my main coding model?
- No. GPT-5 Nano is not a main agent: it will lose the thread on multi step refactors. Add it as "small_model": "openai/gpt-5-nano" and leave "model" on a stronger tier.
- How do I set gpt-5-nano as the Atlas small_model?
- Set OPENAI_API_KEY, run `atlas models openai` to confirm gpt-5-nano, then add "small_model": "openai/gpt-5-nano" to atlas.json. Check the TUI status line to confirm which slot it is filling.
- Is GPT-5 Nano a reasoning model?
- Yes. GPT-5 Nano is a reasoning model despite costing only $0.05 per Mtok input, which is 25x cheaper than GPT-5.
- What is the best model for Atlas session summaries?
- GPT-5 Nano. Use it as the summarizer for long Atlas sessions so compaction costs almost nothing at $0.05 per Mtok input and $0.40 per Mtok output.
- Can GPT-5 Nano really emit 128K output tokens?
- GPT-5 Nano has a 128K max output on paper, but quality degrades long before you reach it. Use its 400K window to read widely rather than to generate at length.
Try Atlas in your terminal
The terminal-native AI coding agent. Free core, single binary.
Install AtlasRelated guides
Atlas for Clojure: A Terminal-Native AI Coding Agent for deps.edn and Kaocha in 2026
Atlas is a terminal-native AI coding agent for Clojure in 2026. It reads deps.edn aliases and namespace requires, runs clojure -M:test with Kaocha, and applies cljfmt.
Atlas vs JetBrains Junie in 2026: Terminal Agent vs IDE Debugger-Driving Agent
Atlas vs JetBrains Junie in 2026: Junie drives the IDE debugger and reports 61.6% resolved on SWE-Rebench, bundled from $10/mo. Atlas is a free, open source terminal agent with permission gates.
Atlas vs Graphite: Terminal AI Coding Agents in 2026
Comparing Atlas and Graphite in 2026: Atlas offers a terminal-native AI coding agent with local privacy, while Graphite focuses on stacked PR workflows and GitHub integration.
Atlas for Fastify in 2026
Atlas is a terminal-native AI coding agent for Fastify in 2026. It reads plugin encapsulation and JSON schemas, then runs node --test behind a permission prompt.
Atlas vs Sourcery: Terminal AI Coding Agents in 2026
Compare Atlas and Sourcery for 2026. Atlas offers terminal-native AI coding with permission-gated tools and local embeddings. Sourcery provides Python-first LLM review and rules-based refactoring, starting at
Atlas vs Devin: AI Coding Agents Compared for 2026
Atlas and Devin offer distinct AI coding experiences in 2026. Atlas provides a terminal-native TUI with local control, while Devin is a cloud-managed engineer with SWE-1.7.
Rename a Symbol Across the Repo with Atlas in 2026
How to rename a symbol across a repo with Atlas in 2026: findReferences gets the true reference set, grep catches strings and docs, and edit refuses ambiguous matches.
Atlas for PHP in 2026
Atlas, the terminal-native AI coding agent, empowers PHP developers in 2026 with intelligent code understanding, secure workflows, and direct integration for Composer and PSR standards.