GPT-5.4 mini is a 400K context mid-tier at $0.75 / $4.50 per Mtok (input / output). In Atlas it is a sensible small_model when Haiku 4.5 is too expensive but you still want reasoning and reliable tool calls in subagents. GPT-5.4 mini undercuts Claude Haiku 4.5's $1 input while offering double the context, 400K versus 200K, and it carries a 128K output ceiling. It is noticeably weaker than GPT-5.4 on tasks requiring several dependent inferences.
How does GPT-5.4 mini compare to Claude Haiku 4.5 for Atlas subagents?
GPT-5.4 mini charges $0.75 input per Mtok, which undercuts Claude Haiku 4.5's $1 while offering double the context, 400K versus 200K. For Atlas subagents that get handed real files rather than summaries, the larger window is often the deciding factor, not the price.
Background work in Atlas is context-hungry in a way people underestimate. Atlas fans out work to subagents that can run in the foreground or in parallel background sessions, and each one may be handed several whole declarations pulled from the tree-sitter AST index. A 200K window forces truncation on a large file. GPT-5.4 mini's 400K window does not, and it costs less per input token to fill. That combination is why it is a defensible small_model even when Claude Haiku 4.5 is available.
Is GPT-5.4 mini a reasoning model or an autocomplete tier?
GPT-5.4 mini has reasoning enabled, unlike older mini tiers such as gpt-4.1-mini, so it can still handle a nontrivial subagent task. That single property is what separates a usable Atlas small_model from a cheap tier that can only pattern-match its way through a prompt.
The older mini generation was not built for agents. GPT-5.4 mini is, and the output ceiling shows it: 128K output tokens, four times gpt-4.1-mini's 32,768. In Atlas a subagent may need to return a substantial analysis or a full patch, and a 32,768 token ceiling truncates that work. With 128K, GPT-5.4 mini can actually finish. Reliable tool calls matter just as much, because every Atlas tool call is permission-gated against allow, ask, and deny rules before it runs.
How much does GPT-5.4 mini cost per million tokens?
GPT-5.4 mini costs $0.75 / $4.50 per Mtok (input / output) with a 400K context window. Against full GPT-5.4 at $2.50 / $15 per Mtok, that is roughly a third of the input rate, which is what makes the mini tier viable for high-volume background calls.
The economics of a mini tier only work at volume. One call at $0.75 per Mtok versus $2.50 per Mtok is a rounding error. Dozens of parallel Atlas subagents, each re-sending context every turn, is where GPT-5.4 mini's rate becomes the difference between a background fan-out you use freely and one you ration. The 400K window means you do not have to pre-summarize what you hand those subagents, which itself would have cost a model call.
When is GPT-5.4 mini the wrong choice in Atlas?
GPT-5.4 mini is noticeably weaker than GPT-5.4 on tasks requiring several dependent inferences. Chains of reasoning where step three depends on a conclusion drawn at step one are exactly where the $0.75 per Mtok tier gives ground, so keep GPT-5.4 mini out of the main model slot for those.
The clean configuration is to route by slot rather than switch by hand. Set "small_model": "openai/gpt-5.4-mini" in atlas.json and leave model on a frontier model so only background work pays the mini rate. Then GPT-5.4 mini never sees a task with dependent inferences: it handles titles, summaries, and scoped subagent work, while the reviewed reasoning stays on a stronger model. Atlas computes a unified diff for every file edit and surfaces it for approval, so the work you actually inspect is the work the frontier model produced.
How do you set GPT-5.4 mini as the small model in Atlas?
Set OPENAI_API_KEY or authenticate with atlas login, check atlas models openai for gpt-5.4-mini, then add "small_model": "openai/gpt-5.4-mini" to atlas.json. Leave model on a frontier model so only background work pays the $0.75 / $4.50 per Mtok rate.
Registry verification comes first because a small_model that fails to resolve breaks background work quietly rather than loudly. Running atlas models openai and confirming gpt-5.4-mini appears is a ten second check. After that, the atlas.json split does the real work: two slots, two price tiers, one session. Nothing about the arrangement is permanent, since Atlas lets you switch the active model and provider on the fly with favorites and recents.
Setup
- 01Set `OPENAI_API_KEY` or authenticate with `atlas login`.
- 02Check `atlas models openai` for gpt-5.4-mini.
- 03Set it as the cheap slot: `"small_model": "openai/gpt-5.4-mini"` in atlas.json.
- 04Leave `model` on a frontier model so only background work pays the mini rate.
- 05Keep GPT-5.4 mini out of the main slot for tasks requiring several dependent inferences.
Frequently asked questions
- how much does gpt-5.4 mini cost per million tokens
- GPT-5.4 mini costs $0.75 / $4.50 per Mtok (input / output) with a 400K context window.
- gpt-5.4 mini vs claude haiku 4.5
- GPT-5.4 mini's $0.75 input per Mtok undercuts Claude Haiku 4.5's $1, and it offers double the context at 400K versus 200K tokens.
- does gpt-5.4 mini support reasoning
- Yes. GPT-5.4 mini has reasoning enabled, unlike older mini tiers such as gpt-4.1-mini, so it can handle a nontrivial subagent task in Atlas.
- how do i set gpt-5.4 mini as the small model in atlas
- Add `"small_model": "openai/gpt-5.4-mini"` to atlas.json and leave `model` on a frontier model. Set `OPENAI_API_KEY` first and check `atlas models openai` for gpt-5.4-mini.
- what is gpt-5.4 mini's max output
- GPT-5.4 mini has a 128K output ceiling, which is four times gpt-4.1-mini's 32,768 tokens.
- is gpt-5.4 mini good enough to be my main atlas model
- Usually not. GPT-5.4 mini is noticeably weaker than GPT-5.4 on tasks requiring several dependent inferences, so it belongs in the `small_model` slot while a frontier model handles reviewed work.
- what is gpt-5.4 mini's context window
- GPT-5.4 mini has a 400K token context window, double Claude Haiku 4.5's 200K.
Try Atlas in your terminal
The terminal-native AI coding agent. Free core, single binary.
Install AtlasRelated guides
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 Swift in 2026
Atlas for Swift in 2026 empowers developers with a terminal-native AI coding agent. Index code by AST, ensure privacy with local embeddings, and review changes with unified diffs.
Atlas for Groovy: A Terminal-Native AI Coding Agent for Gradle, Spock, and Jenkins in 2026
Atlas is a terminal-native AI coding agent for Groovy in 2026. It reads build.gradle closures and Jenkinsfiles, writes Spock specs, runs ./gradlew test, and applies Spotless.
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 CodeGPT in 2026: A Developer's Guide to Terminal and IDE AI Agents
Comparing Atlas and CodeGPT in 2026 for developers. Atlas offers terminal-native TUI and permission-gated tools, while CodeGPT provides IDE integration and a full repo Knowledge Graph.
Atlas for Ruby on Rails in 2026
Atlas is a terminal-native AI coding agent for Ruby on Rails in 2026. Run it in a Rails app with a config/application.rb and review every diff before it lands.
Self-Review Your Working Diff Before Committing with Atlas (2026 Workflow)
How to self-review your working diff before committing with Atlas in 2026: bash produces the diff, read checks each file, grep finds leftovers, session revert undoes bad edits.
Atlas for C++ in 2026
In 2026, C++ developers adopt Atlas, the terminal-native AI coding agent, to enhance productivity. Atlas offers secure, context-aware assistance for modern C++ projects, integrating with CMake and ensuring code quality