The March 2026 GPT-5.4 release brought the 1.05M context window (1,050,000 tokens) down to $2.50 / $15 per Mtok (input / output). In Atlas, GPT-5.4 sits between GPT-5.6 Luna and full GPT-5.6 as the balanced default for day-to-day sessions: half the input cost of GPT-5.6 with the identical 1,050,000 token window and 128K output tokens. The tradeoff is that GPT-5.4 is superseded by GPT-5.6 on raw capability.
What does GPT-5.4 cost and what do you get for it?
GPT-5.4 costs $2.50 / $15 per Mtok (input / output), half the input cost of GPT-5.6 at $5, with the identical 1,050,000 token window. The March 2026 release is what pushed the full GPT-5 context down to a price a developer can leave running all day.
The value case for GPT-5.4 is that nothing about the context budget changes when the price halves. A long Atlas session re-sends its transcript and its retrieved context every turn, and input is where that bill accumulates. At $2.50 per Mtok, GPT-5.4 halves the dominant cost line without asking you to trim the working set. The 128K output tokens is the other half of the story: GPT-5.4 can emit large multi-file patches in a single Atlas edit turn, so a big refactor does not fracture into continuation turns.
Why does settled tool-calling behavior matter for GPT-5.4 in Atlas?
GPT-5.4 is mature enough to have settled tool-calling behavior, which matters directly for Atlas's permission gate. Every Atlas tool call is permission-gated against allow, ask, and deny rules, so fewer malformed calls means fewer wasted approvals and fewer turns billed at $2.50 per Mtok input.
Maturity is an underrated model property in an agent. A newer model with better raw reasoning can still be worse in practice if its tool calls drift out of schema, because Atlas stops and asks before running anything. Each rejected call costs a full round trip. GPT-5.4 has had enough time in production for its tool-calling behavior to settle, and in a permission-gated loop that stability converts directly into fewer interruptions at the approval prompt and less money spent re-sending the same context.
GPT-5.4 vs GPT-5.6: which should drive your Atlas session?
GPT-5.4 costs $2.50 / $15 per Mtok against GPT-5.6's $5 / $30, and both carry the same 1,050,000 token context window. GPT-5.4 is superseded by GPT-5.6 on raw capability, so you pay half but give up the newest reasoning improvements.
That is a clean, honest trade and it usually resolves in GPT-5.4's favor for routine work. Most Atlas turns are not frontier reasoning problems: they are retrieval, a targeted edit, a test run, a fix. GPT-5.4 does those at half the input rate with the same window. Save GPT-5.6 for the turns where the newest reasoning actually decides the outcome. Atlas has a shortcut for exactly this pattern: use model.cycle_recent to flip between GPT-5.4 and a frontier model without reopening the dialog.
How do you flip between GPT-5.4 and a frontier model mid-session?
Use model.cycle_recent to flip between GPT-5.4 and a frontier model without reopening the dialog. Atlas lets you switch the active model and provider on the fly with favorites and recents, so a session pinned to GPT-5.4 at $2.50 / $15 per Mtok can escalate for one hard turn and drop straight back.
The workflow that makes GPT-5.4 the balanced default depends on cheap escalation. Pin "model": "openai/gpt-5.4" in atlas.json so every session starts on the affordable side. When you hit a turn that genuinely needs the newest reasoning, cycle up, take the answer, and cycle back. Because Atlas snapshots file changes as git patches so edits can be diffed and rolled back, an escalation that produces a bad diff costs you nothing permanent either way.
How do you set GPT-5.4 as the default model in Atlas?
Set OPENAI_API_KEY, or authenticate via atlas login, then verify with atlas models openai. Pick GPT-5.4 in the TUI via /models, or pin "model": "openai/gpt-5.4" in atlas.json so every new Atlas session starts on the $2.50 / $15 per Mtok tier.
Registry verification is the step people skip and regret. Running atlas models openai confirms that GPT-5.4 actually resolves with the key you supplied, before you build a session around it. After that, config and session control are independent: the atlas.json pin governs where new sessions start, while /models changes the model for the session in front of you. Both work, and neither locks you in, because Atlas keeps favorites and recents across sessions.
Setup
- 01Set `OPENAI_API_KEY`, or authenticate via `atlas login`.
- 02Verify with `atlas models openai`.
- 03Pick it in the TUI via `/models`, or pin `"model": "openai/gpt-5.4"` in atlas.json.
- 04Use `model.cycle_recent` to flip between GPT-5.4 and a frontier model without reopening the dialog.
- 05Escalate to GPT-5.6 only on turns where the newest reasoning improvements decide the outcome.
Frequently asked questions
- how much does gpt-5.4 cost per million tokens
- GPT-5.4 costs $2.50 / $15 per Mtok (input / output), which is half the input cost of GPT-5.6 at $5 per Mtok.
- what is gpt-5.4's context window
- GPT-5.4 has a 1.05M token context window, exactly 1,050,000 tokens, identical to GPT-5.6. Max output is 128K tokens.
- gpt-5.4 vs gpt-5.6 which is better for coding
- GPT-5.4 is superseded by GPT-5.6 on raw capability, so you pay half ($2.50 / $15 versus $5 / $30 per Mtok) but give up the newest reasoning improvements. Both share the same 1,050,000 token window.
- how do i make gpt-5.4 the default model in atlas
- Pin `"model": "openai/gpt-5.4"` in atlas.json after setting `OPENAI_API_KEY` or authenticating via `atlas login`. Verify it resolved with `atlas models openai`.
- how do i switch models quickly in atlas
- Use `model.cycle_recent` to flip between GPT-5.4 and a frontier model without reopening the dialog. Atlas also keeps favorites and recents so switching is on the fly.
- why do fewer malformed tool calls matter in atlas
- Every Atlas tool call is permission-gated against allow, ask, and deny rules, so a malformed call costs a wasted approval and a re-sent turn. GPT-5.4's settled tool-calling behavior reduces both.
- can gpt-5.4 handle a large multi-file refactor
- Yes. GPT-5.4 has 128K output tokens, so it can emit large multi-file patches in a single Atlas edit turn, and the 1,050,000 token window holds the context to plan them.
Try Atlas in your terminal
The terminal-native AI coding agent. Free core, single binary.
Install AtlasRelated guides
Atlas for Assembly: Registers, Calling Conventions, and nasm in 2026
Atlas is a terminal-native AI coding agent for Assembly in 2026. It reads .asm and .S sources, tracks System V and AAPCS64 calling conventions, and assembles with nasm behind a prompt.
Atlas vs. Goose: Choosing Your AI Coding Agent in 2026
Compare Atlas and Goose for 2026. Atlas offers terminal-native TUI and code-specialized features. Goose provides shareable Recipes and 70+ MCP extensions for general agentic workflows.
Atlas vs Trae: AI Coding Agents in 2026
Atlas and Trae offer distinct AI coding agent experiences in 2026. Atlas provides terminal-native control and local data privacy, while Trae offers a full IDE with SOLO Builder and cloud tasks.
Atlas for C# in 2026
Atlas is a terminal-native AI coding agent for C# and the .NET SDK in 2026. Run it in a solution with a .csproj or .sln and approve every diff before dotnet build.
Atlas for Phoenix in 2026
Atlas is a terminal-native AI coding agent for Phoenix in 2026. It reads contexts, LiveView modules, and Ecto changesets, then runs mix test behind a prompt.
Atlas for Quarkus in 2026
Atlas is a terminal-native AI coding agent for Quarkus in 2026. It reads CDI beans and JAX-RS resources, then runs ./mvnw test behind a permission prompt.
Run the Test Suite and Triage the Failures with Atlas in 2026
How to triage a failing test suite with Atlas in 2026: bash truncates at 2000 lines or 50 KB and saves the full log, then grep groups failures by root cause.
Atlas vs Jules: Terminal AI Coding Agents in 2026
Atlas and Jules comparison for 2026. Atlas offers terminal-native TUI and local code indexing. Jules provides cloud VMs and a Critic agent for PR review.