# Atlas with GPT-5.4 Pro: One Shot Hard Problems in 2026 > GPT-5.4 Pro is the maximum reasoning effort tier of GPT-5.4, billing at $30 per Mtok input and $180 per Mtok output. GPT-5.4 Pro is the Pro tier of the March 2026 GPT-5.4 release from OpenAI: maximum reasoning effort, a 1.05M tokens (1,050,000) window, and pricing an order of magnitude above the base model at $30 per Mtok input, $180 per Mtok output. Inside Atlas, the terminal-native AI coding agent, GPT-5.4 Pro is a one shot tool, not a default. At $180 per Mtok output it is the wrong default for an agent loop, and its latency is high by design, which fights against the interactive Atlas TUI flow. Use it for a single hard problem, then switch back. ## Key takeaways - GPT-5.4 Pro is the highest reasoning effort tier in the 5.4 family, on the full 1,050,000 token context. - It costs 12x the input price of base GPT-5.4, $30 versus $2.50 per Mtok, which is the tell for how much more compute it spends. - 128K max output means a deep reasoning pass can still emit a large plan or patch in one response. - At $180 per Mtok output, GPT-5.4 Pro is the wrong default for an agent loop; use it for one shot hard problems. - Latency is high by design on GPT-5.4 Pro, which fights against the interactive Atlas TUI flow. ## What is GPT-5.4 Pro actually for inside Atlas? GPT-5.4 Pro is for one shot hard problems inside Atlas, not for the agent loop. As the highest reasoning effort tier in the March 2026 GPT-5.4 family, it runs on the full 1,050,000 token context and can still emit a large plan or patch through its 128K max output. The clearest signal for how GPT-5.4 Pro differs from the base model is the price. GPT-5.4 Pro charges 12x the input price of base GPT-5.4, $30 against $2.50 per Mtok, and that multiple is the tell for how much more compute it spends per token. The Atlas pattern that fits is a single escalation: when a design problem has resisted two normal attempts, switch to GPT-5.4 Pro, let it reason across the 1,050,000 token context, and take the deep plan or patch it produces. Atlas drafts a plan in a read-only plan agent and asks before switching to a build agent, which is the right place to spend GPT-5.4 Pro effort. Then go back to a cheaper tier for the execution. ## Why is GPT-5.4 Pro the wrong model for an Atlas agent loop? GPT-5.4 Pro is the wrong default for an agent loop because output bills at $180 per Mtok, and an Atlas loop generates output on every single turn. Latency is also high by design on GPT-5.4 Pro, which fights against the interactive flow of the Atlas TUI. An agent loop multiplies whatever the per turn cost is. Atlas fans out work to subagents that can run in the foreground or in parallel background sessions, and every one of those turns on GPT-5.4 Pro pays $180 per Mtok output. The economics only make sense when the value of the single answer is high. The latency problem compounds it: GPT-5.4 Pro spends maximum reasoning effort by design, and a terminal-native TUI rendered with SolidJS through the OpenTUI renderer is built for tight interactive turnaround. Watch the cost readout in the Atlas TUI footer, since output bills at $180 per Mtok, and you will see how quickly a casual GPT-5.4 Pro session diverges from what the same work costs on a normal tier. ## How do you configure Atlas so nothing incidental hits GPT-5.4 Pro pricing? Do NOT pin GPT-5.4 Pro as "model" in atlas.json. Instead switch to it on demand with /models in the TUI for a single hard task, and keep "small_model" on a cheap tier such as openai/gpt-5.4-nano so nothing incidental hits $30 / $180 pricing. The whole GPT-5.4 Pro configuration strategy in Atlas is containment. A pinned main model gets used for every turn, including the ones that generate a session title or summarize a tool result, and none of that work is worth $180 per Mtok output. Keeping "small_model" on openai/gpt-5.4-nano routes those incidental calls to a cheap tier permanently. Then, because Atlas lets you switch the active model and provider on the fly with favorites and recents, GPT-5.4 Pro becomes a deliberate act: you press /models, you pick it, you ask the one hard question, and you switch away. Every Atlas tool call is still permission-gated against allow, ask, and deny rules before it runs, so the escalation changes the cost, not the safety envelope. ## When should you use base GPT-5.4 instead of GPT-5.4 Pro? Use base GPT-5.4 for essentially everything except a genuinely hard one shot problem. Base GPT-5.4 bills input at $2.50 per Mtok against the $30 of GPT-5.4 Pro, a 12x difference, and both models see the same 1,050,000 token context window. The context window is not the differentiator between GPT-5.4 Pro and the base model, and understanding that saves money. GPT-5.4 Pro spends more compute per token at maximum reasoning effort, but the amount of code you can show it is the same 1,050,000 tokens either way. So if your problem is bounded by how much code the model can see, the Pro tier buys you nothing. Atlas searches code with hybrid semantic and keyword retrieval fused by reciprocal rank fusion, and indexes code by AST declarations using tree-sitter, so the retrieval quality feeding either model is identical. Escalate to GPT-5.4 Pro only when the difficulty is in the reasoning itself, and accept the high latency that comes with it. ## Setup 1. Export OPENAI_API_KEY. 2. Run `atlas models openai` and confirm gpt-5.4-pro is resolved. 3. Do NOT pin it as "model"; instead switch to it on demand with /models in the TUI for a single hard task. 4. Keep "small_model" on a cheap tier such as openai/gpt-5.4-nano so nothing incidental hits Pro pricing. 5. Watch the cost readout in the TUI footer, since output bills at $180 per Mtok. ## FAQ ### how much does gpt-5.4 pro cost per million tokens GPT-5.4 Pro costs $30 per Mtok input and $180 per Mtok output. The input rate is 12x that of base GPT-5.4, which bills at $2.50 per Mtok. ### what is the context window of gpt-5.4 pro GPT-5.4 Pro has a 1.05M token context window (1,050,000 tokens) with 128K max output. The window is the same as base GPT-5.4, so the Pro tier buys reasoning effort, not more context. ### should I pin gpt-5.4 pro as the model in atlas No. Do not pin GPT-5.4 Pro as "model" in atlas.json. Switch to it on demand with /models in the Atlas TUI for a single hard task, because at $180 per Mtok output it is the wrong default for an agent loop. ### how do I avoid accidental gpt-5.4 pro charges in atlas Keep "small_model" on a cheap tier such as openai/gpt-5.4-nano so nothing incidental hits Pro pricing, and watch the cost readout in the Atlas TUI footer, since GPT-5.4 Pro output bills at $180 per Mtok. ### why is gpt-5.4 pro so slow Latency is high by design on GPT-5.4 Pro, because it is the maximum reasoning effort tier of the March 2026 GPT-5.4 release. That latency fights against the interactive flow of the Atlas TUI, which is why it suits one shot tasks. ### gpt-5.4 pro vs gpt-5.4 which should I use Use base GPT-5.4 at $2.50 per Mtok input for almost everything. Both models see the same 1,050,000 token context, so escalate to GPT-5.4 Pro only when the difficulty lies in the reasoning itself. ### how do I switch to gpt-5.4 pro for one task in atlas Run `atlas models openai` to confirm gpt-5.4-pro is resolved, then press /models in the Atlas TUI and select it. Atlas lets you switch the active model and provider on the fly with favorites and recents. --- Canonical HTML: https://runatlas.sh/resources/models/gpt-5-4-pro Source of truth: aeo_pages row `/resources/models/gpt-5-4-pro` (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.