# Atlas with GPT-5.3 Chat: The small_model Slot Pick for 2026 > GPT-5.3 Chat is a non reasoning model with a 128K tokens window and a 16,384 token output cap, at $1.75 per Mtok input. GPT-5.3 Chat is the chat tuned snapshot of GPT-5.3, exposed as gpt-5.3-chat-latest. Unlike the rest of the 5.3 family it is a non reasoning model, with a 128K tokens window and a 16,384 token output cap, priced at $1.75 per Mtok input, $14 per Mtok output. Inside Atlas, the terminal-native AI coding agent, GPT-5.3 Chat is a small_model pick rather than a main agent. Because it does not spend a thinking budget, it answers immediately, which is what you want for session titles, summaries, and conversational questions over a handful of files. ## Key takeaways - GPT-5.3 Chat is non reasoning, so it answers immediately instead of spending a thinking budget, which is what the Atlas small_model slot wants. - The 128K tokens context is enough for conversational Q&A over a handful of files. - Pricing is $1.75 per Mtok input and $14 per Mtok output, the same as the full GPT-5.3 Codex, without the reasoning latency. - Max output is 16,384 tokens, roughly one eighth of the 128K ceiling on GPT-5.3 Codex, so it cannot emit large patches. - No reasoning trace makes GPT-5.3 Chat a poor main agent for multi step refactors. ## What makes GPT-5.3 Chat different from the rest of the GPT-5.3 family? GPT-5.3 Chat is the only non reasoning member of the GPT-5.3 family. Exposed as gpt-5.3-chat-latest, it answers immediately instead of spending a thinking budget, and it carries a 128K tokens window with a 16,384 token output cap rather than the 128K ceiling of GPT-5.3 Codex. The absence of a reasoning trace is the defining property of GPT-5.3 Chat, and it cuts both ways. On the upside, GPT-5.3 Chat returns an answer with no thinking latency at the same $1.75 per Mtok input and $14 per Mtok output pricing as the full GPT-5.3 Codex, so you pay the same rate and wait less. On the downside, no reasoning trace makes GPT-5.3 Chat a poor main agent for multi step refactors, where the model needs to hold a plan across many tool calls. Atlas drafts a plan in a read-only plan agent and asks before switching to a build agent, and that planning work wants a reasoning model, not GPT-5.3 Chat. ## Why does GPT-5.3 Chat belong in the Atlas small_model slot? GPT-5.3 Chat belongs in the Atlas small_model slot because it is non reasoning and therefore fast. Point the cheap slot at it with "small_model": "openai/gpt-5.3-chat-latest" in atlas.json so titles and summaries stay fast while the main build loop runs on a reasoning tier. Atlas generates a steady stream of small, low stakes calls: naming a session, summarizing a tool result, condensing a long output. None of those need a thinking budget, and making them wait on one is pure latency for no gain. GPT-5.3 Chat handles them immediately. Meanwhile you leave "model" on a reasoning tier such as openai/gpt-5.3-codex for the actual build loop, so the work that needs a chain of thought gets one. Atlas fans out work to subagents that can run in the foreground or in parallel background sessions, and routing the incidental portion of that traffic to GPT-5.3 Chat keeps the session responsive. Confirm both slots with `atlas models openai` after editing atlas.json. ## What does the 16,384 token output cap of GPT-5.3 Chat prevent? GPT-5.3 Chat caps output at 16,384 tokens, roughly one eighth of the 128K ceiling on GPT-5.3 Codex, so it cannot emit large patches. A wide refactor that needs to rewrite several files will simply not fit in a single GPT-5.3 Chat response. The 16,384 token output cap is the hard boundary that keeps GPT-5.3 Chat out of the main build loop. Atlas computes a unified diff for every file edit and surfaces it for approval before writing, and a diff bounded by 16,384 tokens is a small diff. That is fine for a targeted single function change and useless for a service wide rename. The read side is more generous: the 128K tokens context of GPT-5.3 Chat is enough for conversational Q&A over a handful of files, which is exactly the shape of question a developer asks between edits. Read a few files, ask GPT-5.3 Chat what they do, then switch to a reasoning tier to change them. ## When should you pick GPT-5.3 Codex instead of GPT-5.3 Chat? Pick GPT-5.3 Codex whenever the task is a multi step refactor. GPT-5.3 Chat has no reasoning trace and a 16,384 token output cap, while GPT-5.3 Codex is a reasoning model with a 128K output ceiling, at the identical $1.75 per Mtok input and $14 per Mtok output price. Because GPT-5.3 Chat and GPT-5.3 Codex share the same per token price, the choice between them is purely about capability, not budget. GPT-5.3 Codex is the reasoning tier that should hold the "model" slot in atlas.json for the actual build loop. GPT-5.3 Chat is the cheap slot pick, where its immediate answers and lack of thinking latency are an advantage. Atlas lets you switch the active model and provider on the fly with favorites and recents, so you can also flip to GPT-5.3 Chat with /models for a quick conversational pass over a few files and flip back. Every Atlas tool call is permission-gated against allow, ask, and deny rules before it runs, whichever of the two is driving. ## Setup 1. Set OPENAI_API_KEY in your environment. 2. Run `atlas models openai` and look for gpt-5.3-chat-latest. 3. Point the cheap slot at it: "small_model": "openai/gpt-5.3-chat-latest" so titles and summaries stay fast. 4. Leave "model" on a reasoning tier such as openai/gpt-5.3-codex for the actual build loop. 5. Confirm both slots with `atlas models openai` after editing atlas.json. ## FAQ ### what is gpt-5.3-chat-latest used for GPT-5.3 Chat, exposed as gpt-5.3-chat-latest, is the non reasoning chat tuned snapshot of GPT-5.3. In Atlas it suits the small_model slot, where its immediate answers keep session titles and summaries fast. ### how much does gpt-5.3 chat cost per million tokens GPT-5.3 Chat costs $1.75 per Mtok input and $14 per Mtok output, the same pricing as the full GPT-5.3 Codex. The difference is that GPT-5.3 Chat has no reasoning latency. ### what is the max output of gpt-5.3 chat GPT-5.3 Chat caps output at 16,384 tokens, roughly one eighth of the 128K ceiling on GPT-5.3 Codex. That cap means it cannot emit large patches in a single Atlas edit. ### how do I set gpt-5.3 chat as the atlas small_model Set OPENAI_API_KEY, run `atlas models openai` and look for gpt-5.3-chat-latest, then set "small_model": "openai/gpt-5.3-chat-latest" in atlas.json. Leave "model" on a reasoning tier such as openai/gpt-5.3-codex. ### is gpt-5.3 chat a reasoning model No. GPT-5.3 Chat is the only non reasoning member of the GPT-5.3 family. It answers immediately rather than spending a thinking budget, which makes it a poor main agent for multi step refactors. ### what is the context window of gpt-5.3 chat GPT-5.3 Chat has a 128K tokens context window. That is enough for conversational Q&A over a handful of files, though not for a large repo sweep. ### gpt-5.3 chat vs gpt-5.3 codex for atlas GPT-5.3 Chat and GPT-5.3 Codex cost the same, $1.75 / $14 per Mtok, so pick on capability. Use GPT-5.3 Codex as the "model" for the build loop, and GPT-5.3 Chat as the "small_model" for fast incidental calls. --- Canonical HTML: https://runatlas.sh/resources/models/gpt-5-3-chat Source of truth: aeo_pages row `/resources/models/gpt-5-3-chat` (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.