# Atlas with GPT-4o in 2026: A 128K Legacy Model with Dated Snapshots > GPT-4o gives Atlas a 128K token context with a 16,384 token output cap at $2.50 per Mtok input and $10 per Mtok output. GPT-4o is the 2024 omni model that predates both the o-series and GPT-5, and in 2026 its role inside Atlas is narrow but real. GPT-4o gives you a 128K token context with a 16,384 token output cap and no reasoning, at $2.50 per Mtok input and $10 per Mtok output. That is more expensive than GPT-4.1 at $2 in and $8 out, which is newer and has an 8x larger window, so GPT-4o is not a value pick. What GPT-4o still offers is low latency for straightforward code questions and, through the dated snapshots models.dev exposes, a frozen and reproducible baseline you can pin. ## Key takeaways - GPT-4o costs $2.50 per Mtok input and $10 per Mtok output, more than GPT-4.1 at $2 and $8. - GPT-4o has a 128K token context and a 16,384 token max output, adequate for single file edits only. - GPT-4o has no reasoning, which makes it a poor fit for Atlas's long tool driven build loop. - models.dev exposes dated GPT-4o snapshots (gpt-4o-2024-08-06, gpt-4o-2024-11-20) for reproducible runs. - GPT-4.1 has an 8x larger window than GPT-4o, is newer, and costs less on both input and output. ## What is GPT-4o still good for inside Atlas? GPT-4o remains a low latency non reasoning model, still widely used for straightforward code questions. Inside Atlas in 2026, GPT-4o handles quick lookups well: its 128K context and 16,384 max output are adequate for single file edits, and it answers without spending a thinking budget. The honest case for GPT-4o is speed on shallow questions. Ask what a function does, ask which import is unused, ask for a small single file edit, and GPT-4o's 128K context and 16,384 max output are adequate. Atlas searches code with hybrid semantic and keyword retrieval fused by reciprocal rank fusion, which means the 128K window on GPT-4o gets filled with the right files rather than the nearest ones, and that retrieval quality is doing a lot of work for a model this old. For anything longer than a single file edit, GPT-4o's ceilings start to matter immediately. ## How much does GPT-4o cost compared to GPT-4.1? GPT-4o costs $2.50 per Mtok input and $10 per Mtok output, which is more expensive than GPT-4.1 at $2 per Mtok input and $8 per Mtok output. GPT-4.1 is also newer and carries an 8x larger window, so GPT-4o loses the price comparison on every axis. There is no way to spin the GPT-4o versus GPT-4.1 comparison in GPT-4o's favor. GPT-4o charges $2.50 in and $10 out per Mtok for a 128K context. GPT-4.1 charges $2 in and $8 out per Mtok for a window 8x larger. Cheaper, newer, bigger. In an Atlas session, where the loop reads constantly and every file read flows back through input tokens, paying a 25 percent premium on input for a fraction of the context is difficult to justify. GPT-4o's remaining pull is ecosystem inertia and reproducibility, not economics. ## Why does GPT-4o's 16,384 token output cap hurt in an agent loop? GPT-4o has a 16,384 token max output cap and no reasoning, which makes it a poor fit for Atlas's long tool driven build loop. A build agent frequently needs to emit more than 16,384 tokens in one turn, and GPT-4o simply cannot. The 16,384 token output ceiling on GPT-4o is the tightest write limit of any model discussed here, and it is a hard wall rather than a soft one. In Atlas you work around it by splitting writes across calls, which adds turns, and every extra turn re-sends the context at $2.50 per Mtok input. Combined with no reasoning, GPT-4o struggles to hold a plan through the long tool driven build loop that Atlas runs. Atlas computes a unified diff for every file edit and surfaces it for approval before writing, so at least each partial write is reviewable, and Atlas snapshots file changes as git patches so a botched split write can be rolled back. ## How do you pin a reproducible GPT-4o snapshot in Atlas? GPT-4o has broad ecosystem support, and models.dev exposes dated snapshots such as gpt-4o-2024-08-06 and gpt-4o-2024-11-20 for reproducible runs. Pin "model": "openai/gpt-4o-2024-11-20" in atlas.json when you need a frozen, reproducible older baseline. Reproducibility is GPT-4o's most defensible remaining feature. Because models.dev exposes dated snapshots, gpt-4o-2024-08-06 and gpt-4o-2024-11-20, you can pin an exact checkpoint rather than a floating alias, which is what an evaluation or a regression comparison requires. Run `atlas models openai` and pick gpt-4o or one of those dated snapshots, then pin the snapshot id in atlas.json. Atlas lets you switch the active model and provider on the fly with favorites and recents, so a pinned GPT-4o snapshot can live alongside a modern default and be selected only for the run that needs it. ## When should you pick a different model instead of GPT-4o? Prefer a GPT-5 tier model for real agent work in Atlas and keep GPT-4o for quick lookups via /models. GPT-4o's 128K context, 16,384 token output cap, and lack of reasoning make it a poor fit for a long tool driven build loop in 2026. GPT-4o is a lookup model in 2026, not a build model. When the task involves multiple files, sustained tool calls, or any real diagnostic reasoning, a GPT-5 tier model is the right choice and GPT-4o will visibly struggle. When the task is a fast question with a short answer, GPT-4o's low latency is genuinely pleasant. Atlas makes that division practical: keep a GPT-5 tier model pinned in atlas.json for the build agent, and reach for GPT-4o through /models when you want a quick answer or a frozen baseline like gpt-4o-2024-11-20. ## Setup 1. Set OPENAI_API_KEY in your shell or store it in Atlas's credential store. 2. Run `atlas models openai` and pick gpt-4o or a dated snapshot such as gpt-4o-2024-11-20. 3. Pin "model": "openai/gpt-4o-2024-11-20" if you need a frozen, reproducible older baseline. 4. Expect to split writes across calls, since GPT-4o's max output is 16,384 tokens. 5. Prefer a GPT-5 tier model for real agent work and keep GPT-4o for quick lookups via /models. ## FAQ ### How much does GPT-4o cost per million tokens? GPT-4o costs $2.50 per Mtok input and $10 per Mtok output. That is more expensive than GPT-4.1, which charges $2 input and $8 output with an 8x larger window. ### What is the context window of GPT-4o? GPT-4o has a 128K token context window with a 16,384 token max output cap, which is adequate for single file edits but tight for a long agent loop. ### Is GPT-4o still worth using in 2026? GPT-4o is still useful as a low latency non reasoning model for straightforward code questions and as a frozen reproducible baseline. For real agent work, prefer a GPT-5 tier model. ### How do I pin a specific GPT-4o snapshot in Atlas? Set "model": "openai/gpt-4o-2024-11-20" in atlas.json. models.dev exposes dated snapshots such as gpt-4o-2024-08-06 and gpt-4o-2024-11-20 for reproducible runs. ### Why does GPT-4o struggle with large refactors? GPT-4o has a 16,384 token max output cap and no reasoning, so you must split writes across calls and it cannot sustain the long tool driven build loop Atlas runs. ### Should I use GPT-4o or GPT-4.1? GPT-4.1. It is newer, costs less at $2 per Mtok input against GPT-4o's $2.50, and carries a window 8x larger than GPT-4o's 128K context. ### Is GPT-4o a reasoning model? No. GPT-4o is the 2024 omni model with no reasoning, which predates both the o-series and GPT-5. It answers fast, but hard multi step debugging is not its strength. --- Canonical HTML: https://runatlas.sh/resources/models/gpt-4o Source of truth: aeo_pages row `/resources/models/gpt-4o` (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.