OpenAI o3 is the April 2025 flagship of OpenAI's o-series reasoning line, and inside Atlas it is the model you reach for when a bug needs to be reasoned about rather than pattern matched. OpenAI o3 gives Atlas a 200K token context with a 100K max output, priced at $2 per Mtok input and $8 per Mtok output after OpenAI's price cut. That price cut is the whole story: o3 made deep reasoning cheap enough to leave running in an agent loop, at 7.5x less than o1 on both input ($15) and output ($60). The tradeoff is speed and window size, and Atlas lets you switch the active model and provider on the fly when either one bites.
What is OpenAI o3 best at inside Atlas?
OpenAI o3 is best inside Atlas at root cause debugging, math, and proofs, not fast code emission. Released in April 2025 as the o-series flagship, o3 carries pure reasoning post training on a 200K token context, priced at $2 per Mtok input and $8 per Mtok output.
Because OpenAI o3 was post trained for reasoning rather than fast code emission, the highest value place to spend it in Atlas is the thinking half of the loop. Atlas drafts a plan in a read-only plan agent and asks before switching to a build agent, and o3 is the model that makes that plan agent worth reading. Point o3 at a failing test and a stack trace and let it reason to the root cause, then hand execution to a cheaper coding model. Atlas searches code with hybrid semantic and keyword retrieval fused by reciprocal rank fusion, so the files o3 reasons over are the right ones, and its $2 per Mtok input price means feeding it that retrieved context is not painful.
How much does OpenAI o3 cost per million tokens?
OpenAI o3 costs $2 per Mtok input and $8 per Mtok output, which since OpenAI's 2025 price cut is 7.5x cheaper than o1 on both sides, where o1 charges $15 input and $60 output. For Atlas users, o3 is the cheap way to buy a full reasoning model.
The pricing on OpenAI o3 is what changed the calculus for agent use. A full reasoning model at $2 per Mtok input and $8 per Mtok output sits in the same band as non reasoning workhorses, so you are no longer paying a reasoning tax just to keep the model thinking between tool calls. In Atlas that matters because the agent loop reads a lot: every file read, every diff, and every tool result rides back through input tokens. At $2 per Mtok input, a long o3 session is affordable. The $8 per Mtok output charge is where a verbose model can still surprise you, so watch the cost footer on any session where o3 is emitting long derivations.
Is a 200K token context enough for OpenAI o3 in a real repo?
OpenAI o3 ships a 200K token context, which is half the 400K available on the GPT-5 line and one fifth of GPT-5.5 and later. For most Atlas sessions 200K is workable, but a whole-monorepo prompt will not fit, and o3 leans hard on retrieval instead.
The 200K token window on OpenAI o3 is the honest constraint of the model, and Atlas is built to work inside it. Atlas indexes code by AST declarations using tree-sitter, not blind line windows, so the chunks that reach o3 are whole functions and whole types rather than truncated fragments that waste a scarce window. On the output side o3 is generous: 100K max output tokens is a lot of room for a 200K window model, so a long derivation, a design document, or a large patch fits in one response. If your work genuinely requires holding an entire large codebase in one prompt, o3 is the wrong tool and a million token context model is the right one.
How do you make OpenAI o3 fast enough for an interactive Atlas session?
OpenAI o3 is slower per turn than GPT-5 class models, and that latency is felt directly in an interactive Atlas agent loop. The fix in 2026 is not to make o3 fast, it is to spend it only where reasoning pays, then hand execution to a cheaper coding model.
Latency is the second real tradeoff on OpenAI o3. Atlas calls o3 through the Responses API, so its reasoning persists across tool calls without extra config, which means you are not re-paying to rebuild a chain of thought on every turn. Even so, a long o3 session feels slower than a GPT-5 class one. The practical pattern is a split: pair o3 with Atlas's plan agent to produce the plan, then switch the active model on the fly with favorites and recents and let a faster model execute. Atlas fans out work to subagents that can run in the foreground or in parallel background sessions, so a single o3 subagent can chew on the hard question while the rest of the session keeps moving.
When should you pick a different model instead of OpenAI o3?
Pick a model other than OpenAI o3 when the job needs a window larger than 200K tokens, or when per turn latency matters more than reasoning depth. OpenAI o3 at $2 in and $8 out is a reasoning buy, not a speed buy, and Atlas switches models on the fly.
OpenAI o3 loses on two axes. First, window: 200K tokens is half of the 400K on the GPT-5 line and one fifth of GPT-5.5 and later, so bulk comprehension across a very large context belongs elsewhere. Second, speed: o3 is slower per turn than GPT-5 class models, and in an interactive Atlas session that is felt on every keystroke of waiting. If neither of those is your bottleneck, o3 at $2 per Mtok input and $8 per Mtok output is hard to beat for the money. Atlas lets you switch the active model and provider on the fly with favorites and recents, so the honest answer is to keep o3 one keystroke away rather than pinning it forever.
Setup
- 01Set OPENAI_API_KEY in your shell or store it in Atlas's credential store.
- 02Run `atlas models openai` and confirm o3 resolves in the model list.
- 03Pin "model": "openai/o3" in atlas.json for reasoning heavy work.
- 04Rely on the Responses API path: Atlas calls o3 through it, so its reasoning persists across tool calls without extra config.
- 05Pair OpenAI o3 with Atlas's plan agent, then hand execution to a cheaper coding model once the plan is approved.
Frequently asked questions
- How much does OpenAI o3 cost in Atlas?
- OpenAI o3 bills at $2 per Mtok input and $8 per Mtok output. Atlas passes provider pricing through directly, so a session costs whatever OpenAI charges for the tokens it consumed.
- What is the context window of OpenAI o3?
- OpenAI o3 has a 200K token context window with a 100K max output. That is half the 400K available on the GPT-5 line and one fifth of GPT-5.5 and later.
- Is OpenAI o3 cheaper than o1?
- Yes. OpenAI o3 is 7.5x cheaper than o1 on input ($2 versus $15 per Mtok) and 7.5x cheaper on output ($8 versus $60 per Mtok), and o3 is the newer model.
- How do I set OpenAI o3 as the default model in Atlas?
- Set OPENAI_API_KEY, run `atlas models openai` to confirm o3 resolves, then pin "model": "openai/o3" in atlas.json.
- Is OpenAI o3 good for agentic coding?
- OpenAI o3 is strong on root cause debugging and planning but slower per turn than GPT-5 class models. A common Atlas pattern is to plan with o3, then hand execution to a cheaper coding model.
- Does Atlas keep o3's reasoning across tool calls?
- Yes. Atlas calls OpenAI o3 through the Responses API, so its reasoning persists across tool calls without extra configuration.
- Can I switch away from OpenAI o3 mid session in Atlas?
- Yes. Atlas lets you switch the active model and provider on the fly with favorites and recents, so you can plan on o3 and execute on something faster in the same session.
Try Atlas in your terminal
The terminal-native AI coding agent. Free core, single binary.
Install AtlasRelated guides
Debug a Single Failing Test with Atlas in 2026
How to debug one failing test with Atlas in 2026: run it in isolation with bash, walk the call graph with the lsp tool, and fix the code, not the assertion.
Atlas vs Windsurf: Terminal AI Coding Agents in 2026
Atlas and Windsurf offer distinct AI coding experiences in 2026. Atlas provides a terminal-native TUI with permission-gated tools, while Windsurf features an in-editor live preview and a dedicated GUI.
Diagnose a Hanging or Long-Running Command with Atlas in 2026
How to diagnose a hanging command with Atlas in 2026: the bash tool races every command against a timeout and tells you whether it is slow or blocked on input.
Atlas vs Bolt.new in 2026: Terminal Agent or In-Browser WebContainer Builder
Atlas is a free, open source terminal-native AI coding agent. Bolt.new runs npm install and your dev server in-browser via WebContainers. Compared for 2026.
Atlas for Nim: A Terminal-Native AI Coding Agent for Nimble Packages and Macros in 2026
Atlas is a terminal-native AI coding agent for Nim in 2026. It reads .nimble requires and asterisk-exported symbols, adds std/unittest suites, runs nimble test, formats with nph.
Atlas vs Sourcegraph Cody: Terminal AI Coding Agents in 2026
Comparing Atlas, the terminal-native AI coding agent, with Sourcegraph Cody, an editor extension for whole-codebase search, for developers in 2026. Explore features, pricing, and workflow.
Atlas for R: A Terminal-Native AI Coding Agent for tidyverse, roxygen2, and testthat in 2026
Atlas is a terminal-native AI coding agent for R in 2026. It reads roxygen2 docblocks and renv.lock, rewrites loops as dplyr or purrr pipelines, and runs devtools::test().
Atlas for Bun: Terminal-Native AI Coding in 2026
Atlas is a terminal-native AI coding agent for Bun. Swap Node compatibility shims for Bun.serve and bun:sqlite, iterate with bun test, and format with bun fmt in 2026.