GPT-5.6 Terra is the mid tier variant of the July 2026 GPT-5.6 line from OpenAI. Terra carries the same 1.05M tokens (1,050,000) window as full GPT-5.6 but bills at $2.50 per Mtok input, $15 per Mtok output instead of $5 / $30. Inside Atlas, the terminal-native AI coding agent, GPT-5.6 Terra is the default GPT-5.6 pick when you want reasoning depth without Sol pricing. It is a reasoning model called through the Responses API, so Atlas keeps reasoning state alive across the tool call loop rather than restarting the chain after each tool result.
What is GPT-5.6 Terra best at inside Atlas?
GPT-5.6 Terra is best inside Atlas as the everyday reasoning model on a very large window. Terra gives you the identical 1,050,000 token context and 128K max output of full GPT-5.6, at half the input price, $2.50 per Mtok instead of $5 per Mtok.
The value of GPT-5.6 Terra in Atlas is that you do not have to choose between window size and budget. A whole service read plus a substantial patch fits in the 1,050,000 token context, and the 128K max output means the patch comes back in one piece. Atlas computes a unified diff for every file edit and surfaces it for approval before writing, so a single large GPT-5.6 Terra response becomes one review pass. Atlas indexes code by AST declarations using tree-sitter, not blind line windows, so the declarations you feed Terra are whole functions and types. At $2.50 per Mtok input, a wide read on GPT-5.6 Terra is genuinely affordable in a way the same read on Sol is not.
How does GPT-5.6 Terra keep reasoning state across Atlas tool calls?
GPT-5.6 Terra is a reasoning model called through the Responses API, and Atlas routes OpenAI through sdk.responses(), so no extra flag is needed to persist reasoning between tool calls. The reasoning chain survives each tool result instead of being rebuilt from scratch every turn.
Reasoning persistence is what makes GPT-5.6 Terra behave well in a long agent loop. Atlas fans out work to subagents that can run in the foreground or in parallel background sessions, and each subagent driven by GPT-5.6 Terra keeps its reasoning trace across the tool calls it makes. Without that persistence, every grep result and every file read would force the model to re derive why it went looking, which costs output tokens at $15 per Mtok and costs coherence on multi step refactors. Because Atlas routes OpenAI through sdk.responses(), GPT-5.6 Terra gets this behavior with no configuration beyond pinning the model. Every Atlas tool call is still permission-gated against allow, ask, and deny rules before it runs.
How much does the 1.05M token context of GPT-5.6 Terra actually cost?
GPT-5.6 Terra bills long context prompts per token like any other model, so a 1M token load costs $2.50 before any output. The 1,050,000 token window on Terra is headroom, not a free lunch, and filling it on every turn is the fastest way to run up a session bill.
The right way to use the GPT-5.6 Terra window is selectively. Atlas searches code with hybrid semantic and keyword retrieval fused by reciprocal rank fusion, which means you can pull the handful of files that matter into the Terra context instead of dumping the repo. Atlas drafts a plan in a read-only plan agent and asks before switching to a build agent, so exploration happens against retrieved declarations rather than against a full 1,050,000 token load. Used that way, GPT-5.6 Terra usually runs well under its ceiling on typical turns, and the big window is there for the rare task that genuinely needs it. At $15 per Mtok output, the write side is still where a long loop accumulates cost.
When should you pick Luna or Sol instead of GPT-5.6 Terra?
Pick GPT-5.6 Luna at $1 / $6 for cheap mechanical work, and pick Sol or full GPT-5.6 when you need the higher effort variants. GPT-5.6 Terra sits between Luna and Sol as the default when you want reasoning depth without Sol pricing at $5 / $30.
GPT-5.6 Terra is explicitly not the top of the 5.6 line. Sol and full GPT-5.6 are the higher effort variants, so a genuinely hard architectural problem is a reason to escalate away from Terra. In the other direction, Luna at $1 per Mtok input and $6 per Mtok output is the cheaper floor, which makes it the right target for the Atlas small_model slot and for high volume mechanical edits. Atlas lets you switch the active model and provider on the fly with favorites and recents, so use /models in the TUI to A/B GPT-5.6 Terra against gpt-5.6-luna on the same task and compare cost in the session footer. That comparison, not a spec sheet, should decide your default.
Setup
- 01Export OPENAI_API_KEY, or run `atlas login` and choose OpenAI.
- 02Run `atlas models openai` and confirm gpt-5.6-terra resolves.
- 03Set "model": "openai/gpt-5.6-terra" in atlas.json.
- 04Rely on the fact that Atlas routes OpenAI through sdk.responses(), so no extra flag is needed to persist reasoning between tool calls.
- 05Use /models in the TUI to A/B GPT-5.6 Terra against gpt-5.6-luna on the same task and compare cost in the session footer.
Frequently asked questions
- how much does gpt-5.6 terra cost per million tokens
- GPT-5.6 Terra costs $2.50 per Mtok input and $15 per Mtok output. That is half the input price of full GPT-5.6, which bills at $5 / $30, and more than GPT-5.6 Luna at $1 / $6.
- what is the context window of gpt-5.6 terra
- GPT-5.6 Terra has a 1.05M token context window (1,050,000 tokens) with 128K max output, identical to full GPT-5.6. Terra is the mid tier variant of the July 2026 GPT-5.6 line.
- how do I set gpt-5.6 terra as the atlas model
- Export OPENAI_API_KEY or run `atlas login` and choose OpenAI, run `atlas models openai` to confirm gpt-5.6-terra resolves, then set "model": "openai/gpt-5.6-terra" in atlas.json.
- gpt-5.6 terra vs sol which should I use in atlas
- GPT-5.6 Terra gives the same 1,050,000 token window at $2.50 / $15, while Sol is the higher effort variant at $5 / $30. Pick Terra for reasoning depth without Sol pricing, and escalate to Sol for the hardest tasks.
- does gpt-5.6 terra need a flag to keep reasoning between tool calls in atlas
- No. Atlas routes OpenAI through sdk.responses(), so GPT-5.6 Terra persists reasoning between tool calls with no extra flag. Pinning "model": "openai/gpt-5.6-terra" is enough.
- is a 1m token prompt on gpt-5.6 terra expensive
- A 1M token load on GPT-5.6 Terra costs $2.50 in input before a single output token, since long context prompts still bill per token. Atlas hybrid semantic and keyword search keeps you from filling the window unnecessarily.
- what is the max output of gpt-5.6 terra
- GPT-5.6 Terra supports 128K max output, the same ceiling as full GPT-5.6. In Atlas that means a whole service read plus a large patch can fit in a single turn.
Try Atlas in your terminal
The terminal-native AI coding agent. Free core, single binary.
Install AtlasRelated guides
Atlas for SQL in 2026
Atlas is a terminal-native AI coding agent for SQL in 2026. Run it in a repo with your migrations or .sql files, optimize a query, and review the diff before applying.
Atlas vs Cline: Terminal AI Coding Agents in 2026
Compare Atlas and Cline, two leading AI coding agents for 2026. Atlas offers terminal-native TUI and permission-gated tools, while Cline integrates with VS Code and provides checkpoint rollbacks.
Atlas vs Magic.dev: Terminal AI Coding Agents in 2026
Compare Atlas, the terminal-native AI coding agent with permission-gated tools and diff review, against Magic.dev's research claims of 100 million token context models in 2026.
Atlas for Symfony in 2026
Atlas is a terminal-native AI coding agent for Symfony in 2026. It reads autowired services and Doctrine mappings, and shows migration SQL before anything runs.
Atlas for Axum in 2026
Atlas is a terminal-native AI coding agent for Axum in 2026. It decodes tower trait-bound errors, adds IntoResponse types, and runs cargo nextest run.
Atlas vs Poolside: Terminal AI Coding Agents in 2026
Compare Atlas and Poolside, two terminal AI coding agents in 2026. Atlas offers a free core and robust safety features, while Poolside provides open-weight Laguna models and on-prem deployment options.
Atlas for Erlang in 2026
Atlas is a terminal-native AI coding agent for Erlang/OTP in 2026. Run it in an app with a rebar.config, map supervisors and gen_server modules, review every diff.
Atlas vs Traycer in 2026: Terminal Agent That Writes Code vs a Planning Layer Above One
Atlas vs Traycer in 2026. Traycer plans and verifies but writes no code itself, so you pay it on top of an agent. Atlas plans and writes, free and open source.