Grok 4.5 is xAI's July 2026 flagship, and inside Atlas it is the model to reach for when a single turn has to produce an enormous patch. Grok 4.5 runs on a 500K tokens context window with a symmetric 500,000 max output tokens budget at $2 / $6 per Mtok (input / output). Atlas routes xAI through the Responses API, the same path it uses for OpenAI, calling `sdk.responses(modelID)` so reasoning state carries across tool calls rather than being rebuilt each turn.
What makes Grok 4.5's 500,000 output token budget matter in Atlas?
Grok 4.5 offers 500,000 max output tokens, roughly four times the 128K ceiling of Claude Opus 4.8 and the GPT-5 family. Inside Atlas, where every file edit is rendered as a unified diff before it is written, that budget is what lets Grok 4.5 emit an entire refactor in one turn.
Output budget is the constraint that actually breaks agentic refactors, and Grok 4.5 is the model that removes it. Atlas computes a unified diff for every file edit and surfaces it for approval before writing. On a sweeping change, the diff itself is the model's output, so a 128K output cap forces the work to be split across turns and reviewed piecemeal. Grok 4.5's 500,000 max output tokens, four times that ceiling, lets Atlas present the full patch once. Atlas snapshots file changes as git patches, so even a very large single-turn edit from Grok 4.5 can be diffed and rolled back cleanly.
How does Atlas connect to Grok 4.5?
Atlas connects to Grok 4.5 through the xAI provider, loaded via `@ai-sdk/xai` and authenticated with `XAI_API_KEY`. Atlas calls `sdk.responses(modelID)` for xAI, the same Responses API path it uses for OpenAI, so reasoning state carries across tool calls rather than being rebuilt each turn.
The transport detail is not cosmetic for Grok 4.5. Because Atlas routes xAI through the Responses API, a Grok 4.5 session that runs a chain of tool calls keeps its reasoning state instead of reconstructing it on every turn. That matters in Atlas specifically, where every tool call is permission-gated against allow, ask, and deny rules before it runs, so a long agentic loop can involve many round trips with approvals in between. Export `XAI_API_KEY=...`, run `atlas models xai` to confirm the model resolved, and select Grok 4.5 from `/models`. To make it the default, pin `"model": "xai/grok-4.5"` in atlas.json.
How much does Grok 4.5 cost compared to Claude Opus 4.8?
Grok 4.5 costs $2 / $6 per Mtok (input / output). At $6 per Mtok output, Grok 4.5 is less than a quarter of Claude Opus 4.8's $25 while still sitting at frontier positioning, which is what makes its 500,000 token output budget financially usable rather than theoretical.
The pricing and the output ceiling reinforce each other on Grok 4.5. A model that can emit 500,000 tokens but charges $25 per Mtok output would make big single-turn patches painful; Grok 4.5 charges $6, less than a quarter of Claude Opus 4.8's rate, so the large-diff workflow is affordable. On the input side, $2 per Mtok covers the context Atlas assembles by searching code with hybrid semantic and keyword retrieval fused by reciprocal rank fusion. For refactor-heavy Atlas sessions where output volume dominates the bill, Grok 4.5 is the cost-efficient frontier option.
What are the tradeoffs of running Atlas on Grok 4.5?
Grok 4.5's 500K context is half the 1M offered by Sonnet 5, GPT-5.6, and Gemini 3.1 Pro. Grok 4.5 also has a shorter track record on agentic coding evaluations than the Anthropic or OpenAI lines, since xAI shipped it in July 2026.
Two honest limits on Grok 4.5. The context window, at 500K tokens, is generous but half of what Sonnet 5, GPT-5.6, and Gemini 3.1 Pro provide, so a session that wants to hold an entire large monorepo in context will hit the wall sooner. And the evaluation record is thinner: the Anthropic and OpenAI lines have been measured on agentic coding for longer, so there is less public evidence about how Grok 4.5 behaves on the hardest tasks. Atlas gives you a safety valve here, since every Atlas tool call is permission-gated against allow, ask, and deny rules before it runs, so you can evaluate Grok 4.5 with edits set to ask.
When should you pick a different model than Grok 4.5?
Pick a different model than Grok 4.5 when you need a 1M token context window or a longer proven record on agentic coding. Sonnet 5, GPT-5.6, and Gemini 3.1 Pro all offer 1M against Grok 4.5's 500K, and the Anthropic and OpenAI lines have deeper evaluation coverage.
Grok 4.5 wins on a specific axis: output volume per dollar. If your Atlas work is dominated by generating large diffs, Grok 4.5's 500,000 max output tokens at $6 per Mtok is the strongest combination available. If your work is dominated by reading, holding a huge repository in context, or squeezing the last few points of quality out of a hard agentic task, a 1M context frontier model from Anthropic, OpenAI, or Google is the better choice. Atlas lets you switch the active model and provider on the fly with favorites and recents, so Grok 4.5 can be the refactor model without being the only model.
Setup
- 01Export `XAI_API_KEY=...` in your shell. Atlas loads xAI via `@ai-sdk/xai`.
- 02Confirm the model resolved by running `atlas models xai`.
- 03Select Grok 4.5 from `/models` in the Atlas TUI.
- 04Optional: pin `"model": "xai/grok-4.5"` in atlas.json to make it the session default.
- 05Keep Atlas's permission rules on ask for edit and bash while you evaluate Grok 4.5's tool-calling behavior.
Frequently asked questions
- How much does Grok 4.5 cost per million tokens?
- Grok 4.5 costs $2 / $6 per Mtok (input / output). The $6 output rate is less than a quarter of Claude Opus 4.8's $25 at comparable frontier positioning.
- What is Grok 4.5's context window and output limit?
- Grok 4.5 runs a 500K tokens context window with a symmetric 500,000 max output tokens budget, roughly four times the 128K ceiling of Claude Opus 4.8 and the GPT-5 family.
- How do I use Grok 4.5 with Atlas?
- Export `XAI_API_KEY=...`, confirm with `atlas models xai`, then select Grok 4.5 from `/models`. To pin it, set `"model": "xai/grok-4.5"` in atlas.json.
- Does Atlas use the Responses API for Grok 4.5?
- Yes. Atlas routes xAI through the Responses API, the same path it uses for OpenAI, calling `sdk.responses(modelID)` so reasoning state carries across tool calls rather than being rebuilt each turn.
- Is Grok 4.5 good for large refactors?
- Yes. Grok 4.5's 500,000 max output tokens let it emit a very large unified diff in a single turn, and at $6 per Mtok output that volume stays affordable.
- Does Grok 4.5 have a 1M context window?
- No. Grok 4.5 has a 500K tokens context window, half the 1M offered by Sonnet 5, GPT-5.6, and Gemini 3.1 Pro.
- Is Grok 4.5 proven on agentic coding benchmarks?
- Grok 4.5 has a shorter track record on agentic coding evaluations than the Anthropic or OpenAI lines. xAI shipped it in July 2026, so evaluation coverage is still thinner.
Try Atlas in your terminal
The terminal-native AI coding agent. Free core, single binary.
Install AtlasRelated guides
Atlas vs Tabnine: Choosing Your Terminal AI Coding Agent in 2026
Comparing Atlas and Tabnine for developers in 2026. Atlas offers a terminal-native AI agent with diff review, while Tabnine provides privacy-first code completion and chat.
Atlas for Assembly: Registers, Calling Conventions, and nasm in 2026
Atlas is a terminal-native AI coding agent for Assembly in 2026. It reads .asm and .S sources, tracks System V and AAPCS64 calling conventions, and assembles with nasm behind a prompt.
Atlas vs Aider: Terminal AI Coding Agents in 2026
Comparing Atlas and Aider for 2026 developers. Atlas offers a rich TUI, permission-gated tool calls, and plugin support, while Aider provides a repo-map and commits every change as a discrete git revision.
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 for Python in 2026
Atlas is a terminal-native AI coding agent for Python in 2026. Run it in a repo with a pyproject.toml or requirements.txt and review every diff before it lands.
Document a Module with a README Using Atlas (2026 Workflow)
How to document a module with a README using Atlas in 2026: the lsp tool's documentSymbol enumerates the real exports, read supplies the behavior, write emits the README.
Atlas for COBOL: Copybooks, PIC Clauses, and GnuCOBOL in 2026
Atlas is a terminal-native AI coding agent for COBOL in 2026. It reads your divisions and copybooks, explains a paragraph before touching it, and compiles with cobc under GnuCOBOL.
Atlas for Perl: A Terminal-Native AI Coding Agent for CPAN Distributions in 2026
Atlas is a terminal-native AI coding agent for Perl in 2026. It reads cpanfile deps and @EXPORT lists, writes Test2::V0 cases, runs prove -lr t/, and runs perltidy on the diff.