AI 程式開發工具
Grok 4.5 Comes to GitHub Copilot With a 500K Context Window and Cross-IDE Reasoning Controls
GitHub has begun rolling out Grok 4.5 across Copilot’s IDE integrations, CLI, cloud agent, and desktop app, offering a context window of up to 500,000 tokens, text and image input, and three reasoning effort levels. Enterprise administrators must explicitly enable the model policy, while GitHub has so far disclosed only observations from internal terminal-based tasks, with no comparable Copilot benchmark results.

GitHub began deploying SpaceXAI’s Grok 4.5 to Copilot on July 28. The new model is available from the model selector in VS Code, Visual Studio, JetBrains IDEs, Xcode, Eclipse, Copilot CLI, the cloud agent, and the Copilot App. It supports text and image input, a context window of up to 500,000 tokens, and three reasoning effort levels: low, medium, and high. Eligible plans include Pro, Pro+, Max, Business, and Enterprise. Administrators of Business and Enterprise organizations must separately enable the Grok 4.5 policy, which is disabled by default.
The technical significance of this update goes beyond adding another model name. A 500K-token capacity can accommodate larger portions of codebases, terminal output, and agent histories, while adjustable reasoning effort lets teams balance latency and cost against deeper, multi-step analysis. GitHub says Grok 4.5 performed particularly well at dispatching tools in parallel and taking direct action in internal terminal-based coding tasks. This aligns with SpaceXAI’s positioning of the model as an asynchronous coding agent. Vendor data indicates a serving speed of about 80 tokens per second and an average output of 15,954 tokens on SWE-bench Pro—claimed to be about 4.2 times lower than Opus 4.8 max. API pricing is $2 per million input tokens and $6 per million output tokens.
However, these figures cannot be applied directly to Copilot. GitHub meters usage based on vendor list pricing, while agents also add prompts, tool results, and retries; actual cost and end-to-end latency therefore depend on the client and the task. SpaceXAI’s comparisons also mix vendor harnesses from different models rather than using a fully controlled, head-to-head evaluation. Engineering teams should first compare success rates, change scope, tool-call counts, and token consumption using a fixed repository workload. They should also verify the data-governance requirements for images, source code, and terminal content under the new model provider. The key questions going forward are whether GitHub will publish reproducible Copilot evaluations and whether long context genuinely improves cross-file fixes instead of merely increasing input costs.