Grok Code Fast 1

What Is Grok Code Fast 1

xAI introduced Grok Code Fast 1 on August 28, 2025 as a coding-focused reasoning model for agentic workflows. According to the xAI announcement, its training and post-training emphasized programming content, real pull requests, coding tasks, and common development tools such as repository search, terminal commands, and file editing.

The model was initially positioned as a responsive option for day-to-day agentic coding. Its launch integrations included tools such as GitHub Copilot, Cursor, Cline, Roo Code, Kilo Code, OpenCode, and Windsurf. Those integrations supplied the surrounding agent loop: they collected repository context, exposed tools, applied edits, and displayed the result.

That distinction matters. A model can propose a patch or decide which tool to call, but the surrounding coding agent controls:

• which files and commands the model can access;
• how repository context is selected;
• whether edits run in an isolated workspace;
• how tests, linting, and type checks are executed;
• how retries, approvals, diffs, and rollbacks are handled.

In current xAI documentation, the older identifiers grok-code-fast-1, grok-code-fast, and grok-code-fast-1-0825 appear as aliases for grok-build-0.1. The current model page should be treated as the source of truth for the active endpoint and rate information.

Coding Benchmarks & Performance

xAI reported a 70.8% score on the full SWE-bench Verified subset for Grok Code Fast 1 at launch. The result came from xAI's own internal harness, so it should be cited as a vendor-reported result—not as a universal score that every coding agent will reproduce.

SWE-bench Verified measures whether a system can resolve selected real-world GitHub issues. It is relevant to agentic coding, but the final result depends on more than the base model. The harness, tool permissions, context retrieval, prompt, retry policy, test environment, and time budget can all change the outcome.

For a practical evaluation, use one fixed repository state and compare the following:

1. Task understanding: Did the model identify the correct files, constraints, and acceptance criteria?
2. Diff quality: Is the patch focused, readable, and free from unrelated changes?
3. Verification: Do the relevant tests, lint checks, and type checks pass?
4. Repair effort: How many retries or manual edits were needed after the first result?
5. Latency: How long did the complete task take, including tool calls and repairs?
6. Task cost: What was the total provider cost after context, output, retries, and caching?

Grok Code Fast 1 was designed around short feedback loops and tool use. That can make it a useful candidate for bounded bug fixes, codebase questions, targeted refactors, test generation, and other tasks where the agent can verify each step. Larger architectural changes still need planning, repository-specific evaluation, and human review.

Grok Code Fast vs Devstral 2 vs Claude Code

Grok Code Fast 1, Devstral 2, and Claude Code are not three versions of the same product. The first two are coding models; Claude Code is a coding-agent product that combines Claude models with repository access, file editing, command execution, and development workflows.

Choose Grok Code Fast 1 when

• you want to test xAI's coding-specific model inside an existing agent or API workflow;
• low-friction tool loops and bounded coding tasks are the priority;
• you can verify the active alias and provider route before deployment;
• you are prepared to measure full task cost rather than compare only token prices.

Choose Devstral 2 when

• an open-weight model and deployment control are important;
• your team can support its infrastructure requirements or use Mistral's hosted API;
• you need to evaluate licensing, fine-tuning, or on-premises deployment;
• multi-file repository work is central to the test plan.

Mistral's Devstral 2 release states that the full 123B model requires data-center-class hardware for self-deployment. Devstral Small 2 is the separate, smaller option intended for more accessible local deployment.

Choose Claude Code when

• you want an end-to-end coding agent instead of a model endpoint;
• the workflow needs repository exploration, file edits, test execution, Git operations, and pull-request support in one tool;
• your team already uses a supported Claude subscription, the Anthropic Console, or a supported enterprise cloud route;
• agent permissions and workflow integration matter more than selecting a standalone model API.

The Claude Code setup guide documents these access paths. A direct price comparison with Grok Code Fast 1 or Devstral 2 would be misleading unless the Claude model, billing route, context use, and agent workload are held constant.

Pricing & API Access

Grok Code Fast 1 pricing has changed since its original release. Use current xAI documentation for budgeting.

The current xAI model page lists a 256,000-token context window and identifies grok-code-fast-1 as an alias for grok-build-0.1. When prompt length exceeds 200K tokens, budget the request using the long-context row for input, cached input, and output. That means an existing integration may continue to recognize the older name, while new work should follow the active identifier documented by xAI.

For direct API access:

1. Create or use an xAI developer account and API key.
2. Confirm the currently supported model ID in the xAI model documentation.
3. Follow xAI's current Responses API example rather than copying an older launch snippet.
4. Test a small request before sending a large repository context.
5. Record input, cached input, output, retries, and tool-loop cost separately.

The xAI model page is the source of truth for current rates, aliases, regional availability, and long-context pricing.

Use this migration checklist before relying on an older integration:

Using Grok Code Fast in Verdent

Grok Code Fast 1 is not listed as a built-in model in Verdent's current model-pricing documentation. Verdent currently documents BYOK support for Anthropic, OpenAI, and OpenRouter—not direct xAI keys.

You may be able to evaluate Grok Code Fast 1 through OpenRouter BYOK if the active xAI alias appears in the model list available to your OpenRouter account. This is conditional provider access, not native xAI support inside Verdent.

To check availability:

1. Open Settings → Models → Configure Models in Verdent Desktop.
2. Select OpenRouter and add a valid API key.
3. Review the available model list for grok-build-0.1, grok-code-fast-1, or another current alias documented by the provider.
4. Enable the model only if it appears and key validation succeeds.
5. Run a bounded repository task before using it in a broader workflow.

Verdent's BYOK documentation says enabled BYOK models can appear in the chat model selector, presets, subagent settings, and reviewer settings. It also notes that Smart Suggestions and automatic compression do not support BYOK API keys.

A controlled model trial should use the same repository state, prompt, acceptance criteria, and checks for every candidate. In Verdent, you can isolate the task in a workspace, inspect the plan and diff, run relevant tests, and compare repair effort against a current built-in model. If the Grok alias does not appear through OpenRouter, use xAI's API in a compatible external agent instead of describing the model as available in Verdent.

If you are reviewing older xAI fast-model routes, the Grok 4.1 Fast guide explains that model's retirement status and why it should not be treated as the current coding endpoint.