How Do I: Choose the Right AI Assistant for AL Development?

GitHub Copilot’s move to usage-based billing with model multipliers taking effect June 1 is a good moment to step back and ask a question most AL developers have not had to think about seriously:

Is Copilot the right tool for every AI-assisted task or just the most convenient one?

The answer depends heavily on what you are actually trying to do. AL development involves at least two distinct modes of work:

writing code inside VS Code with full access to your project
and thinking about code, architecture, error analysis, and documentation away from the editor.

These modes have different requirements, and no single tool is the obvious winner across both.

GitHub Copilot in VS Code: strong where context matters

Copilot’s structural advantage for AL development is editor integration.

It sees your open files, your workspace symbols, your AL object names, your event publishers and subscribers.
For anything that depends on awareness of your actual codebase, that context is genuinely valuable and not easily replicated by copying and pasting into a chat window.

The Chat and Agent modes in VS Code extend this further.

Agent mode can read multiple files, run the AL compiler, interpret error output, and iterate.
For a contained task within a known codebase, that is a meaningful capability.
The question is what it costs under the new pricing model.

Under the new multiplier table, a long agentic session using a frontier reasoning model is no longer priced the same as a quick inline completion.

Claude Haiku sits at roughly 0.33x
Claude Sonnet at around 1x
Claude Opus at approximately 27x

For reasoning-heavy tasks, you are paying a significant premium to keep that work inside VS Code rather than moving it to a separate tool.

Claude.ai: the case for moving reasoning work outside the editor

Claude.ai on a Pro subscription gives you direct access to Claude Sonnet and Opus without a Copilot multiplier on top.

The tasks that work well here are those where codebase context matters less than reasoning quality:

Why does this codeunit behave differently when called from a job queue versus a page action?
How should I structure this AL interface to avoid the coupling problem I currently have?

You paste the relevant code, ask the question, and get a substantive answer.

The bottleneck is never context. It is understanding Business Central itself, and neither Copilot nor Claude is particularly strong there.

Use Copilot in VS Code for code generation within your project.
Use Claude.ai for sustained reasoning, architecture discussion, and code review.
For reasoning-heavy work it is often cheaper to switch tools than to stay in VS Code.

Claude desktop app: the same model, closer to your workflow

The Claude desktop app is Claude.ai with tighter OS integration:

a global shortcut
desktop extensions
and direct access to local files without manual copy-pasting.

The subscription cost is the same as Claude.ai Pro. The difference is convenience and file access.

Careful: Security is a question here: Researcher claims Claude Desktop installs “spyware” on macOS | Malwarebytes

Other AI assistants

ChatGPT with a Plus or Pro subscription offers comparable reasoning quality to Claude for most AL questions, and its code interpreter adds occasional value for data analysis tasks adjacent to BC development such as parsing telemetry exports.
Microsoft Copilot, available through many existing Microsoft 365 subscriptions, has some Dynamics 365 awareness but adds little for AL development specifically over ChatGPT or Claude.
Gemini Advanced is capable on general coding questions but has less BC community usage and no strong reason to prefer it for AL work.

A practical decision framework

Task	Best tool	Why
Inline completions, boilerplate, field declarations	Copilot in VS Code	Sees your open files, symbols, and AL workspace
Architecture questions, code review, error analysis	Claude.ai / Desktop	Frontier model without Copilot multiplier cost
Agentic multi-file refactoring or test writing	Copilot Agent / Claude Code	Editor context is genuinely necessary; accept the higher cost
Documentation, AL procedure explanation	Claude.ai / Desktop	Reasoning quality matters more than codebase context
General BC functional questions, quick lookups	Any model, including free tier	No frontier model needed for straightforward questions

The right tool depends on the task, not on loyalty to a single subscription.

For inline completions, boilerplate, and routine code generation inside your project, GitHub Copilot in VS Code with a cost–appropriate model is the right choice.
For architecture questions, complex error analysis, code review, and documentation, Claude.ai or the Claude desktop app on a Pro subscription gives you frontier model access without Copilot multipliers.
For agentic multi-file refactoring or test writing where editor context is genuinely necessary, GitHub Copilot Agent mode or Claude Code is the right tool, accepting the higher cost.
For general BC functional questions or quick lookups, any capable model is sufficient, including free tiers.

The June 1 pricing change does not make Copilot a bad tool. It makes the cost of using frontier models inside Copilot visible in a way it was not before.

GitHub is running a usage preview through May so you can see your projected bill before the switch.
Check your Copilot usage dashboard, identify which tasks are driving the most premium model usage, and decide whether those tasks are better handled inside VS Code or in a separate tool at a flat subscription cost.

At this point this is how I read the impact. In practice it may turn out differently once usage-based billing has been in effect for a while and real cost patterns become visible.

If you want to dig into the details:

Update: A follow-up post clarifies two billing models that the original version of this post conflated. If you are evaluating the cost of using frontier models in Copilot, read GitHub Copilot Billing: Setting the Record Straight first.