x-High and AI Credits¶

For / Key Points

For: GitHub Copilot (Pro/Pro+/Enterprise) users who want to understand what the reasoning level selector actually controls

Key Points:

Reasoning levels control how much the model "thinks" — same model, different thinking token budgets
Under usage-based billing after June 2026, higher reasoning levels can consume more AI Credits
Start with defaults, then raise effort for design decisions, multi-file work, and debugging

What Are Reasoning Levels?¶

VS Code 1.113 (March 25, 2026) introduced the Thinking Effort selector — a control that adjusts how deeply the model reasons before generating a response. Today, the model picker is the primary control surface, and the old github.copilot.chat.responsesApiReasoningEffort setting is deprecated.¹

The model picker shows a submenu with levels such as Low / Medium / High when a selected model supports reasoning. Available levels vary by model and surface. Non-reasoning models do not show the Thinking Effort submenu.

The key point: changing the reasoning level does not change the model itself. The same GPT-5.4 at Low produces a minimal reasoning chain for an instant response, while at x-High it allocates significantly more thinking tokens for deeper analysis.

Supported Models and Levels¶

Model family	Thinking Effort	Practical judgment
GPT-5.4 / GPT-5.3-Codex and similar reasoning models	Available when supported	Complex design and debugging
Claude Sonnet / Opus reasoning models	Available when supported	Quality-focused implementation and review
GPT-4.1 / GPT-4o / GPT-5 mini and similar models	Often not shown	Lightweight questions and routine chat
Claude Haiku / Gemini Flash and similar models	Often not shown	Speed- and cost-sensitive work

Unsupported models (no reasoning level selector):

GPT-4.1, GPT-4o, GPT-5 mini — non-reasoning models
Claude Haiku 4.5, Gemini 3 Flash — speed-optimized

Do not freeze this as a static model table

GitHub Copilot models, extended context, and reasoning controls change quickly. Prefer the live model picker and official model documentation over static model names in this article.³

Cost: Reasoning Level Now Matters for AI Credits¶

This is the most important fact.

Under usage-based billing after June 2026, raising reasoning effort can increase AI Credits consumption.

GitHub explains Copilot cost as a function of the model and token consumption: input tokens, output tokens, and cached context are converted into AI Credits. GitHub's June 4, 2026 changelog also states that larger context windows or higher reasoning levels consume more AI Credits per interaction.²³

In the legacy premium request model, it was sometimes reasonable to say "same model, same request multiplier." In the current AI Credits model, the same model can cost more if higher effort creates more thinking tokens or output.

Cost lens	Legacy premium requests	AI Credits after June 2026
Unit	Request count and model multiplier	Model price, input, output, cached context, thinking tokens
High effort	Often treated as the same multiplier for the same model	Can increase per-interaction consumption
Savings tactic	Reduce round trips	Reduce round trips, context size, and unnecessary effort

So High is not a free quality switch. Use it when the task justifies deeper reasoning, but do not pin every lightweight chat to High or x-High.

What Each Level Actually Does¶

Level	Behavior	Latency	Best For
Low	Minimal reasoning. Instant response	Fastest	Boilerplate, syntax questions
Medium / default	Balanced	Fast	General coding
High	Deep reasoning chain	Slightly slower	Refactoring, design decisions
x-High / extended thinking	Highest reasoning budget	Slowest and more likely to consume credits	Security audits, large migrations

Watch Credit Consumption, Not Only Speed¶

Higher effort can improve results on complex work. The VS Code AI Credits optimization guide also warns that higher effort produces more thinking tokens, increasing both latency and credit consumption.⁴

Low is genuinely faster, but response quality drops noticeably on certain tasks. Multi-file refactoring at Low, for example, can miss cross-file dependencies.

When to Raise Effort¶

Raise effort based on quality risk.

Situation	Recommendation
Naming, simple API questions, short explanations	Default or lightweight model
Small one- or two-file edits	Start with default, raise only if it fails
Multi-file refactoring	Consider High
Security review, migration design, incident analysis	Consider High or above
Long logs or repository-wide context	Narrow the scope first, then raise effort only if needed

The key step is narrowing scope before raising effort. If you send a large context window and also raise effort, quality may improve, but AI Credits consumption is more likely to rise too.

Does Low Have a Place?¶

Yes, but it's limited.

Low works well for:

Simple code completions and snippet generation
"What's the signature for this function?" type questions
Rapid feedback loops during prototyping
Auto-generating doc comments

Avoid Low for:

Multi-file refactoring
Security review and vulnerability analysis
DB schema changes or framework migrations
Architecture decisions

The rule is simple: tasks that don't require thinking → Low or default. Tasks that need real reasoning → High or above.

But "when in doubt, use High" is no longer the right cost default. If you care about AI Credits, start with the default and raise effort only when the result is not good enough.

Recommended Configuration¶

Use task-based switching, not always-on High.

Use defaults for routine chat, small edits, and explanation tasks.
Use High for design decisions, multi-file edits, and root-cause analysis.
Reserve x-High or extended thinking for security audits, large migrations, or failed prior attempts.
Before combining long context with High effort, trim files, logs, and task scope.

Switch from the model picker

VS Code 1.113+ lets you switch Thinking Effort from the model picker for supported models. Prefer checking the UI per task instead of pinning the deprecated settings.json setting.

For broader cost design, read GitHub Copilot AI Credits Optimization.

Summary¶

Reasoning levels control how much the model "thinks." The model itself doesn't change
Under AI Credits, higher reasoning levels can consume more credits
High is worth using for complex work where quality risk is high
Start routine work at the default, then raise effort only when needed

Reasoning effort is a quality control. It is also a credit-consumption control. Treat it as a per-task decision, not a permanent maximum setting.

GitHub Copilot AI Credits Optimization — Optimize usage-based billing and AI Credits consumption
GitHub Copilot Hooks Complete Guide — Workflow automation
GitHub Copilot Custom Instructions Guide — Improve prompt quality