← Home Model Selection

Which Copilot model · and when

v1 · sourced from docs.github.com

Default to Auto. Override when the task demands it.

GitHub Copilot Business and Enterprise expose 18+ models across OpenAI, Anthropic, Google, and xAI. Customers don't need to memorise them — they need a clear rule: start with Auto model selection, then escalate or downshift based on the task. This page maps GitHub's official task taxonomy to recommended models for the two UBB plans, with the per-token AI Credit pricing (effective June 1, 2026) overlaid on every card.

Default
Auto
GitHub picks the best model
for each task
1 · Pick the task, see the model
GitHub's docs group all models into five task categories. Filter by task and by plan to see what's recommended and what it costs.
Task category
All General coding & writing Fast / repetitive Deep reasoning & debugging Agentic development Visuals (diagrams, screenshots)
Plan
Business Enterprise
Lightweight — fast / cheap (output ≤ ~$5/1M) Versatile — general workhorse (output ~$5–$15/1M) Powerful — deep reasoning / premium (output ~$10–$25/1M) Frontier — highest cost (output $25–$30/1M) Included — no AI Credits consumed
Pricing source — All rates shown are GitHub's published Usage-Based Billing per-token prices (per 1M tokens), effective June 1, 2026 for Copilot Business and Enterprise. Tokens are converted to GitHub AI Credits (1 AI Credit = $0.01 USD) and drawn from the per-user pooled allowance. Code completions and next-edit suggestions are not billed in AI Credits. See Models and pricing for the authoritative tables.
2 · Anti-patterns to call out
The four most expensive customer mistakes — and how to avoid them.
⚠️ Frontier model on autocomplete-style tasks
Using Claude Opus 4.7 ($25 / 1M out) or GPT-5.5 ($30 / 1M out) for "rename this variable" or "add a docstring" burns 10–15× more AI Credits than GPT-5 mini ($2 / 1M out, included) for no quality gain. Reach for Powerful-tier models only when reasoning depth is the bottleneck.
⚠️ Powerful model left on for agent loops
Agent mode often sends dozens of follow-up prompts per task — each one is billed per token. Pair agent workflows with GPT-5.2-Codex / GPT-5.3-Codex / GPT-5.4 mini (Lightweight–Versatile tier) instead of Sonnet 4.6 or Opus when the task doesn't demand frontier reasoning.
⚠️ No user-level budget set
Without a user-level budget, a single power user on Opus 4.7 can drain a meaningful share of the org's AI Credit pool in days. Set per-user budgets for everyone; raise the cap explicitly for senior / agentic users.
✅ Use Auto, then promote
Start every team on Auto model selection. Only allow-list specific Powerful-tier models for users / teams whose work justifies it. This is the single highest-leverage governance move — it prevents accidental Opus / GPT-5.5 spend across the org.
3 · Authoritative references
All recommendations on this page are sourced from public GitHub documentation.
This content has been authored by Microsoft Asia Developer GBB and is based on publicly available GitHub Copilot documentation. It has not been validated across all supported configurations. While the team makes continuous efforts to stay aligned with the latest documentation, users are advised to verify details against official GitHub documentation before making any decisions.