Usage-Based Billing with Confidence

01 What Changes on June 1, 2026

GitHub Copilot Business and Copilot Enterprise usage is measured with GitHub AI Credits. Interactions consume tokens — input, output, and cached — and GitHub converts that usage into AI credits.

Important. GitHub says the pricing and billing methods described here start on June 1, 2026. Review the live GitHub docs before making billing or procurement decisions, because model availability and pricing can change.

The goal is not to use Copilot less

The goal is to use Copilot with clearer intent: choose the right mode, give the right context, verify output quickly, and avoid expensive loops that do not improve the result.

Key facts to remember

• Copilot Business includes 1,900 AI credits per assigned user per month.
• Copilot Enterprise includes 3,900 AI credits per assigned user per month.
• Existing Business and Enterprise customers receive promotional monthly included credits from June 1 through September 1, 2026.
• Included AI credits are pooled at the billing entity level, not held in isolated buckets per user.
• After pooled credits are exhausted, additional usage either continues at published rates or is blocked, depending on policy.
• User-level budgets can halt a user's Copilot access even when the organization still has pooled credits available.
• Copilot Code Review consumes AI credits and GitHub Actions minutes.

02 Included AI Credits & Pooling

Each Copilot license contributes its AI credit value to a single shared pool. All licensed users draw from this pool first; usage beyond it is additional spend.

Example: an organization with 100 Copilot Business users receives a shared pool of 190,000 standard AI credits per month. During the promotional period, that same organization receives 300,000 AI credits per month.

Tokens drive cost

03 How Model Pricing Works

All prices below are per 1 million tokens — the usage-based billing rates GitHub lists for June 1, 2026. A small prompt to a lightweight model can cost a fraction of a credit; a long agent session on a frontier model across many files costs more.

OpenAI models

GPT-4.1 and GPT-5 mini are included models. GPT-5.4 pricing applies to prompts with 272K tokens or fewer.

Anthropic models

Anthropic models include a cache write cost in addition to cached input.

Google models

Gemini 2.5 Pro and Gemini 3.1 Pro pricing applies to prompts with 200K tokens or fewer. Gemini 3 Flash has no long-context surcharge.

xAI & fine-tuned GitHub models

Raptor mini uses GPT-5 mini pricing. Goldeneye uses GPT-5.1-Codex pricing.

04 User Guide — Token-Smart Habits

Keep Copilot useful while being thoughtful about AI credits. The main habit is simple: start narrow, give Copilot the context it needs, and expand only when the task requires it.

Choose the right Copilot mode

Use a prompt template

Goal: <specific outcome>
Scope: <files, folders, or APIs Copilot should consider>
Constraints: <what must stay unchanged>
Verification: <tests, commands, or manual checks to run>
Before editing: summarize the likely change in two sentences.

Prompt for less waste

Use specific prompts that reduce unnecessary context and retries. Avoid vague prompts like Look at this repo and improve it or Fix everything and make the code better, which trigger broad exploration.

Review only this diff for functional regressions, missing tests, and surprising
behavior. Prioritize concrete findings with file references.

05 Admin Guide — Budgets & Controls

Prepare budget controls and operating habits so high-value Copilot usage stays available while spend stays visible and manageable. Budgets are set in USD; usage appears in AI credits ($10 covers 1,000 credits).

Budget levels

Recommended policy starting points

Coaching signals

Monthly review template

Month:
Total AI credits used:
Included pool:
Additional spend:
Top repositories by usage:
Top workflows by usage:
High-value examples:
Coaching opportunities:
Budget changes for next month:
Reusable prompts or instructions to publish:

06 Training — Agent Quality & Token Optimization

When tokens are cheap, agent accuracy is not important. Once they are not, quality becomes the better focus. Instead of counting tokens, make every token count.

The compound error problem. Even at 99% accuracy per step, a 50-step workflow only lands near 60% success. Agent errors compound quickly, so quality work upstream protects both ROI and credits.

The two biggest levers

Research → Plan → Implement

Persistent guidance that pays off

• copilot-instructions.md — keep it small and human-written; use it as an agent-miss log and to trim output.
• Custom agents, skills, MCP servers, and subagents — attach the right tools and context only when needed.
• Scoped & prompt files — conditional instructions and reusable manual starting points.
• Apply good architecture — clear domains and guardrails make agent discovery easier and reduce wasted sessions.

07 Quick Transition Checklist

08 References

• Models and pricing for GitHub Copilot
• Usage-based billing for organizations and enterprises
• README — billing model, included credits, and model pricing tables
• User guide · Admin guide · Training guide
• 📄 Download the slide deck (PDF) · Version française (PDF)