By ByteBot
GitHub switched Copilot to usage-based AI credit billing on June 1. By day two, some developers on the $39 Pro+ plan had burned through most of their monthly allocation before lunch. The model-price gap between cheapest and priciest options is 24x. New session limit controls arrived July 1. If you have not checked your usage dashboard, stop reading and go look.
What Costs Credits and What Does Not
The billing model is more nuanced than a flat per-use charge, and the nuance matters. Code completions and next edit suggestions remain completely unlimited on all paid plans. What does consume credits is everything else: Copilot Chat, Copilot CLI, the cloud agent, Copilot Spaces, Spark, and any third-party coding agents integrated through the platform.
That distinction trips up a lot of developers. The assumption is that a $39/month subscription covers Copilot as a whole. It covers autocomplete. The moment you open a Chat session or run an agent task, you are drawing from your monthly credit pool.
Plan allocations for the current promotional period (June through August 2026):
- Copilot Pro ($10/mo): 1,500 AI credits included
- Copilot Pro+ ($39/mo): 7,000 AI credits included
- Copilot Max ($100/mo): 20,000 AI credits included
One AI credit equals $0.01. Credits reset on UTC day one each month and do not roll over. Check the official GitHub Copilot billing documentation for the full breakdown.
The Model Price Spread Is 24x — and That Is Your Biggest Lever
The single most impactful variable in your Copilot bill is not how much you use it — it is which model handles the request. The spread is enormous:
- GPT-5.4 nano: $0.20/M input, $1.25/M output
- GPT-5 mini: $0.25/M input, $2.00/M output
- MAI-Code-1-Flash: $0.75/M input, $4.50/M output
- Claude Sonnet 4.6: $3.00/M input, $15.00/M output
- GPT-5.5: $5.00/M input, $30.00/M output
To make that concrete: a heavy agent iteration session with 250K input and 20K output tokens costs $0.28 on MAI-Code-1-Flash and $1.85 on GPT-5.5 — the same task, 6.7x the price. A Pro+ subscriber with 7,000 credits can run 136 medium GPT-5.5 agent steps or 909 MAI-Code-1-Flash steps on the same monthly budget.
The recommendation is not to default to the cheapest model on everything — that trades cost for quality at the wrong times. Use Auto model selection instead. It picks the appropriate model for the task and gives you a 10% discount on top of that. Reserve manual model selection for the cases where you explicitly need a specific capability.
One more thing on cost: output tokens are the expensive side of the equation. Capping response length where appropriate delivers 20 to 70 percent cost reduction, depending on the model.
Session Limits: The Control That Shipped July 1
GitHub shipped session limits for Copilot CLI and SDK on July 1, 2026. The feature is available in Copilot CLI 1.0.66 and later, and Copilot SDK 1.0.5 and later. Run copilot update if you are not current.
Session limits cap how many AI credits an agent or conversation can consume in one run. In an interactive session:
/limits set max-ai-credits 200
/limits unsetFor scripted or automated runs:
copilot -p "refactor auth module" --max-ai-credits 200These are soft caps — if a model response is already processing when the limit is reached, that response finishes and then the session stops. The tracking covers model calls, subagents, and background compaction work, which matters for agentic tasks where background operations add up. Set your limit above 30 credits; most individual model calls cost more than 20.
For SDK users, the equivalent is:
sessionLimits: { maxAiCredits: 200 }Set it when creating or resuming a session. Usage is checked after model calls return, so the soft cap behavior applies here too.
Admin Controls: The Dangerous Default You Need to Change
If you are managing Copilot for a team, there is a default setting that looks like a guardrail but is not. When you create a spending budget at the cost center, organization, or enterprise level, the option called “Stop usage when budget limit is reached” is off by default. Without enabling it, your budget is an alert, not a hard stop. Charges will continue to accrue past the limit.
User-Level Budgets, by contrast, always enforce hard stops — no toggle required. The GitHub budget controls documentation covers all six layers of the budget hierarchy: universal ULB, cost center ULB, individual ULB, cost center budget, organization budget, and enterprise budget. The lowest remaining headroom wins at any given moment, so a user-level hard stop will override even a permissive enterprise budget.
September 1: The Credit Cliff
The included credit amounts you see today are promotional. They run through August 2026 and they are higher than the standard amounts that take effect September 1. GitHub has not published the exact post-promotional figures, but if your team adopted Copilot after June 1, your baseline understanding of monthly usage is built on numbers that will drop.
Use the next six weeks to establish your actual usage patterns and set spending controls before the floor shifts.
GitHub’s billing model rewards developers who treat model selection as part of their workflow — the same way you would choose between running a quick linter locally versus spinning up a full CI pipeline. The tools to control costs exist and most of them are a single CLI flag or dashboard toggle away. The developers getting surprise bills were the ones who assumed the subscription handled everything. For autocomplete, it does. For agents and chat, the meter is running.
