Every token counts against your credits now. These habits cut the bill without making the work worse, sorted by how much they usually save. Many follow GitHub's own optimization guide.
Copilot charges each model's API rate. Claude Haiku 5.5 and GPT-6 Luna cost $0.10 in / $0.50 out per million tokens; Claude Fable 5.1 and GPT-6 Astra cost $10 / $50. That is a 100× spread. Find the cheapest model on the frontier that scores high enough for the task.
Model picker → choose per taskHigher effort means more thinking tokens. GPT-6.1 Sol uses about 14K tokens per benchmark task at medium effort and 68K at max, for a score of 46 vs 55. Claude Opus 5.5 goes from 44K to 173K. Save high effort for debugging and architecture.
Reasoning level → medium (VS Code, CLI)Every agent turn re-sends the whole conversation as input. A chat that wandered through three bugs pays for all three on every new message. Start fresh when you switch problems.
New chat · CLI: /new or /clearCached input costs a fraction of fresh input: $0.20 vs $4.00 per million for Claude Opus 5.5, $0.10 vs $2.00 for GPT-6.1 Sol. The cache belongs to one model, so a switch re-sends the whole context at full price.
Research and plan with a strong reasoning model, write the plan to a file, then implement it in a fresh session with a mid-tier model. The build session starts small and stays focused.
VS Code: Plan agent · CLI: /planCode completions and next edit suggestions are not billed in AI credits on paid plans. For a rename or a one-line fix, typing and accepting a suggestion costs nothing. Asking chat does.
Tab to acceptAuto model selection gets a 10% discount on model costs on paid plans, in Chat, the CLI, the Copilot app and the cloud agent. It is a solid default when you don't need a specific model.
Model picker → AutoSearching the codebase and summarizing files rarely needs a frontier model. A custom agent pinned to a Haiku- or Luna-class model does the reading; your main session receives only the summary.
Toolkit → scout agentWhen a session has to continue, summarize instead of carrying the full history. Check what fills the context window before adding more.
CLI: /compact and /contextAttaching the two files that matter gives the model exactly what it needs. Broad codebase searches pull in many chunks you pay for on every turn.
#file:lib/pricing.tsThe definitions of every enabled tool and MCP server are sent with each request. Ten idle MCP servers are a fixed tax on every turn. Turn on toolsets per task.
Chat → Configure toolscopilot-instructions.md and AGENTS.md are added to every request. Keep them to the layout, the commands and the few rules that matter. Long style guides belong in skills, which load only when relevant.
.github/copilot-instructions.mdSome models charge more once a single request passes a threshold. GPT-6.x doubles its input price above 272K tokens. Claude Haiku 5.5 costs 5× more above 100K. Starting fresh is cheaper than crossing the line.
Changing the reasoning level, context size, tools or MCP servers mid-session also breaks the cache. Decide on the setup first, then work.
Give every agent task a clear goal, the relevant context up front and a stopping condition: "stop when the tests pass; don't touch unrelated files." Agents without one keep going and keep spending.
A type checker, linter or test run catches mistakes for free. Without them, the agent spends another round of tokens finding the problem by reading code.
GitHub's own estimates per review: Lite about $0.25, Balanced about $1.50, Max about $10, plus Actions minutes. Cost grows with PR size, so small PRs are cheaper to review.
Caches expire after inactivity: about 24 hours for OpenAI models and 1 hour for most others. After a long break, a fresh session with a short summary can cost less than resuming.
The Copilot icon in the VS Code status bar shows usage and the reset date. On github.com, the AI usage page breaks credits down by model, which shows where the money goes.
Billing → AI usageThe Copilot CLI and SDK accept an AI credit limit per session, a soft cap that keeps an unattended agent from running away.