By GitHub Agentic Workflows entered public preview on June 11 alongside a welcome security upgrade: agentic workflows no longer require a Personal Access Token, switching to GitHub Actions’ built-in GITHUB_TOKEN instead. Two announcements, both real. But the more consequential change landed two days earlier, on June 8, when gh aw quietly migrated from Effective Tokens to AI Credits as its primary billing metric. One AI Credit equals $0.01. Agentic runs burn far more of them than chat completions. If you haven’t opened your usage dashboard since June 1, now is the time.
What Actually Shipped on June 11
GitHub Agentic Workflows is now available to all GitHub users — no early access waitlist required. The product lets you define repository automations in plain Markdown files. Describe what you want in natural language, run gh aw compile, and the tool generates standard Actions YAML. Issue triage, CI failure analysis, PR review, documentation updates — all expressible in a few paragraphs of Markdown instead of hundreds of lines of YAML.
The PAT elimination matters operationally. Long-lived personal access tokens are an audit and rotation headache at scale. The new model uses scoped, ephemeral GITHUB_TOKEN credentials — per the June 11 changelog — with a sandboxed container, Agent Workflow Firewall, and read-only permissions by default. Org-level workflows bill AI credits to the organization once you enable the “Allow use of Copilot CLI billed to the organization” policy.
The June 8 Change That Actually Changes Your Budget
Effective Tokens (ET) was a normalized proxy metric — useful for comparing runs across models, but disconnected from what you actually pay. AI Credits (AIC) map directly to dollar cost. The formula: input tokens × input rate plus output tokens × output rate, divided by one million, times 100. Every gh aw audit and gh aw logs output now shows AIC as the primary figure, per the June 8 migration post. Effective Tokens remains only as a legacy compatibility field.
This matters because the price gap between models is enormous — and ET obscured it. An agentic task that costs 1 AI Credit on GPT-4o costs 46 on o3. A nightly cron workflow running issue triage with o3 as default is not a productivity choice; it’s negligent cost management. The same workflow on GPT-4o costs less than one cent per run. On o3, you burn half a dollar before you’ve checked Slack in the morning.
The Real Cost of an Agentic Session
Chat completions are cheap. Agentic runs are not. A task that plans a multi-file change, reads a dozen source files, drafts a diff, runs tests, and iterates can consume over 100,000 tokens in orchestration alone — and one to 3.5 million tokens across the full run with retries. On Claude Sonnet 4.6, 100,000 output tokens costs $1.50. Stack a few of those per day on a Pro plan ($10/month, approximately 1,000 credits) and you’ve exhausted your monthly allotment before Wednesday.
GitHub’s own data makes the scale concrete. Over 30 days, their internal agentic workflows consumed 215,998,387 tokens across 333 runs — an estimated $21.60. That’s modest for an organization GitHub’s size, but only because they ran an aggressive optimization program first.
GitHub’s 62% Cost Reduction Playbook
GitHub published three techniques that drove the savings, detailed in their token efficiency post:
- Prune MCP tools. Each MCP server with 40+ tools adds 10–15 KB of schema per request. Most workflows use a fraction of those tools. Pruning unused tools saves 8–12 KB per call — trivial individually, significant at scale.
- Replace MCP calls with
ghCLI. Where a native CLI invocation suffices, use it. MCP routing carries protocol overhead; direct CLI calls do not. - Deploy a Daily Auditor agent. A workflow that scans all other workflows, identifies expensive patterns, and files GitHub Issues with specific fixes. The feedback loop is automated — the agent reports on the agents.
Results: Auto-Triage Issues workflow down 62%, Smoke Claude down 59%, Security Guard down 43%.
What to Do Now
# Check AI Credit spend per workflow
gh aw audit
# Inspect model calls and token breakdown for a specific workflow
gh aw logs auto-triage
# Compile a workflow with explicit model selection
gh aw compile .github/workflows/auto-triage.mdRun gh aw audit today. If any workflow defaults to o3 or Claude Opus, switch it to GPT-4o unless the task genuinely requires frontier-level reasoning. Set a hard budget cap — the default behavior is notification only, not enforcement. Enable “Stop usage when budget limit is reached” in GitHub’s billing settings explicitly. Prune your MCP tool lists to what each workflow actually needs.
The public preview is a good time to treat agentic workflows as the production infrastructure they now are — which means cost governance from day one, not capability exploration followed by a billing surprise.
