The Vibe Father
Pricing analysis
GitHub Copilot changed how it bills in 2026, and the change is bigger than a line item. The old "premium requests" model — a monthly bucket of allowances for the more expensive AI actions — was retired in favor of usage-based AI credits, reported at roughly a cent per credit. If you use Copilot casually, you may barely notice. If you lean on it hard, or if you were counting on a fixed monthly ceiling, this reshapes your bill and possibly your tool choice. Here is what actually changed, who wins, who pays more, and how it stacks up against bringing your own keys.
What the old model was
For most of Copilot's paid life, the pitch was a flat subscription with inline completions included and a monthly allotment of "premium requests" — a capped number of the pricier AI operations (agent actions, premium models, the heavier stuff) baked into your plan. It was easy to reason about: pay your monthly fee, get your allowance, and the meter mostly stayed out of sight. The catch was that heavy users hit the allowance wall and had to top up, while light users effectively subsidized a headline number they never approached.
What replaced it
In 2026, GitHub moved to usage-based AI credits and retired the premium-requests framing. Instead of a fixed bucket of allowances, your heavier AI actions draw down credits, reported at approximately $0.01 each, and you pay for what you consume. Inline completion — the classic ghost-text autocomplete — remains part of the base experience, but the metered AI work now scales with how much of it you actually do rather than being pre-portioned into a plan tier.
The full details live on the Copilot product page, and the specifics of which actions cost how many credits will keep evolving, so treat any exact per-action figure as reported rather than fixed. The structural shift is the story: from a fixed allowance to a variable meter.
Who wins under the new model
Usage-based pricing is genuinely better for a large group of people. If you use Copilot lightly — mostly autocomplete, the occasional agent action — you now pay closer to what you actually consume instead of subsidizing a premium bucket you never emptied. Occasional and intermittent users, people who dip into the heavier features a few times a week, and anyone whose usage is spiky rather than sustained tend to come out ahead. You are no longer paying for headroom you don't use.
Who pays more
The people who feel this are the heavy users, and heavy usage is exactly what agentic coding encourages. Running an agent through a multi-file refactor, letting it iterate on a failing test suite, driving long autonomous sessions — these consume the metered actions, and under a per-credit meter, more work means a bigger bill with no ceiling to lean on. The fixed-cost predictability that made the old model easy to budget is gone. For an individual power user or a team where everyone runs agents daily, the monthly number can now move in ways that are harder to forecast, and it climbs precisely as you get more ambitious with the tool. That is the fundamental tension of any metered model: the moment it becomes indispensable is the moment it starts costing real money.
The BYOK comparison
This is the fork in the road worth thinking through, because a usage-based meter invites a direct comparison to the alternative model: bring your own keys. With BYOK, you pay the model provider directly for inference at their published rates, and the tool sitting on top charges you a flat fee — or nothing, in the case of open-source agents — rather than reselling that inference with a margin on top.
The two structures aren't strictly better or worse; they optimize for different users. A metered, resold model like credits is convenient — one bill, one vendor, no key management — and for a light user it can be the cheapest path. BYOK trades that convenience for control and, for heavy users, cost: when you buy inference at wholesale and pay a flat fee for the harness, the margin that a reseller would take stays in your pocket, and the more you use it the more that matters. We laid out the full case in stop paying token markup, and the economics generally in the economics of AI coding.
The honest summary: if you are a heavy user, a usage-based meter is the model most likely to surprise you on the bill, and BYOK is the model most likely to reward your volume. If you are a light user, the meter's convenience may genuinely be worth it.
What this signals about the market
Copilot's shift is part of a broader pattern. As agentic features move from occasional novelty to constant use, every vendor has to decide who pays for the inference and how. Some are moving to metered credits (Copilot). Some bundle a house meter into a subscription (the editor-first tools). And some hand you the keys and take a flat fee (the BYOK and open-source camp). None of these is a scam; they are different answers to the same rising cost. But the direction of travel is clear — the "unlimited-feeling" era of a flat fee with a generous allowance is ending, and usage is increasingly what you pay for.
That is exactly why the model-agnostic, keys-in-your-own-hands approach is having a moment. When inference is the cost, controlling how you buy it becomes the lever. It is the reasoning behind our own pricing: The Vibe Father is BYOK at a flat $20/mo, so we earn nothing when you burn tokens and our incentives don't diverge from yours as your usage grows. Whether that fits you depends on how heavy a user you are — which is precisely the question Copilot's new meter forces everyone to answer.
Bottom line
GitHub Copilot's 2026 move to usage-based AI credits (reported at about a cent each) retired the premium-requests allowance and replaced predictability with a meter. Light users mostly win; heavy users mostly pay more and lose their fixed ceiling. If your usage is climbing — and agentic coding pushes it up — it is worth pricing out BYOK before the meter decides for you. Next reads: Copilot vs Cursor, stop paying token markup, and the best free AI coding tools. Model costs and scores at /benchmarks.
