Loading today’s stories
Build1 distinct publisher2 min readPublished
Premium requests are now labelled legacy and AI Credits bill at one cent each. Annual Pro and Pro+ seats keep the old meter, so one org ends up budgeting in two units with a four-to-one conversion.
The Engineer · Build desk
Compiled by The EngineerSomething wrong?How this is made
Start with the exchange rate. A premium request at the standard 1x rate bills at four cents once overages are switched on [5], while an AI Credit is one cent [2]. One legacy request is therefore worth four credits, and the older meter has no way to express anything finer than four [1]. That is a real gap at the bottom of the model table: the lightweight tier, Claude Haiku 4.5 and Gemini 3 Flash among them, draws about 0.33x [14], which the request meter charges out at roughly 1.3 cents a prompt [2].
Now price each plan's included allowance at that same four-cent overage rate. Copilot Pro costs $10 per user per month and includes 300 premium requests [6], which comes to $12 of metered capacity [3]. Pro+ at $39 includes 1,500 [7], or $60 [4]. Copilot Business at $19 per user includes 300 [8], the same allowance as a $10 Pro seat [6], worth $12 [5]. Enterprise at $39 per user includes 1,000 [9], or $40 [7]. Only the individual plans hand back more overage-priced capacity than they charge for [8].
The legacy unit is also not sitting still. Copilot code review deducted a single premium request when the feature shipped; since June 1, 2026 it carries a 13x multiplier [4], which makes one review 52 cents at the standard overage rate [9] and puts an Enterprise seat's 1,000 requests at roughly 76 reviews a month [10].
Running dry is not a stop on paid plans. The 0x tier, including GPT-5 mini, GPT-4.1 and GPT-4o, keeps answering after the allowance is gone, slower at peak but free [12], and inline completions never touch the meter at all [17]. If overages were never enabled, premium-model access pauses instead until the next reset [13], and that reset lands on the 1st of the month at 00:00 UTC with nothing carried over, on a date that does not have to match the invoice cycle [11].
So the finance problem is not the conversion, which is clean arithmetic. It is that an org holding annual Pro and Pro+ seats alongside credit-metered ones is forecasting against two denominators for one headcount [3], and the variable in both is the multiplier table rather than the seat price. A team can hold its plan, its price and its allowance constant for a year and still watch the cost of a pull request review move by an order of magnitude [4].
Ranked by verification strength, evidence, and original report placement.
As of June 1, 2026, GitHub moved Copilot from request-based billing to usage-based billing, and premium requests are labelled "legacy" throughout GitHub's own documentation.
Annual Copilot Pro and Pro+ subscribers who stayed on their existing plans are still billed in premium requests.
Copilot code review originally consumed one premium request per review; since June 1, 2026 it carries a 13x multiplier, so a single review deducts 13 premium requests.
Model multipliers run from 0x for base models through roughly 0.33x for lightweight models such as Claude Haiku 4.5 and Gemini 3 Flash, 1x for the frontier tier, 3x for Claude Opus-class models, and 15x to 30x for the heaviest preview tiers.
The replacement for premium requests is GitHub AI Credits, metered at one cent per credit.
If overages are enabled, additional premium requests bill at $0.04 per request at the standard 1x rate, so a 3x model interaction costs $0.12.
Follow any of these and your For You feed starts watching them — no settings page required.
Evidence-backed comparisons of source perspectives and observed adoption signals. Read the methodology
Which Builder, Operator, and Investor concerns the observed source mix emphasized—not a truth score.
Evidence, demonstrated adoption, hype gap, incentives, and confidence are assessed independently, each on its own current evidence. How these are measured.
Specific numbers, single secondary source
Every claim is concrete and checkable — dates, per-tier prices, allowances, multipliers, a $0.04 overage rate and a one-cent credit — and the derived arithmetic follows cleanly from the stated figures. But the entire cluster is one dev.to community explainer that quotes GitHub's documentation and model multiplier reference without linking either, no primary GitHub changelog or pricing page is present, no second publisher corroborates the June 1, 2026 cutover, and the article's own credits section is cut off mid-sentence. There is also unresolved internal tension in applying a new 13x code review multiplier to a meter the same article calls legacy as of the same date.
No uptake data
The cluster documents vendor-side pricing actions only. There is no disclosure of how many organizations or seats migrated to AI Credits, how many remain on the legacy premium request meter, how much credit or request volume is actually consumed, or any deployment, benchmark or customer usage figure. Adoption cannot be measured without inferring facts the source does not provide.
Framing outruns the evidence base
The mechanics themselves are stated soberly and are not inflated. The overstatement is in scope and certainty: the cluster's framing that half an organization's seats are still priced in the old unit is asserted without any seat-mix or migration data, and a four-to-one request-to-credit conversion is presented as a budgeting equivalence when the source itself says credits meter tokens processed rather than flat per-prompt multipliers, so the two units are not interchangeable. A single uncited community post is also being read with more authority than it can carry on live pricing.
SEO explainer with cost-advisory framing
The item is a keyword-shaped reference post ('if you are searching for how they work in mid-2026') on a community publishing platform, a format that rewards comprehensiveness and freshness over verification. It also carries a light commercial tell, referring to 'a pattern we see across every vendor in seat and usage right-sizing', which points toward cost-optimization advisory interests. Separately, all pricing facts originate with the vendor whose pricing is under discussion and are relayed without citation. No paid placement, affiliate relationship or sponsorship is disclosed or evident, so the distortion risk is moderate rather than acute.
Low: one uncited secondary account
The internal arithmetic is reliable and the claims are precise, which supports moderate confidence in the derived economics conditional on the inputs. Confidence in the inputs themselves is low: one publisher, no primary GitHub source, unverifiable model version names, a truncated section covering the current credit allowances, an unreconciled overlap between the legacy label and a new multiplier on the same date, and zero adoption evidence. Treat as a lead to verify against GitHub's billing documentation, not as a settled pricing record.
build
AI-written code fails the same four ways, and every gate you own reports green1 distinct publisher
build
Grok 4.6 lands in Copilot two days after launch, and the model picker becomes a procurement problem1 distinct publisher
build
244 kB, 500 a minute, 5 percent: three ceilings that fail for the same reason1 distinct publisher
build
The failure modes of AI agents in low-level code have names, and names can be gated in CI1 distinct publisher
Distinct publishers with included, body-backed reporting in this cluster.
dev.to
1 article · August 25, 2026