Loading today’s stories
Build1 distinct publisher3 min readPublished
Completions stay free and the overage meter never cuts you off, so the choice between GitHub's credits and Claude Code's five-hour windows turns on how often your team reaches for an Opus-class model.
The Engineer · Build desk
Compiled by The EngineerSomething wrong?How this is made
Here is what happens when you send a chat turn under the new scheme. Copilot counts the input and output tokens, prices them at that model's published API rate, converts the dollars into credits at a hundred credits per dollar, and draws them from your monthly pool [2]. Nothing in that path caps the size of a turn. An agent task that pulls a large repository context into the prompt bills every token of that context, each time it re-reads it.
That is why the most quoted number in the backlash needs unpacking. The dev.to writeup cites a community user who calculated Opus at roughly 27x more expensive under the new math [14]. The published rates do not get you there: Opus-class models run $5 per million input and $25 per million output against Sonnet 4.6's $3 and $15 [11][10], which is 1.67x per token in both directions [1]. The distance between 1.67x and 27x is a claim about how many tokens that user's Opus turns carried relative to what one premium request used to cost. For it to transfer to your team, your context sizes and agent loop lengths have to look like theirs.
The same calculation puts a Pro+ allowance at about 140 Opus requests a month [15]. Pro+ carries 7,000 credits [6], so the implied request costs 50 credits, or 50 cents [2]. At $25 per million output tokens, 50 cents buys 20,000 output tokens [3]. That is an agent turn, not a chat question, which is worth knowing before you compare it with anything. Carry the figure up the ladder and Copilot Max's 20,000 credits covers roughly 400 such requests [4].
Claude Max 5x costs the same $100, adds Opus-class access, and allows roughly 50 to 225 prompts per five-hour window [19]. That looks like more work per dollar than 400 requests a month. But the weekly cap is denominated in compute hours [18], and nothing published converts compute hours into prompts, so the two $100 plans cannot be reconciled on paper. What does settle is the direction of the decision. If Copilot is mostly autocomplete for you, nothing changed, because inline completions and Next Edit suggestions stay unlimited and never touch credits [9]. If you live in Opus agent loops, Copilot bills you in dollars with no cutoff [16] while Claude Code bills you in waiting, on a rolling five-hour reset and a weekly window that clears Monday [21], with reduced limits on weekday mornings from 5 to 11 Pacific [22].
One structural detail in the tier tables deserves a second read: $10 returns $15 of metered usage, $39 returns $70, and $100 returns $200 [5][6][7], so the match climbs from 1.5x to 2x as you spend more [5]. GitHub is discounting the exact usage pattern that generated the complaints.
A note on sourcing. All of this comes from one writeup whose author says he pays per token through APIs on his own rented infrastructure, has not subscribed to the credit system, and worked from published rate cards [23]. The 10x to 50x jumps and the 27x are user reports rather than audited invoices [4][14]. The announcement thread's 958 thumbs down against 24 thumbs up [3] is a ratio near 40 to 1 [6], which is not a survey, though it is the cheapest telemetry GitHub will ever get on who reads a rate card before subscribing.
Ranked by verification strength, evidence, and original report placement.
Copilot Pro at $10 per month includes $10 in base credits (1,000) plus $5 in flex credits, for 1,500 credits total, about $15 of metered usage.
Copilot Pro+ at $39 per month includes 3,900 base credits plus 3,100 flex credits, for 7,000 credits total, about $70 of usage.
Copilot Max at $100 per month includes 10,000 base plus 10,000 flex credits, for 20,000 credits and $200 of usage, effectively a 2x match on the subscription price.
On June 1, 2026, GitHub switched Copilot billing from counting requests to metering tokens.
One GitHub AI Credit equals $0.01 of metered model usage. Every chat message, agent task and code review consumes tokens, the tokens are priced at each model's published API rate, and the dollar total converts to credits drawn from a monthly pool.
The GitHub community discussion announcing the change has 24 thumbs up and 958 thumbs down.
Distinct publishers with included, body-backed reporting in this cluster.
dev.to
1 article · August 31, 2026
Follow any of these and your For You feed starts watching them — no settings page required.
build
AI-written code fails the same four ways, and every gate you own reports green1 distinct publisher
build
Claude Code now outruns Copilot roughly two to one in JetBrains' survey of 15,000 developers1 distinct publisher
build
Copilot's meter changed on June 1, and half your seats are still priced in the old unit1 distinct publisher
security
The model is now choosing the extortion targets, not just writing the malware1 distinct publisher
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.
Rate cards firm, cost shocks hearsay
The pricing skeleton is the strong part: credit conversion, per-tier allowances, per-token rates and Anthropic's window limits are all the kind of figure you can hold against a published price page. The weak part is exactly the part readers will quote. The 10x-to-50x jumps and the 27x Opus figure come from developer posts and one community user's sums, and dev.to's own rate table undercuts the framing, since Opus costs 1.67x Sonnet per token, not 27x. The author also concedes that published sources disagree slightly on what each plan contains.
Already billing, reaction hostile
This is not a preview. Metering has been the live billing basis for Copilot since June 1, which makes uptake involuntary for every paid seat rather than a matter of developer enthusiasm. Two more changes are already scheduled onto one date: the org credit boost lapses and code review starts drawing Actions minutes. What is missing is scale. The only quantified user signal anywhere in our coverage is a vote tally on GitHub's announcement thread, 958 down to 24 up, and vote counts are not usage.
Sober piece, borrowed alarm
The analysis itself resists drama: it concedes completions are unchanged, that light users spend under their subscription, and that the crossover to Claude's flat rate sits around daily agent use. The overstatement is imported. Repeating 10x-to-50x and 27x near the top lends measured authority to numbers the piece never reconciles with its own rate table, where the widest published gap between model classes is 1.67x per token. The direction of the change is real; the magnitude is louder than the arithmetic on the page.
Stake declared, no vendor tie
The disclosure is better than most: two years of paying per token on his own rented infrastructure, and an admission that he has never subscribed to the Copilot credit plan he is pricing. Both cut the same way, toward someone already comfortable with meters, which is worth remembering when the piece calls the disappearance of a predictable ceiling a philosophical difference rather than a problem. No vendor relationship, affiliate arrangement or sponsorship appears anywhere in this reporting, and neither GitHub nor Anthropic is quoted defending its model.
One writer, two rate cards
Everything here comes from a single developer's write-up, so nothing has been checked by a second desk and the vendors' own words are absent. That caps how far the assessment can go, even though the structural facts are the sort that a reader can verify in ten minutes against two pricing pages. Confidence would move sharply on a GitHub billing document, an Anthropic limits page, or any second account that reproduces the 27x figure with real invoices.