GitHub Copilot's token-metered billing vs Claude Code flat rates: the new cost math

A meter replaces the request counter

On June 1, 2026, GitHub stopped counting Copilot usage as requests and began metering tokens. Every chat message, agent task and code review now draws from a pool of AI Credits, where one credit equals $0.01 of metered usage and tokens are priced at each model's published API rate. According to a dev.to analysis of GitHub's rate sheets, the community discussion announcing the change collected 24 upvotes against 958 downvotes, and some developers reported effective costs jumping 10x to 50x overnight depending on which models they had selected.

What each plan now buys

GitHub structures the plans as base credits plus a flex allotment, per the analysis:

• Copilot Pro, $10 a month: 1,000 base credits plus $5 of flex credits, or 1,500 total, roughly $15 of metered usage.
• Copilot Pro+, $39 a month: 3,900 base plus 3,100 flex, 7,000 credits total, about $70 of usage.
• Copilot Max, $100 a month: 10,000 plus 10,000, or 20,000 credits, effectively a 2x match worth $200 of usage.
• Business ($19 a seat) and Enterprise ($39 a seat): credits pooled across the organization, temporarily raised to 3,000 and 7,000 per user.

The analysis flags a step-down teams should plan for: when the promotional window ends on September 1, 2026, per-user credits fall to 1,900 and 3,900.

Completions stay free; model choice is the bill

Three details matter more than the headline allowances. First, inline completions and Next Edit suggestions remain unlimited on paid plans and never touch credits, so autocomplete-heavy users see no change at all.

Second, model choice dominates cost. The Copilot catalog contains twenty-plus models with widely varying rates: the analysis cites Claude Sonnet 4.6 at $3 per million input tokens and $15 per million output, Opus-class models at $5 and $25, and lightweight options such as GPT-5 mini included at zero credits. One community calculation cited there puts Opus at roughly 27x more expensive under the new math, with a Pro+ allowance covering about 140 Opus requests a month.

Third, overage behavior flipped. The old request caps cut users off; the credit system keeps working and bills the excess at $0.01 per credit, trading a predictable ceiling for a variable bill that depends on model discipline. From September 1, code review also consumes GitHub Actions minutes on top of AI Credits, a double charge teams will notice on one invoice.

Claude Code rations instead of metering

Anthropic's Claude Code never left flat-rate pricing, according to the analysis; it throttles instead. Claude Pro at $20 a month provides full Claude Code access with roughly 10 to 45 prompts per five-hour rolling window, a weekly cap and Sonnet-class models only. Max 5x at $100 offers five times the per-session usage, Opus-class access and roughly 50 to 225 prompts per window; Max 20x at $200 targets full-time agent users. Since March 2026, Anthropic has also reduced five-hour limits during weekday peak hours, 5 to 11 AM Pacific. There is no overage invoice, but also no guarantee you can finish a task before the cooldown resets.

Pricing one real session

Plan tables hide the decision, so the analysis prices a representative agentic session, one that reads repo context, plans and writes a change, at about 50,000 input and 20,000 output tokens, based on tracked Copilot billing breakdowns. At Sonnet 4.6 rates that works out to $0.45 per session, or 45 credits. Output tokens are only 40 percent of the volume but 60 percent of the cost, because output is priced 5x input on Anthropic models.

At that rate, four sessions a week costs about $7.80 a month and fits within Pro's allowance. Thirty sessions a month runs roughly $13.50, which exceeds Pro but sits comfortably inside Pro+. Thirty Opus-class sessions land near $22.50. The crossover point, the analysis argues, is daily agent usage: below it Copilot's meter is cheaper because you spend less than the subscription price; above it Claude's flat rate wins because each extra session costs nothing until the throttle bites. The author, who discloses running token-metered API infrastructure but not a Copilot credit subscription, recommends auditing a real week of usage, checking which models the workflow actually needs, and setting spend caps either way.

Why it matters

The switch turns AI coding assistance from a flat subscription into a cloud-style variable cost, where identical behavior can be cheap or ruinous depending on which model is selected by default. Budget discipline becomes an engineering skill: lightweight models for routine work, frontier models reserved for tasks that need them, and caps configured before the September 1 team-credit step-down. It also marks a fork in the market. GitHub prices usage and allows unlimited work; Anthropic caps usage and charges nothing extra. Every team now has to decide whether a bigger invoice or a locked-out afternoon is the worse failure mode. As agent workflows multiply token consumption, metered billing looks like where the industry is heading, and Copilot's rough reception suggests the transition will be turbulent.